Manufacturing (Non-Automotive)AI Implementation & Best Practices

AI layout optimisation for manufacturing plants: from a drawing on the wall to a layout that earns every move

AI layout optimisation for manufacturing plants is the use of machine learning and mathematical search to decide where machines, cells, buffers, storage and aisles sit, and how material moves between them. Models generate and rank candidate arrangements; simulation and encoded safety rules certify them; and people, standards and shutdown windows decide which ones are ever built.

Manufacturing (Non-Automotive) · AI Implementation & Best Practices

Atomic Loops EngineeringIndustrial AI practiceReviewed by Atomic Loops Solutions ArchitecturePublished 20 August 202661 min read

Key takeaways

Plant layout optimisation is a search problem, not a prediction problem. Its classical formulation — the quadratic assignment problem — is NP-hard, and as the authors of a 2023 deep-reinforcement-learning study on it state plainly, no method is known that solves instances larger than about 30 locations exactly. Anything above that size is a ranked shortlist, never a proven optimum.
Sort every layout change by reversibility before you sort it by value. Slotting, staging and AMR traffic rules can change in a shift and be undone in a shift; moving a machine waits for a shutdown window; walls, docks and cleanroom grading are a capital project. AI should be allowed to run the first tier continuously, propose the second, and only ever advise on the third.
The binding constraint is almost never the optimiser. It is the spatial model: a plant's machine-readable geometry drifts away from the floor within months, and once the model is wrong every candidate layout generated from it is confidently wrong too. Track model-to-floor divergence as a first-class KPI or the whole capability rots quietly.
Hard constraints — egress and aisle clearance, hygienic or ATEX zoning, cleanroom grading, crane envelopes, floor loading — belong in the generator as filters, not in the objective function as penalties. A tool that proposes crossing a high-care boundary once will be switched off permanently, and it will deserve to be.
The forecast is not the outcome. A layout change is only proved by a post-move measurement against a comparable unchanged area, with disruption, requalification and ramp-back priced in the same figure as the capital cost. Programmes that stop at the business case are the reason plant engineering distrusts the next proposal.

Abbreviations used on this page

MES: Manufacturing execution system
ERP: Enterprise resource planning
WMS: Warehouse management system
CMMS: Computerised maintenance management system
RTLS: Real-time location system (typically UWB or BLE tags on assets and MHE)
AGV: Automated guided vehicle — fixed-path material transport
AMR: Autonomous mobile robot — free-path material transport
MHE: Material handling equipment (forklifts, tuggers, conveyors, AMRs)
DES: Discrete-event simulation
FLP: Facility layout problem — the formal problem of placing departments to minimise flow cost
QAP: Quadratic assignment problem — the classical mathematical formulation of the FLP
OEE: Overall equipment effectiveness (availability × performance × quality)

What AI layout optimisation actually does in a plant

A definition, the three things the software really has to produce, and the path a candidate layout travels from evidence to floor.

AI layout optimisation in a manufacturing plant is the use of machine learning and mathematical search to decide where things go and how material moves between them — which machines sit in which bay, where cells face, how big the buffers are, where staging and kitting happen, which way the aisles run and where the vehicles are allowed to drive. It is a search and simulation problem, not a prediction problem: the model is not forecasting a number, it is proposing arrangements and being told, by a simulator and by a rulebook, which ones are feasible and which are good.

The classical formulation is the facility layout problem, usually written as a quadratic assignment problem (opens in a new tab) — place n departments in n locations so that the sum of flow multiplied by distance is minimised. It is NP-hard, and the practical ceiling is lower than most buyers assume: the authors of a 2023 study applying deep reinforcement learning to it note that, unlike the travelling salesman problem, no method is known that solves instances larger than about 30 locations exactly. Every real plant is larger than that. So the honest output of any layout tool is a ranked shortlist with a stated evaluation, never a proven optimum, and a vendor claiming otherwise is describing something that does not exist.

That constraint shapes everything downstream. Because the search cannot be exhaustive and its scores cannot be trusted absolutely, the value of a layout capability sits in three artefacts rather than in the optimiser: an accurate machine-readable model of the floor, a measured from-to matrix of what actually moves, and a set of hard constraints written so a machine can check them. Given those three, several different search methods work adequately. Missing any one of them, no search method works at all — it produces confident arrangements that are wrong in ways nobody on the floor can immediately articulate but everybody can immediately feel.

How a candidate layout travels from evidence to floor

The three lanes are evidence, generation and execution. Most plants build the first lane and stop; the value is released in the third, and the loop back from reconciliation to the spatial model is what keeps the whole thing from decaying. Note where the risk sits: not in the optimiser, but in the divergence between the model and the floor.

Data & feeds
Human in the loop
AI / model
System-of-record action
Where value leaks

The process, in words

Evidence lane: the plant's real geometry is captured as a machine-readable as-built model rather than a commissioning drawing; MES and WMS transactions plus RTLS or AMR telemetry produce a from-to matrix of what actually moves, by product family and shift; and each move is priced from MHE hours, labour and lost throughput so distance can be traded against capital and floor space.
Generation lane: the encoded constraint set defines the feasible region before any search begins. A solver with a learned surrogate ranks thousands of arrangements cheaply, the constraint screen removes anything that breaches egress, clearance, zoning or loading rules, discrete-event simulation certifies the survivors against the current mix, and engineering and EHS review a shortlist that already carries a move sequence and a reversion plan.
Execution lane: soft-layer changes — slotting, staging, AMR traffic rules — run immediately under a written policy because they are reversible within a shift. Cell and structural moves wait for an approved shutdown window. Both feed as-built reconciliation, which corrects the spatial model. Skip the reconciliation and model-to-floor divergence grows until the generator is optimising a plant that no longer exists.

Step-by-step insights

The as-built model is the load-bearing artefact, not the optimiser: Almost every failed layout programme fails here, and it fails silently. A plant makes hundreds of small undocumented changes a year — a bench moved a metre, a rack added, a conduit rerouted, a pallet position taken permanently by a tote stack — and none of them individually matters. Collectively they mean the model's clearances, adjacencies and free areas stop being true, and a candidate generator will happily place a machine in a space that is occupied by something nobody recorded. The failure mode is not an error message; it is a proposal that survives review and then cannot be built. Reconciliation is dull, cheap and the single highest-return discipline on this page: update the model from the work order, and sample the floor against the model on a cadence so you know your divergence rate rather than assuming it is zero.
The from-to matrix must be transactional, not observational: Observation weeks measure what an observer can see and what happens while they are watching, which biases toward day shift, toward visible congestion, and toward whatever the plant is running that week. Transaction history from the MES and WMS measures what moved, from where, to where and when, across every shift and every product family, and it is already recorded. The practical work is joining rather than collecting: mapping storage locations and work centres onto coordinates in the spatial model, then aggregating movements into a matrix segmented by product family and shift. Segmentation is what stops the matrix lying: a plant with two product families running on alternating weeks has two different flow patterns, and the average of the two describes neither.
Constraints are filters, and regulated constraints are hard filters: There is a persistent temptation to express constraints as penalties in the objective function, because it makes the search mathematically tidier and lets the optimiser trade a small breach for a large gain. Do not do this for anything safety-related or regulated. Egress width, fire-compartment boundaries, allergen or high-care segregation, cleanroom grade boundaries, ATEX zone edges and crane envelopes are not tradeable at any price, and a tool that proposes trading them once will be — correctly — switched off. Encode them as filters applied before scoring. Soft preferences, such as keeping noisy operations away from the quality lab, belong in the objective; hard rules belong in the feasible region's definition.
Simulation certifies; it does not decide: The generator's score is a cheap proxy — flow times distance, plus whatever penalties were encoded. It cannot see queueing, blocking, starving, changeover interactions or the way a shared forklift couples two areas that look independent on a flow matrix. Discrete-event simulation can, which is why the pipeline is generate-cheaply-then-certify-expensively rather than a single scoring pass. The number that makes the certification honest is the fidelity gap: simulated throughput minus actual throughput for the same period and mix, published with every result. A model that is consistently 8% optimistic is usable, because the reader can correct for it. A model whose gap has never been computed is a drawing that moves.
The soft layer is where AI is actually allowed to run: Notice that only one arrow in the diagram goes straight from the constraint screen to execution without human approval, and it carries only the reversible layer: which pallet goes in which position, where staging is allocated, which routes vehicles prefer, how the kitting point is arranged. These decisions can be re-made every week and undone in a shift, which is precisely why they can be governed by a written policy rather than by an approval queue. Everything bolted down goes the long way round, through certification and review and a shutdown window. Programmes that try to give the optimiser authority over the bolted-down layer do not fail on safety — they fail on approval, and they consume a year discovering it.
The reconciliation loop is the difference between a capability and a study: The arrow from execution back to the spatial model is the one that turns a project into an asset. Each executed change — soft or hard — corrects the model, so the next generation run starts from a truer picture than the last one did, and the capability improves with use. Without it, every study begins by rebuilding the model, which is why plants that have run three layout studies in a decade have paid for the same survey three times. The organisational version of this is equally simple and equally often skipped: make the model update part of the work order's completion criteria, so nobody has to remember to do it separately.

One deliberate exclusion: where the search and simulation compute actually runs — on a workstation, on plant edge hardware, or burst into a cloud tenancy for a weekend — is a real question and a separate one. It is an infrastructure decision with its own data-residency and OT-segregation arguments, and it does not change any of the reasoning on this page. Treat it as a scheduling and architecture matter once the three evidence artefacts exist, not as a prerequisite for starting.

The five rungs of the layout ladder, in detail

Drawn, Measured, Simulated, Optimised, Adaptive — what each looks like on the ground, the signals a reviewer can check in an afternoon, and the mistake that traps plants there.

The ladder below describes how a plant's capability to change its own layout matures, and it is deliberately not a generic AI maturity model. Each rung is defined by what the plant can actually do — describe, measure, evaluate, generate, adapt — and by which layer of the layout that capability reaches. The hallmarks are observable conditions, the diagnostic signals are checks you can run against your own systems this week, and the anti-pattern is the specific error most often made trying to leave that rung.

Value released along the layout ladder

Value stays close to flat through Drawn and Measured — where most plants are — because measurement produces arguments rather than options. It inflects at Simulated, when being wrong becomes cheap, and again at Optimised, when generating an option costs almost nothing. The final rung is flatter than it looks: Adaptive releases value continuously but only across the reversible layer.

Flow cost removed, and kept removed by stage

Stage 1 · Drawn — 31% of operators. The layout exists as a drawing rather than a model, material flow is described from memory, and layout decisions are made in a workshop.
Stage 2 · Measured — 36% of operators. Material flow is quantified from transactions and telemetry, so layout arguments now carry evidence — but the plant still generates its options by hand.
Stage 3 · Simulated — 22% of operators. A calibrated model of the plant exists, so candidate layouts are evaluated rather than argued — and the hard constraints are written down as rules.
Stage 4 · Optimised — 8% of operators. Layout generation is a standing capability: candidates are generated, screened against encoded constraints, certified in simulation, and priced with their disruption.
Stage 5 · Adaptive — 3% of operators. The reversible layer of the layout re-optimises continuously inside a written policy, while physical moves are scheduled from certified evidence.

Curve shape: logistic, plotted from the stage data above. Distribution: Stage framing consistent with the acatech Industrie 4.0 Maturity Index.

Select a rung

Every rung's full detail is in the page source — the selector only changes which panel is visible, so nothing here depends on JavaScript to exist.

Stage 1

Drawn

31% of operators sit here

The layout exists as a drawing rather than a model, material flow is described from memory, and layout decisions are made in a workshop.

Stage 1 is not ignorance of the plant — the people on the floor usually know it intimately. It is the absence of any representation of the plant that a computer could reason about. The geometry lives in a drawing that cannot be queried, the flow lives in the heads of the supervisors who walk it, and the cost of the flow lives nowhere at all. Every layout question therefore has to be re-answered by argument, and arguments are won by the person who has been there longest.

The tell is what happens when someone proposes moving a machine. At stage 1 the discussion is about whether it will fit and whether the services reach — questions about the building — and never about what it does to throughput, congestion or handling cost, because nobody has a number for those. The change is then justified by a benefit that was estimated on the back of the proposal itself, which means the plant has no way of knowing afterwards whether the move worked.

This is a cheaper stage to leave than most people assume, because the first move out of it is measurement rather than construction. It is also expensive to stay in: without a flow record, every layout decision is a fresh guess, and each guess is embedded into concrete for a decade. The compounding cost is not the bad decisions, it is that good ones cannot be recognised as good.

In practice

The staging area nobody chose

In a mid-sized food plant, the pallet staging in front of the packing hall grew where it did because a line was extended in 2014 and the space in front of the old wall was empty. Eleven years later, every pallet from the raw store crosses the packing hall's main pedestrian route to reach it. Two supervisors will tell you it is a nuisance. Nobody has counted the crossings, so nothing has ever been proposed and nothing has ever been costed.

What it looks like

The authoritative layout is a PDF or a printed general arrangement drawing on the wall
Nobody can say how far material travels per unit, or what a metre of travel costs
Layout changes are argued from experience and settled by seniority
The drawing was last reconciled with the floor at commissioning, or at the last major project

Diagnostic signals you can check this week

Ask for the current layout in a machine-readable form. If you receive a PDF or a paper print, you are here
Ask what the plant's material travel distance per unit is. At stage 1 there is no answer, not even a rough one
Walk the busiest route with a supervisor and count crossings and reversals — then ask whether anyone has ever counted them before
Check when the general arrangement drawing was last reconciled against the floor; commissioning is the usual answer

Anti-pattern · Buying layout software first

The instinctive fix is a layout or simulation tool, on the theory that the tool will produce the answer. It will not, because every one of them consumes two inputs the plant does not yet have: an accurate machine-readable geometry and a real from-to flow matrix. The tool then gets fed the 2014 drawing and last year's product mix, produces a beautiful result, and the result is quietly ignored by everyone who knows the floor. Spend the first quarter producing the two inputs; the tool question answers itself afterwards, and the shortlist will be shorter and cheaper than the one you would have bought from.

What holds you here

There is no queryable model of the floor and no measured flow, so every layout proposal is an opinion and no result can be attributed afterwards.

Highest-leverage next move

Build the two inputs for one hall: a machine-readable as-built geometry, and a from-to material-flow matrix derived from MES and WMS transactions rather than from memory.

Cost of leaving

Effort: 2–4 months
Team: One industrial engineer, one data engineer, part-time; a day of surveying per hall
Risk: Low — nothing in production changes, and the outputs are useful even if the programme stops
To next stage: 2–4 months

If this is you, the next step is

A two-week exercise: survey the geometry, extract the flow, price a metre of travel.

Get one hall measured

Stage 2

Measured

36% of operators sit here

Material flow is quantified from transactions and telemetry, so layout arguments now carry evidence — but the plant still generates its options by hand.

Stage 2 is where the conversation changes character. Once a from-to matrix exists, the question stops being whether the staging area is a nuisance and becomes how many pallet-metres a year it costs, which is a question with an answer. Proposals start carrying evidence, and the evidence usually surprises people: the route everyone complains about is rarely the most expensive one, because complaint tracks visibility and cost tracks volume.

The limitation is that measurement produces an argument, not a search. An engineer can hold perhaps three candidate layouts in mind and compare them fairly; the plant has thousands of feasible arrangements, and the good ones are frequently counter-intuitive because they trade a longer path for a shorter queue. So a stage-2 plant reliably improves on where it started and reliably stops well short of what the same floor could do, and nobody knows by how much, because no baseline of what was possible was ever computed.

This is the mode of the industry and the plateau worth naming. Measurement is genuinely valuable and genuinely insufficient, and because it feels like progress it can absorb years. The signature of a plant stuck here is a shelf of good flow studies and a layout that has changed twice in a decade, both times for reasons unrelated to any of them.

In practice

The study that proved the wrong route was the problem

A speciality chemicals site ran a six-week flow study after complaints about tanker-to-blender transfers blocking the main aisle. The transfers turned out to be 4% of the site's total handling metres. The dominant cost was drum movements between the blending hall and the finishing store, invisible because they happened on night shift when the aisle was empty. The study was correct, the complaint was correct, and the two had almost nothing to do with each other.

What it looks like

A from-to flow matrix exists, built from MES or WMS transaction history
Spaghetti diagrams and travel-distance figures appear in layout proposals
Someone can state travel distance and handling touches per unit for the main product families
Candidate layouts are still drawn by an engineer and compared two or three at a time

Diagnostic signals you can check this week

Ask to see the from-to matrix and check what period and product mix it covers; anything older than the last major SKU change is describing a plant that no longer exists
Ask how many candidate layouts were compared in the last change proposal. Two or three means the options were drawn, not generated
Check whether the flow figures came from MES/WMS transactions or from a manual observation week
Ask what a metre of travel costs on the busiest route. A stage-2 plant usually has a distance but not a cost

Anti-pattern · Treating the spaghetti diagram as the answer

A spaghetti diagram is an excellent instrument of persuasion and a poor instrument of design: it shows where flow is tangled without showing which untangling is worth the disruption. Plants that stop here tend to make the visually obvious change — straighten the route in the picture — and find the benefit smaller than expected, because the tangle was absorbing variability that now surfaces somewhere else. Convert the diagram into a costed from-to matrix and a candidate-generation problem before you convert it into a move.

What holds you here

Options are drawn by hand, so the plant compares three candidates out of thousands and cannot know what it is leaving on the floor.

Highest-leverage next move

Build a discrete-event simulation of one hall, calibrate it until it reproduces last quarter's actual throughput, and encode the hard constraints — clearance, egress, zoning, services — as rules the model enforces.

Cost of leaving

Effort: 4–8 months
Team: Industrial engineer, simulation engineer, plant engineering sponsor
Risk: Low to medium — the work is analytical, but the first simulation will expose disagreements about how the plant really runs
To next stage: 4–8 months

If this is you, the next step is

We build the discrete-event model, calibrate it against last quarter's actuals, and show you the fidelity gap.

Turn the flow study into a model

Stage 3

Simulated

22% of operators sit here

A calibrated model of the plant exists, so candidate layouts are evaluated rather than argued — and the hard constraints are written down as rules.

Stage 3 is the first stage at which the plant can be wrong cheaply. A calibrated model lets a proposal fail in an afternoon instead of failing in concrete, and the value of that is easy to underestimate because the avoided cost never appears in any account. The discipline that makes it real is calibration: a model that has never been made to reproduce a known period is a drawing that moves, and it will be believed more than it deserves precisely because it moves.

The second half of stage 3 is constraint encoding, and it is the half most plants skip. Every plant runs on rules that live in two or three long-serving engineers: this aisle must stay clear because it is the fire route, that wall is the boundary between high-care and low-care, this bay cannot take a pallet stack because of the floor loading, that machine needs three metres of maintenance access on the drive side. Until those are written as machine-checkable rules, no candidate generator can be trusted, because its first genuinely creative suggestion will be one that violates a rule nobody told it about.

The constraint that emerges here is throughput of a different kind: the plant can now evaluate a candidate layout in an afternoon but can still only invent three of them. Simulation without generation is a much better version of stage 2 — faster, fairer, better evidenced — and it is where a great many capable plants settle, because it is genuinely comfortable and it produces defensible answers to the questions somebody thought to ask.

In practice

The model that finally settled the buffer argument

A pharmaceutical packaging site had argued for two years about whether the inter-line buffer before the cartoner was too small. The calibrated model showed that doubling it improved throughput by almost nothing, because the real constraint was changeover sequencing upstream — but that halving it freed floor space with no throughput penalty at all. The buffer shrank, the argument ended, and the freed bay took a second labeller the following year.

What it looks like

A discrete-event model reproduces last quarter's actual throughput within a stated tolerance
Clearance, egress, zoning and services constraints exist as checkable rules, not as engineering memory
Proposals arrive with a simulation run attached, against the current product mix
The model has a named owner and a recalibration cadence tied to mix changes

Diagnostic signals you can check this week

Ask what the model's fidelity gap was on the last release — simulated throughput minus actual, same period and mix. If nobody has computed it, the model is uncalibrated
Ask to see the constraint rules. If the answer is a design-standard document rather than something the model enforces, they are not encoded
Count how many candidate layouts were evaluated in the last study. Under ten means options are still being drawn by hand
Ask what triggers recalibration. Calendar-based recalibration on a plant whose mix changes quarterly is a slow drift into fiction

Anti-pattern · Modelling the plant you designed instead of the plant you run

The most common simulation failure is not technical, it is editorial: the model encodes the standard cycle times, the planned changeover durations and the nominal availability, so it reproduces the plant's design intent rather than its behaviour. It then reports that every proposed change works, because in the designed plant nothing ever starves. Calibrate against a real period including its bad weeks, and publish the fidelity gap alongside every result — a model that is honestly 8% optimistic is far more useful than one that is silently perfect.

What holds you here

The plant can evaluate candidates but not generate them, so the search space explored is whatever three arrangements an engineer thought of.

Highest-leverage next move

Put a candidate generator in front of the simulation — a solver with a learned surrogate to screen arrangements — and make the encoded constraint set a hard filter that runs before any human sees a candidate.

Cost of leaving

Effort: 6–12 months
Team: Simulation engineer, industrial engineer, an EHS reviewer for the constraint set, plant engineering owner
Risk: Medium — encoding constraints surfaces disagreements about what the rules actually are, which is uncomfortable and necessary
To next stage: 6–12 months

If this is you, the next step is

Two workshops with plant engineering and EHS turn tribal rules into machine-checkable ones.

Encode your constraint set

Stage 4

Optimised

8% of operators sit here

Layout generation is a standing capability: candidates are generated, screened against encoded constraints, certified in simulation, and priced with their disruption.

At stage 4 the marginal cost of asking a layout question collapses. Because generation, screening and certification are a pipeline rather than a project, the plant can afford to ask questions it previously could not justify: what if the second labeller went in the north bay, what if despatch marshalling moved outside, what does the layout look like if the new SKU family is 30% of volume in two years. The character of the conversation shifts from defending a proposal to choosing between certified options, which is a much healthier argument to be having.

The discipline that separates a real stage 4 from an impressive stage 3 is move economics. A generated layout has a benefit and a price, and the price is not the capital cost: it is the capital cost plus the output lost while the move happens, plus requalification or revalidation where the process is regulated, plus the ramp-back period where the line runs below rate while operators relearn it. Plants that price only the capex systematically over-move, discover the disruption afterwards, and spend the following two years unable to get any layout proposal approved.

The remaining constraint is human and structural. Every physical move still queues behind a shutdown window, a capital committee and, where the process is regulated, an authority. That is usually correct rather than a defect: it is the mechanism that keeps a search algorithm from rearranging a plant that people work in. What stage 4 changes is that the queue is now fed by certified, priced, comparable options instead of by whoever made the most persuasive case.

In practice

The shortlist that cost less than the study it replaced

A metals fabricator planning a bay reconfiguration generated and screened several thousand arrangements over a weekend, of which about sixty passed the constraint filter, of which six were simulated in detail. The chosen arrangement was not the one with the lowest travel distance — it was fourth on that measure — but it was the only one in the top ten that could be executed in two shutdown weekends instead of one continuous three-week outage. The move sequence, not the end state, decided it.

What it looks like

Thousands of candidate arrangements are generated and screened per study, not three
Every shortlisted candidate carries a simulation run, a constraint-compliance record and a reversion plan
Move economics include disruption, requalification and ramp-back, not capital cost alone
Realised benefit is measured after the move and fed back into the cost model

Diagnostic signals you can check this week

Ask how many candidates the last study generated and how many survived the constraint screen; the ratio tells you how much of the constraint set is really encoded
Ask to see the disruption line in the last move's business case. If there is only a capex line, the economics are half-built
Ask whether the realised benefit of the last completed move was measured against a comparable unchanged area
Check whether shortlisted candidates carry a move sequence, or only an end state. End-state-only proposals are what turn two weekends into three weeks

Anti-pattern · Optimising the end state and improvising the transition

The generated layout is the destination; the move is the journey, and the journey is where the money goes. Optimising purely for the steady-state flow cost produces arrangements that require the plant to be substantially empty to execute — every machine's new position occupied by the machine that has not moved yet. Model the move sequence as part of the candidate: which assets move in which window, what runs at reduced rate meanwhile, and what the reversion looks like if a window overruns. Arrangements that are 3% worse at rest and executable in a weekend beat the optimum you cannot reach.

What holds you here

Physical moves still queue behind shutdown windows and capital approval, so the value released is bounded by how often the plant can execute rather than by what it can find.

Highest-leverage next move

Move value into the reversible layer: put storage assignment, staging, kitting points and AMR traffic rules under a bounded, continuously re-optimised policy, so the layout improves between shutdowns rather than only at them.

Cost of leaving

Effort: 12–18 months
Team: Optimisation engineer, simulation engineer, plant engineering, finance partner for move economics, EHS reviewer
Risk: Medium to high — the generator will propose things the plant has never considered, and the governance around what it is allowed to propose has to exist first
To next stage: 12–18 months

If this is you, the next step is

We build the disruption and requalification model that turns a capex number into a decision.

Price a move properly

Stage 5

Adaptive

3% of operators sit here

The reversible layer of the layout re-optimises continuously inside a written policy, while physical moves are scheduled from certified evidence.

Stage 5 is narrower than the phrase 'self-optimising factory' suggests, and the narrowness is the point. What re-optimises continuously is the soft layer: where things are put down, which routes vehicles prefer, how staging is allocated, how much buffer sits where. What does not re-optimise continuously is anything bolted to the floor, connected to a service, or inside a regulated boundary. A plant that has genuinely reached stage 5 is precise about which decisions sit inside the policy and which sit outside it, and can show you the document.

The second characteristic is that mobility becomes a specification. Once a plant knows what a move costs and how often the optimum shifts, it starts buying differently: skidded modules instead of grouted plinths, quick-disconnect services, castored benches, drop-in cell frames. This is a slower and more consequential change than any algorithm, because it converts a fraction of the structural layer into the cell layer and a fraction of the cell layer into the soft layer, which is where the compounding value actually is.

Sustaining stage 5 is a governance problem rather than a technical one, and it regresses more easily than it advances. The policy that was correct for last year's mix silently stops being correct; the spatial model drifts as small changes accumulate; the reconciliation audit gets skipped for two quarters during a busy period. The leading indicators are unglamorous — divergence rate, fidelity gap, the age of the policy document — and watching them is the whole job.

In practice

The bounded soft layer

A consumer-goods site re-optimises pallet slotting and AMR route preference for the following week every Sunday night, inside written bounds: no assignment may cross the allergen segregation boundary, no route may use the pedestrian-priority aisle during shift change, and no more than 15% of positions may change in one run so that operators are not relearning the store every week. Anything outside the bounds is a proposal for the Monday engineering meeting, not an action.

What it looks like

Storage assignment, staging and AMR traffic policy re-optimise on a schedule inside stated bounds
The bounds are a versioned document reviewed like code, not settings in a tool
The spatial model is reconciled continuously and its divergence rate is monitored as a leading indicator
Some assets are deliberately specified for mobility — skidded, castored, quick-disconnect — so the cell layer becomes semi-reversible

Diagnostic signals you can check this week

Ask for the policy document that states what may re-optimise and inside what bounds; if it is a settings screen, there is no policy
Ask when the reversion path was last exercised — reverting last week's slotting on a live shift should be a drill, not a theory
Check the model-to-floor divergence rate trend; a rising rate is the earliest warning that the capability is decaying
Ask which assets were specified for mobility in the last capital cycle. If none, the cell layer is still as rigid as it was at stage 3

Anti-pattern · Letting the soft layer churn

Continuous re-optimisation without a change budget produces a plant where nothing is where it was last week. Pick rates fall because operators lose their spatial memory, errors rise, and the measured benefit of the optimisation is consumed by the human cost of the churn — usually while the dashboard reports an improvement, because travel distance fell. Cap the proportion of positions that may change per run, measure pick rate and error rate alongside travel distance, and treat operator relearning as a real cost with a real number.

What holds you here

Sustaining adaptivity is a governance discipline: policy bounds silently expire, the spatial model drifts, and the capability decays without anyone deciding to stop it.

Highest-leverage next move

Treat the policy bounds and the spatial model as versioned, reviewed artefacts with owners and audit cadences — and specify mobility into the next capital cycle so more of the plant can move cheaply.

Cost of leaving

Effort: Continuous
Team: Optimisation owner, plant engineering, EHS, plus a standing review of the policy bounds
Risk: Concentrated — low frequency and high consequence: a policy that drifts out of validity fails on a segregation or safety boundary, not on throughput

If this is you, the next step is

We run your written bounds against real scenarios and the last two quarters of mix change.

Stress-test your policy bounds

Where manufacturing plants actually sit on the ladder

The distribution across the five rungs, why the Measured plateau is the deepest one, and what the published research does and does not support.

Most plants are on the Measured rung, and a large minority have not left Drawn. That is a different distribution from most AI capability curves, and the reason is structural rather than cultural: layout work has an unusually long feedback loop. A forecasting model can be improved forty times in a year; a plant layout might be materially changed twice in a decade, so the organisational learning rate is slow, the people who ran the last reconfiguration have often moved on, and each new study starts closer to the beginning than it should.

Illustrative distribution of manufacturing plants across the five rungs

Illustrative, not measured: a synthesis of published maturity-model stage distributions, the small number of sites recognised for advanced manufacturing practice relative to the industrial base, and industry survey reporting on simulation and digital-twin adoption. Treat the shape as an argument about where the plateau is, not as a census.

Share of plants

31% — 1 · Drawn
36% — 2 · Measured (the plateau)
22% — 3 · Simulated
8% — 4 · Optimised
3% — 5 · Adaptive

Source: Illustrative distribution, synthesised from acatech, World Economic Forum and MHI reporting

Largest layout (QAP) instance size any known method solves exactly

arXiv, QAP deep-RL study

7.5%

Average gap from a strong local-search baseline for a trained deep-RL layout policy, out of sample

arXiv, QAP deep-RL study

8 s

Target cycle time Siemens reports its Amberg line reached after the fix was made in the digital twin

Siemens

It is worth being precise about what the research supports, because layout is a field where vendor language runs a long way ahead of published results. The deep-reinforcement-learning study cited above reports solutions that are, out of sample, on average within 7.5% of a high-quality local search baseline (opens in a new tab), beating it on roughly 1.2% of instances. That is a real and useful result — a trained policy generalises to new instances without retraining, which local search cannot do — and it is emphatically not evidence that learning-based methods dominate classical optimisation on layout problems. Earlier work combining genetic search with simulation for unequal-area layouts under stochastic flow (opens in a new tab) makes the same point from the other direction: the hard part is representing the problem faithfully, not choosing the search algorithm.

No methods are known to exactly solve QAP instances of size greater than 30.
Solving the Quadratic Assignment Problem using Deep Reinforcement Learning, arXiv

For a view of what advanced practice looks like at the top of the distribution, the World Economic Forum's Global Lighthouse Network (opens in a new tab) profiles sites recognised for deploying advanced manufacturing technology at scale, and MHI's annual industry report (opens in a new tab) tracks adoption of simulation, robotics and material-handling automation across the wider industrial base. The gap between the two is the honest picture: a small number of exemplary sites, a wide middle that has measured its flow and stopped there, and a large tail still working from drawings.

Sort layout changes by reversibility before you sort them by value

Three horizons — soft, cell and structural — with different approvers, different lead times, and radically different costs of being wrong.

The single most useful thing a plant can do before it buys any layout technology is to sort its layout decisions into three horizons by reversibility, because reversibility determines who may approve a change, how quickly it can happen, and how much evidence it needs to carry. A soft change — which pallet goes where, how staging is allocated, which route a vehicle prefers — can be made on Monday and undone on Tuesday. A structural change is the next decade of flow, poured into a floor slab. Treating both as 'layout' and running both through the same process is why plants either move too slowly or move too rashly, and frequently both at once in different halls.

Horizon	What changes	Who signs it off	Lead time	Cost of being wrong	What AI should be allowed to do
Soft — flow rules	Storage assignment and slotting, staging and marshalling allocation, kitting points, AMR and forklift traffic rules, one-way aisles, buffer sizing, line-to-cell allocation	Area or shift supervisor, under a standing written policy	Hours to days	A shift of throughput; fully reversible	Re-optimise continuously inside stated bounds, with a one-instruction reversion
Cell — moveable assets	Moving a machine or workstation, changing cell orientation, re-racking, re-routing conveyors and chutes, reconfiguring a packing line, relocating a kitting bench	Plant engineering plus EHS, through change control	Weeks, and usually a shutdown window	Weeks of disrupted output plus services rework; partly reversible at a price	Generate and rank candidates; simulate the move sequence, not only the end state
Structural — the building	Walls and fire compartments, dock count and orientation, drainage, cleanroom grading, crane rails, floor loading, utility mains, roof plant	Capital committee, building control, and where regulated the competent authority	Quarters to years	The next decade of flow; effectively irreversible	Evaluate long-horizon scenarios under demand uncertainty — advisory only, never automated

The three layout horizons. The right-hand column is the operating rule that follows from the rest of the row: AI runs the soft horizon under a written policy, proposes into the cell horizon, and only ever advises on the structural one.

Which change should you make now?

Plot the change you are considering against two axes: how reversible it is, and how strong the flow evidence behind it is. Three of the four answers are not 'run the optimiser', and the bottom-left quadrant is where most regretted capital goes.

Just try it, then measure

Reversible change, weak evidence
The cheapest experiment a plant can run
One shift, one area, one comparable control area

Run it now

Reversible change, measured evidence
Slotting, staging, traffic rules, buffer sizes
Governs by written policy, not by approval queue

The capex trap

Permanent change, weak evidence
Where a decade of flow cost gets committed on a hunch
Buy evidence first — it is orders of magnitude cheaper

Simulate, stage, schedule

Permanent change, measured evidence
Certify in simulation and screen against constraints
Optimise the move sequence, not only the end state

Reversibility of the change — top: Soft, undone within a shift, bottom: Structural, effectively permanent

Strength of the flow evidence — left: Estimated from experience, right: Measured from transactions and telemetry

The strategic implication of the table is that the goal of a layout programme is not a better plant layout. It is to move as much value as possible out of the structural horizon and into the soft one, where it can be re-decided continuously and cheaply. Some of that migration is analytical — realising that a staging rule achieves most of what a wall move was proposed for. A great deal of it is procurement: specifying skidded modules rather than grouted plinths, quick-disconnect services rather than hard-piped ones, castored benches, drop-in cell frames. A plant that has made those specification choices for five years has a structurally different optimisation problem from one that has not, and no algorithm can compensate for the difference.

Where layout optimisation lands in a plant

Receiving, process, packing, internal transport, despatch, services and the building — the decisions worth optimising, the system each one lives in, and the KPI it moves.

Layout value in a non-automotive plant concentrates in seven domains, and each has a natural horizon. A decision is a good first candidate when three things are true: the data behind it already exists in a system you control, the change sits in the soft or cell horizon, and it moves a KPI that a plant manager already reports. The map below is how we scope first and second pieces of work with manufacturers — food and beverage, pharma, chemicals, consumer goods, metals, electronics and industrial equipment alike, because the domains are common even where the processes are not.

Plant domain	Layout decisions worth optimising	System of record	KPI it moves	Horizon
Receiving and raw store	Dock allocation, put-away and slotting by velocity and mix, inbound staging lanes, decant point placement	WMS / ERP	Dock-to-line time, MHE hours per 1,000 units	Soft
Process and make area	Cell arrangement and adjacency, vessel-to-filler routing, inter-stage buffer sizing, CIP and changeover staging	MES / process historian	Throughput per hour, blocked and starved time	Cell
Packing and finishing	Line balance and station spacing, changeover staging, case-packer and palletiser placement, reject-loop routing	MES / WMS	OEE performance loss, changeover minutes	Cell
Internal transport	AMR and AGV route network, one-way aisles, crossing and priority policy, charging and parking positions	AMR fleet manager / RTLS	Travel distance per unit, congestion wait, near-miss count	Soft
Finished goods and despatch	Pallet slotting, pick-face arrangement, marshalling lane allocation, loading-dock assignment	WMS	Picks per labour hour, order lead time	Soft
Utilities, services and maintenance	Service and compressed-air drop positions, spares store location, maintenance access clearance, lifting points	CMMS / EAM	Mean time to repair, permit and access time	Cell to structural
Site and building	Dock count and orientation, wall lines and fire compartments, cleanroom grading, drainage and floor loading	CAD / BIM and the capital plan	Throughput per square metre, capex per unit of capacity	Structural

The plant layout decision landscape. 'Horizon' is the reversibility class from the previous section, and it is the field that determines the approval path and the evidence required — not the size of the prize.

Internal transport and the two storage domains are where nearly every plant should begin, and the reason is not that the prize is largest — often it is not — but that the whole loop closes inside one quarter. The data is in the WMS and the fleet manager, the change is a rule rather than a move, the reversion is an instruction, and the KPI moves inside a week. That produces the one thing a layout programme most needs and most rarely has: a measured, attributable result early enough to fund the harder work in the process and packing halls.

The systems the layout data actually lives in
Layout work is unusual in that it spans the whole stack: geometry in CAD or BIM, work centres and routings in the MES, storage locations in the WMS, asset footprints and access requirements in the CMMS, and movement telemetry in the fleet manager or RTLS. The integration model that keeps this coherent is the standard one — ISA-95 (opens in a new tab) for the enterprise-to-control hierarchy and the definition of a work centre, OPC UA (opens in a new tab) for equipment-level information exchange, and the MESA model (opens in a new tab) for how operations functions relate. You do not need a new architecture for layout work; you need the coordinate mapping between the work-centre identifiers in your MES and the geometry in your spatial model, and that mapping is usually the single most valuable half-day of the project.
The KPI definitions that stop the argument
Layout benefits are contested more than most because several teams have a legitimate claim on the same minutes. Adopting a standard KPI definition set — ISO 22400-2 (opens in a new tab) defines KPIs for manufacturing operations management, including the components of OEE and the time model beneath them — removes an entire class of dispute about whether a saved minute was availability, performance or simply a shorter walk. Agree the definitions before the first study, not while reviewing its results.
The constraints that are not negotiable
Aisle and passageway clearance is a regulated matter in most jurisdictions — in the United States, for example, OSHA's materials handling and storage rule (opens in a new tab) requires sufficient safe clearances for aisles used by mechanical handling equipment and that aisles be kept clear and in good repair. In the European Union the Machinery Regulation (EU) 2023/1230 governs the safety requirements of machinery and assemblies, which is what a reconfigured production line becomes. Regulated process boundaries add more: cleanroom grading in pharmaceutical manufacture, high-care and low-care segregation and allergen separation in food, and ATEX zone boundaries in chemicals and dust-handling. All of these belong in the constraint filter, and moving equipment across one of them typically triggers requalification that dwarfs the move cost.
Where the reference work is published
For teams building the measurement and modelling side, NIST's manufacturing programme (opens in a new tab) and its smart manufacturing systems design and analysis work (opens in a new tab) are the most useful public reference points on system modelling, performance measurement and simulation interoperability, and they are vendor-neutral in a field where most published material is not.

What the transitions look like in public

Three publicly reported programmes, read against the ladder. None is an Atomic Loops engagement — each links to the operator's own published material.

The most instructive public examples are not the ones with the biggest numbers; they are the ones that show which artefact the operator invested in. In each case below the differentiator was a representation of the plant — a simulated line, a scanned building, a decades-old modelling capability — rather than a particular algorithm. That is the pattern worth copying, because algorithms are procurable and representations are not.

Three programmes read against the layout ladder

Outcomes as reported by the operators themselves; verify any figure against the linked source before reusing it, as we have not independently audited them. The card images are generated industry scenes from our media library, not photographs of the named sites, and no operator endorsement is implied.

Generated industry scene: an electronics production line with automated placement machines and inline test stations

SiemensElectronics manufacturing · Amberg Electronics Works, Germany34

Challenge: Designing a production line for Simatic controller components on a floor that Siemens reports carries roughly 1,200 different products and about 350 production changeovers a day — a mix profile under which a line designed for the average product is wrong for nearly every actual one.
Approach: The line was built in a digital twin before it was built on the floor, and the twin was used to identify which machine modules were limiting the design rather than accepting the first simulated result as the achievable rate.
Reported outcome: Siemens reports that the initial simulation promised an eleven-second cycle time, that analysis attributed the shortfall largely to machine modules that were planned for the line but did not work optimally, and that after those components were replaced in the digital twin the target cycle time of eight seconds was achieved.; What it shows about the curveThis is the Simulated-to-Optimised transition in one artefact. The value came from being able to be wrong in the model — the first simulated design was rejected, not shipped — which is only possible when the model is trusted enough to overrule an existing plan.

Siemens — Electronics, Digital Enterprise and future technologies (opens in a new tab)

Generated industry scene: a consumer-goods production hall with filling and packing lines, viewed as a connected floor layout

HenkelConsumer goods · laundry, home care and adhesives plants23

Challenge: Building a usable representation of factories that already existed, across a network of sites, without the luxury of designing them digitally from scratch.
Approach: Henkel describes creating a virtual factory by scanning the plant and connecting it — a complete copy of all assembly lines and rooms, published as cloud-based 3D replicas that simulate operations and surface optimal process parameters to operators, with sensor data feeding the model.
Reported outcome: Henkel reports digital twins in use as a working representation of its production facilities, including its Somat production at the Kruševac plant, and positions them as the basis for optimising processes and detecting damage or errors early.; What it shows about the curveThis is the artefact most layout programmes skip. Reality capture of an existing building is how a brownfield plant gets a spatial model that matches the floor rather than the commissioning drawing — the load-bearing input for everything above it on the ladder.

Henkel — Digital twins are paving the way for the factory of the future (opens in a new tab)

Generated industry scene: a high-speed consumer-goods packing hall with palletising and marshalling areas

Procter & GambleConsumer goods · global manufacturing and supply network34

Challenge: Sustaining a modelling and simulation capability across a large manufacturing network for long enough that it becomes an institutional asset rather than a series of consultancy studies.
Approach: P&G describes a long-running collaboration with the University of Cincinnati that began as a computer modelling lab used to improve manufacturing processes and expanded, over fifteen years and roughly 500 students, into a broader digital capability spanning manufacturing digital twins and supply-network modelling.
Reported outcome: P&G reports the partnership now covers manufacturing digital twins and supply-network modelling aimed at an efficient yet resilient product supply, with the modelling work embedded in how the business solves operational problems.; What it shows about the curveThe Optimised rung is a capability, not a project. The distinguishing feature here is duration and continuity — models, people and conventions that persist between studies — which is precisely what a plant loses when it buys each layout study from a different supplier.

P&G — Digital Accelerator partnership with the University of Cincinnati (opens in a new tab)

The reference architecture for a layout capability

Six layers, each annotated with the rung that first requires it — and the one layer that everybody defers and shouldn't.

A layout optimisation capability needs six layers, and the order in which they are built decides whether the plant ends up with an asset or with a stack of reports. The architecture below is deliberately vendor-neutral: every layer is defined by what it must guarantee rather than by what product provides it, because the products change every few years and the guarantees do not. Note where the rung annotations fall — three of the six layers are already required before any optimiser appears.

Layers required by rung

A programme buying generation and certification without the spatial record, flow record and constraint set is buying an engine with no fuel, no map and no brakes. The reconciliation layer is the one most often deferred and the one whose absence decays everything above it.

Spatial record
Stage 1+
- As-built geometrySurvey or reality capture, not the commissioning drawing
- Asset register with envelopesFootprint, maintenance access, service connections
- Zone and boundary mapFire compartments, hygiene zones, cleanroom grades, ATEX zones
→
Flow record
Stage 2+
- Transaction extractMES and WMS movements, mapped to spatial coordinates
- Movement telemetryRTLS, AMR and forklift logs — the routes actually driven
- From-to matrixSegmented by product family and shift, not averaged
→
Cost model
Stage 2+
- Cost per metre movedMHE hours, labour, energy, by route and vehicle type
- Space costWhat a square metre of production floor is worth in this hall
- Disruption modelLost output, requalification, retraining, ramp-back
→
Constraint set
Stage 3+
- Safety and egress rulesAisle clearance, escape routes, vehicle-pedestrian separation
- Process and hygiene rulesSegregation, grading, zone boundaries, cross-contamination
- Structural and services limitsFloor loading, crane envelopes, drop positions, headroom
→
Generation and certification
Stage 3+
- Candidate generatorSolver plus learned surrogate; thousands of arrangements, cheaply scored
- Constraint screenHard filter applied before scoring; every rejection logged
- Discrete-event simulationCertification against the current mix, with a published fidelity gap
→
Execution and reconciliation
Stage 4+
- Move sequencingWhich asset moves in which window, and what runs meanwhile
- Reversion pathThe previous arrangement or rule, one instruction away
- As-built reconciliationModel updated from the work order; divergence sampled and tracked

Pipeline described

Spatial record (stage 1+) — As-built geometry: Survey or reality capture, not the commissioning drawing; Asset register with envelopes: Footprint, maintenance access, service connections; Zone and boundary map: Fire compartments, hygiene zones, cleanroom grades, ATEX zones
Flow record (stage 2+) — Transaction extract: MES and WMS movements, mapped to spatial coordinates; Movement telemetry: RTLS, AMR and forklift logs — the routes actually driven; From-to matrix: Segmented by product family and shift, not averaged
Cost model (stage 2+) — Cost per metre moved: MHE hours, labour, energy, by route and vehicle type; Space cost: What a square metre of production floor is worth in this hall; Disruption model: Lost output, requalification, retraining, ramp-back
Constraint set (stage 3+) — Safety and egress rules: Aisle clearance, escape routes, vehicle-pedestrian separation; Process and hygiene rules: Segregation, grading, zone boundaries, cross-contamination; Structural and services limits: Floor loading, crane envelopes, drop positions, headroom
Generation and certification (stage 3+) — Candidate generator: Solver plus learned surrogate; thousands of arrangements, cheaply scored; Constraint screen: Hard filter applied before scoring; every rejection logged; Discrete-event simulation: Certification against the current mix, with a published fidelity gap
Execution and reconciliation (stage 4+) — Move sequencing: Which asset moves in which window, and what runs meanwhile; Reversion path: The previous arrangement or rule, one instruction away; As-built reconciliation: Model updated from the work order; divergence sampled and tracked

Step-by-step insights

Spatial record — capture reality, not intent: The commissioning drawing describes the plant somebody intended to build; the floor describes the plant that exists. On a brownfield site the gap is usually years of accumulated small changes, and it is not recoverable by reading the drawing more carefully. Reality capture — laser scanning, photogrammetry or a disciplined survey — produces a base model that is true on the day it is taken, which is the only kind of base model worth having. The zone and boundary map deserves separate mention because it is the part most often left implicit: fire compartments, hygiene zones and cleanroom grade boundaries are invisible in geometry but decisive in feasibility, and a model without them will generate candidates that are geometrically perfect and operationally illegal.
Flow record — the join is the work: The movements are already recorded; nobody has to instrument anything new to get a first flow matrix. What has to be built is the join: MES work-centre identifiers and WMS storage locations mapped to coordinates in the spatial model, so a transaction becomes a vector rather than a row. Expect the mapping to expose surprises — locations that exist in the WMS and not on the floor, work centres that share a physical position, movements recorded against a hall rather than a point. Those surprises are themselves findings. Add telemetry second: RTLS and fleet logs describe the routes actually driven, which is how you learn that the shortest path and the used path differ, usually for a reason worth knowing.
Cost model — a metre is not a metre: Distance is the wrong unit for a decision because a metre travelled by a tugger at three in the morning on an empty aisle is not the same cost as a metre travelled by a counterbalance truck through a congested packing hall at shift change. Build the cost per metre by route and vehicle type, including the labour attached to the movement and the throughput lost when the route is blocked. The disruption model is the other half and the one that changes decisions most: lost output during a move, requalification or revalidation where the process is regulated, operator retraining, and the ramp-back period during which the line runs below rate. Plants that carry only capex in the business case reliably discover this arithmetic afterwards.
Constraint set — write it down before you need it: The constraint set is an act of documentation more than of engineering, and the reason to do it early is that the people who know the rules are a small and shrinking group. Run it as two workshops: plant engineering states the physical rules, EHS and quality state the regulated ones, and the output is a list of checkable predicates rather than prose. Keep the regulated constraints separately and treat them as absolute filters — a candidate that crosses a grade boundary or narrows an escape route is not a low-scoring candidate, it is not a candidate. Log every rejection with the rule that caused it; the rejection log is how you discover that a rule everybody believes in was actually retired three years ago.
Generation and certification — cheap search, expensive proof: The pipeline shape that works is generate-cheaply, screen hard, certify expensively. A solver with a learned surrogate can score thousands of arrangements per minute on a proxy objective; the constraint screen removes the infeasible ones deterministically; discrete-event simulation then spends real compute on the handful that survive, because only simulation sees queueing, blocking, starving and shared-resource coupling. Publish the fidelity gap with every certification — simulated throughput minus actual, same period and mix — because a stated bias is correctable and an unstated one is corrosive. And keep the generator's proxy objective honest by periodically checking its ranking against the simulator's on a sample; when the two diverge, the surrogate has drifted and needs retraining against recent runs.
Execution and reconciliation — where capability compounds or rots: Two components carry disproportionate weight. The reversion path is the political key: plant engineering will approve a soft-layer change that can be undone with one instruction and will not approve one that cannot, and this is entirely rational. As-built reconciliation is the technical key: unless the executed change updates the spatial model, the model decays, and the decay is invisible until a proposal reaches the floor and does not fit. Make the model update a completion criterion of the work order and sample the floor against the model quarterly, so the divergence rate is a number you watch rather than a discovery you make.

The layer most often deferred is the cost model, usually on the grounds that it can be added once there is something to cost. That sequencing is backwards. Without a cost per metre and a disruption estimate, generated candidates can only be ranked on distance, which systematically favours arrangements that are marginally shorter and enormously more disruptive to reach — and the plant learns to distrust the whole output because its first three recommendations were undeliverable.

The four dimensions that set your rung

Layout maturity is not one number. Four dimensions gate each other, and the lowest of them is the real rung.

Layout maturity is not a single number, and treating it as one is how plants end up with an expensive optimiser sitting on a wrong model. A plant is scored on four dimensions — spatial model fidelity, flow evidence, constraint encoding and move economics — and the lowest of the four is the real rung, because each gates the others. Generation built on a drifted spatial model produces confidently undeliverable proposals; a perfect model with no flow evidence has nothing to optimise; encoded constraints without move economics produce feasible arrangements nobody can afford to reach.

Spatial model fidelity
Whether a machine-readable model of the floor exists, and whether it still matches the floor. The binding question is not how the model was made but how it is maintained: a scanned model with no reconciliation process is a very accurate description of one day in the plant's history. The measurable form of this dimension is the model-to-floor divergence rate — the share of sampled spatial assertions that fail a physical check — and plants that have never sampled it invariably guess low.
Flow evidence
Whether material movement is measured from transactions and telemetry rather than described from experience, and whether it is segmented by product family and shift rather than averaged into a single matrix. This is the dimension that most often looks stronger than it is, because a plant with an impressive flow study from two years ago and three SKU launches since then has evidence about a plant that no longer runs.
Constraint encoding
Whether the rules that make an arrangement legal are machine-checkable or resident in two engineers' heads. This is the dimension that decides whether generation can be trusted at all, and it is overwhelmingly the lowest-scoring one in the reviews we run, because encoding constraints is unglamorous documentation work that nobody's targets reward until the first bad proposal makes the case for it.
Move economics
Whether the plant can price a change honestly — capital, lost output, requalification, retraining, ramp-back — and whether it measures what the last change actually delivered. This is the dimension that converts a technically excellent capability into an organisationally trusted one, and its absence is why so many plants have a well-regarded simulation team whose recommendations do not get funded.

Dimension	The question it answers	The artefact that proves it	What a weak score caps
Spatial model fidelity	Does our model of the floor match the floor?	A sampled divergence-rate report with a date on it	Everything — a wrong model makes every candidate confidently wrong
Flow evidence	Do we know what actually moves, and what it costs?	A from-to matrix segmented by mix and shift, plus a cost per metre	Ranking. Without it you can generate arrangements but not prefer one
Constraint encoding	Can a machine tell whether an arrangement is legal?	A checkable rule set plus the rejection log from the last run	Trust. One illegal proposal ends the programme's credibility
Move economics	Can we price a change and prove what the last one delivered?	A disruption model and a realised-versus-forecast record	Funding. Unpriced and unproven changes stop getting approved

Each dimension, the question it answers, the artefact that proves it, and what it caps if it is weak. The right-hand column is the practical reason to score them separately rather than averaging them.

A 90-day plan: cut travel and congestion in one packing hall

The Measured-to-Simulated transition made concrete on one problem — staging assignment and vehicle traffic rules in a single hall, with no machine moved.

Moving one rung takes about 90 days when it is scoped to a single hall and a single reversible decision, and several years when it is scoped to a site. To make that concrete, the plan below runs the transition on a specific and very common problem: a packing hall where pallets from the raw store cross the main vehicle route to reach staging, congestion peaks at changeover, and nobody has ever costed either. Nothing in this quarter moves a machine, requires a shutdown window or spends capital — every change sits in the soft horizon and reverts with one instruction.

Measured to Simulated on one packing hall, in one quarter

One hall, one shift pattern, one named owner. If a phase needs longer than its window, narrow the scope — fewer SKU families, one shift instead of three — rather than extending the plan.

Days 1–15
Make the model match the floor
Survey or scan the hall and reconcile it against the current general arrangement drawing. Record every discrepancy and compute an honest model-to-floor divergence rate for the hall — this number is the baseline for a KPI you will keep. Map the hall's WMS storage locations and MES work centres onto coordinates in the model. Name a plant engineering owner; travel, congestion and near-misses in that hall are their numbers.
A true spatial model and a first divergence rate
Days 16–40
Extract the flow and price a metre
Pull six months of movement transactions from the WMS and MES for the hall, plus fleet-manager or RTLS logs where they exist, and build a from-to matrix segmented by product family and shift. Compute cost per metre by route and vehicle type from MHE hours, labour and blocked throughput. Publish the top ten routes by cost — expect the ranking to differ from the complaint ranking.
A costed from-to matrix for the hall
Days 41–60
Encode the constraints and build the model
Two workshops turn tribal rules into checkable ones: aisle clearance and escape routes, vehicle-pedestrian separation, allergen or hygiene segregation where relevant, floor loading, and the positions that cannot change. Build a discrete-event model of the hall and calibrate it until it reproduces last quarter's actual throughput; publish the fidelity gap rather than tuning until it disappears.
A checkable constraint set and a calibrated model
Days 61–80
Generate, certify and run one shift
Generate staging assignments and vehicle traffic rules against the costed matrix, screen every candidate through the constraint set, and certify the top three in simulation. Run the winner on one shift, in one hall, with the previous assignment one instruction away and the supervisor briefed to revert without asking. Cap the proportion of positions that change so operators are not relearning the hall.
A certified change running under a written reversion
Days 81–90
Attribute against a comparable area
Hold a comparable hall, aisle or shift on the previous rules. Report travel distance per unit, congestion wait, pick rate and error rate — all four, because a travel reduction bought with a pick-rate loss is not a gain. Feed the variance between forecast and realised benefit back into the cost model. This is the number that funds the process hall.
A measured, attributable delta and a corrected cost model

The order matters

Model fidelity before optimisation
A merely adequate optimiser on a true model beats an excellent optimiser on a drifted one, every time and without exception. The first fifteen days are the highest-return fifteen days in the quarter, and they are the ones most often compressed because they produce no visible output.
Reversible before permanent
Everything in this quarter can be undone in a shift. That is what makes it approvable without a capital case and what makes it safe to be wrong. Save the machine moves for the second quarter, when you have a calibrated model, a costed matrix and one measured result to argue from.
Constraints before candidates
Generate nothing until the constraint set is written. The first illegal proposal a plant sees is remembered for years, and it does not matter that the rule was never written down — it will still be the optimiser's fault.
Four KPIs before one
Travel distance alone is the metric most likely to show a false win, because it improves when work is pushed onto operators or when churn is rising. Report travel, congestion wait, pick rate and error rate together from day one, so nobody has to relitigate the result later.

Instrumenting layout KPIs: formula, source, cadence

Where each layout metric actually comes from — the formula, the system that produces it, and the rung at which it first measures something real.

A layout KPI you cannot name a source system for is an opinion with a decimal point. Every metric below reduces to timestamps, counts and coordinates that the MES, WMS, CMMS, fleet manager or the spatial model already holds — the instrumentation work is joining them, not creating them. Two of these metrics are unusual and worth adopting deliberately: model-to-floor divergence, which is the health of the capability itself, and the simulation fidelity gap, which is the honesty of the evidence everything else rests on.

KPI	Formula / read	Source	Cadence	Honest from
Travel distance per unit	Sum of movement distances ÷ units produced, by product family	WMS/MES transactions joined to the spatial model; fleet logs where available	Weekly	Rung 2
Handling touches per unit	Count of discrete pick, place and transfer events ÷ units produced	WMS / MES transactions	Weekly	Rung 2
Cost per metre moved	(MHE hours × rate + attached labour + blocked throughput) ÷ metres travelled	CMMS, payroll and fleet logs joined to the flow matrix	Monthly	Rung 2
Blocked and starved time	Duration in blocked or starved state ÷ scheduled run time, per station	MES / process historian	Per shift	Rung 2
Congestion wait	Vehicle time stationary in an aisle awaiting clearance ÷ total travel time	AMR fleet manager or RTLS telemetry	Daily	Rung 3
Model-to-floor divergence	Sampled spatial assertions failing a physical check ÷ assertions sampled	Audit log against the spatial model	Quarterly	Rung 3
Simulation fidelity gap	Simulated throughput − actual throughput, same period and mix	DES output versus MES actuals	Per model release	Rung 3
Throughput per square metre	Units produced ÷ occupied production floor area	MES joined to the spatial model	Monthly	Rung 3
Constraint-screen pass rate	Candidates passing the automated screen ÷ candidates generated	Generation and rejection log	Per run	Rung 4
Realised versus forecast benefit	Post-move KPI delta ÷ forecast delta, against a comparable control area	MES actuals plus the project record	Per completed move	Rung 4
Soft-layer churn	Positions or rules changed ÷ total positions or rules, per optimisation run	WMS assignment history; fleet policy versions	Per run	Rung 5

Instrumentation build sheet for plant layout work. 'Honest from' is the rung at which the metric starts measuring something real; reporting it earlier produces a number that exists but means nothing.

Two disciplines make the whole sheet trustworthy. First, travel-based metrics are always reported alongside a human-performance metric — pick rate and error rate — because a reduction in travel bought by making the operator's job harder shows up as a win on the distance line and a loss everywhere else. Second, every value claim is measured against a comparable unchanged area rather than against the same area before the change, because product mix, seasonality and staffing all move at the same time as your layout does and will otherwise take the credit or the blame.

Rung 3 readiness checklist

If you cannot tick all eight, you are still on the Measured rung regardless of what software you own. Tick as you go — this list works without JavaScript.

A machine-readable model of the floor exists and has been checked against the floorChecked, with a date and a divergence rate — not assumed.
The from-to flow matrix is derived from transactions, not from observation or memorySegmented by product family and shift, covering the current mix.
You can state the cost of a metre moved on your busiest routeMHE hours, attached labour and blocked throughput, by vehicle type.
Hard constraints are written as machine-checkable rulesClearance, egress, zoning, floor loading, service positions, access envelopes.
Regulated boundaries are hard filters, not weighted penaltiesGrade boundaries, segregation lines and escape routes are not tradeable.
A discrete-event model reproduces a known past period within a stated tolerance
The simulation fidelity gap is published with every resultA stated bias is correctable; an unstated one is corrosive.
Executed changes update the spatial model as part of the work orderThis is what stops the capability decaying between studies.

0 of 8 ticked

Tick honestly — an empty list is a clean starting point

Zero ticks is the Drawn rung, and it is more common than the industry admits. Do not start with tooling. Pick one hall and run the first fifteen days of the 90-day plan above: survey the geometry, reconcile it against the drawing, and write down the divergence rate. Everything else on this list becomes tractable once that exists.

Scope the first hall with us

Failure modes that send a layout programme backwards

Layout maturity is not monotonic. Four regressions account for nearly all of it, and three are invisible until a proposal reaches the floor.

Layout capabilities regress, usually without anyone noticing, because the conditions that made them trustworthy quietly stopped holding. What makes layout regression distinctive is its latency: a drifted spatial model or an expired constraint keeps producing plausible output for months, and the failure surfaces only when somebody tries to build the recommendation. Four patterns account for nearly all of it.

Likelihood: highImpact: high

The model quietly stops matching the floor

Two hundred small undocumented changes a year — a bench moved, a rack added, a conduit rerouted, a bay permanently occupied by totes — and the spatial model becomes a description of a plant that no longer exists. Every candidate generated afterwards is confidently wrong, and the error is only discovered when a proposal is taken to the floor and does not fit.

PreventionMake the model update a completion criterion of the work order, and sample the floor against the model quarterly so divergence is a tracked rate rather than a discovery.

Likelihood: highImpact: medium

The simulation is calibrated to a mix that has gone

The model was calibrated against a quarter whose product mix, changeover pattern and staffing no longer apply. It keeps producing confident throughput figures, and the fidelity gap widens silently because nobody recomputes it outside a release.

PreventionKey recalibration to mix events — new SKU families, line reallocations, shift-pattern changes — rather than to the calendar, and recompute the fidelity gap on every release.

Likelihood: mediumImpact: high

A rule that was never written down gets broken

The generator proposes an arrangement that crosses a hygiene boundary, narrows an escape route or blocks a maintenance access envelope, because that rule lived in an engineer who has since retired. One such proposal typically ends the programme's credibility regardless of the quality of everything else it produced.

PreventionTreat the constraint set as a living document with an owner and an annual review, and log every screen rejection with the rule that caused it so retired and missing rules both surface.

Likelihood: mediumImpact: high

The move is priced on capex and delivered into disruption

A change is approved on capital cost, executed, and the plant then absorbs weeks of lost output, an unbudgeted requalification and a long ramp-back. The realised benefit never gets measured against the forecast, so the organisation learns only that layout projects hurt — and the next proposal, however well evidenced, does not get funded.

PreventionCarry disruption, requalification and ramp-back in the same figure as capital, and measure realised versus forecast benefit against a comparable unchanged area after every completed move.

Glossary

Hover a term for its definition — or expand the map full screen. The full definitions are written out below.

Facility layout problem (FLP): The formal problem of assigning departments, machines or cells to locations so that a cost — usually the sum of material flow multiplied by distance — is minimised, subject to space and adjacency constraints. Plant layout optimisation is an applied instance of it.
Quadratic assignment problem (QAP): The classical mathematical formulation of the facility layout problem, due to Koopmans and Beckmann. It is NP-hard, and no known method solves instances larger than about 30 locations exactly, which is why real layout tools return ranked shortlists rather than proven optima.
From-to matrix: A table of material movement volumes between every pair of locations in a plant, ideally derived from MES and WMS transactions and segmented by product family and shift. It is the primary input to any layout optimisation and the artefact most often missing.
Spaghetti diagram: A drawing of the paths material or people actually take through a facility, overlaid on the layout. Excellent for persuasion and poor for design, because it shows where flow is tangled without showing which untangling is worth the disruption.
Model-to-floor divergence: The share of sampled spatial assertions in a plant's digital model that fail a physical check on the floor. The health metric of a layout capability: as it rises, every generated candidate becomes less deliverable, and it rises silently.
As-built reconciliation: The practice of updating the spatial model from executed work so that the model continues to match the floor. Best implemented as a completion criterion of the work order rather than as a separate task somebody must remember.
Soft layout: The reversible layer of a plant's arrangement — storage assignment and slotting, staging and marshalling allocation, kitting points, vehicle traffic rules and buffer sizes — which can be changed within a shift and undone within a shift. The only layer AI should run continuously.
Constraint screen: A deterministic filter applied to generated candidate layouts before scoring, removing any arrangement that breaches clearance, egress, zoning, loading or access rules. Regulated boundaries belong here as hard filters, never in the objective function as weighted penalties.
Discrete-event simulation (DES): A model that advances through discrete state changes — arrivals, starts, completions, blockages — to reproduce a system's dynamic behaviour. In layout work it is what certifies a candidate, because it sees queueing, blocking, starving and shared-resource coupling that a flow-times-distance score cannot.
Fidelity gap: Simulated throughput minus actual throughput for the same period and product mix. Published alongside every simulation result, it makes a model's bias correctable; left uncomputed, it makes the model's outputs quietly untrustworthy.
Cost per metre moved: The cost of moving material one metre on a given route, including material-handling equipment time, the labour attached to the movement and the throughput lost when the route is blocked. It converts distance into a unit that can be traded against capital and floor space.
Ramp-back: The period after a layout change during which a line runs below its rated output while operators relearn the arrangement and residual problems are cleared. A real and routinely unbudgeted component of move economics, and often larger than the physical downtime.

Frequently asked questions

The questions plant engineering and industrial engineering teams ask most often when scoping layout optimisation work.

Can AI design a factory layout on its own?

No, and the limitation is mathematical rather than commercial. The classical formulation of plant layout is the quadratic assignment problem, which is NP-hard; a 2023 deep-reinforcement-learning study of it notes that no method is known to solve instances larger than about 30 locations exactly. Every real plant is bigger than that. What AI does well is generate and cheaply rank thousands of candidate arrangements so that simulation and human review can be spent on the handful worth examining. The output is a certified shortlist, never a proven optimum.

What data do we actually need before starting?

Three artefacts, in this order. A machine-readable model of the floor that matches the floor — surveyed or scanned, not the commissioning drawing. A from-to material-flow matrix derived from MES and WMS transactions and segmented by product family and shift. And a cost per metre moved, built from material-handling equipment hours, attached labour and blocked throughput. Given those three, several different optimisation approaches work adequately. Missing any one of them, none of them works, because the tool will confidently optimise a plant that does not exist or a flow that is not representative.

How long does it take to move from Measured to Simulated?

About 90 days when scoped to one hall and one reversible decision, and several years when scoped to a whole site. The work in that quarter is reconciliation, extraction, constraint encoding and calibration — not model development — because a Measured plant already has the flow data it needs. The most common way to turn 90 days into two years is to start with the process hall rather than a storage or transport area, because process-hall changes almost always sit in the cell horizon and inherit a shutdown window and a change-control queue.

Do we need a digital twin of the whole plant?

No, and the whole-plant framing is a common way to spend a year without a result. Start with one hall: its geometry, its movements, its constraints and a discrete-event model calibrated against a known past quarter. A hall-level model produces decisions inside a quarter and teaches you what your data is actually like. Whole-site twins are worth building once two or three hall models exist and share conventions, at which point the site model is largely composition rather than a new project.

How do we stop the optimiser proposing something unsafe?

By encoding constraints as filters applied before scoring, not as penalties inside the objective function. Aisle clearance, escape routes, vehicle-pedestrian separation, hygiene and allergen segregation, cleanroom grade boundaries, ATEX zone edges, floor loading and maintenance access envelopes are not tradeable at any score. Write them as checkable rules, run every generated candidate through them before a person sees it, and log each rejection with the rule that caused it. The rejection log is also how you discover rules that were retired years ago and are still constraining your options.

What is the difference between layout optimisation and slotting optimisation?

Slotting is a subset — the reversible one. Slotting decides which item occupies which storage position and can be changed and undone within a shift, which is why it is the right place to start and the only layer that should re-optimise continuously under a written policy. Layout optimisation also covers cell arrangement, machine placement, buffer sizing, aisle direction and building geometry, which sit in the cell and structural horizons with far longer lead times, far higher costs of being wrong, and approval paths through engineering, EHS and capital committees.

How do we measure whether a layout change actually worked?

Against a comparable unchanged area, not against the same area before the change. Product mix, seasonality, staffing and demand all move at the same time as your layout does, and they will otherwise take the credit or the blame. Report four numbers together — travel distance per unit, congestion wait, pick or throughput rate, and error rate — because a travel reduction bought by making an operator's job harder shows up as a win on the distance line and a loss everywhere else. Then compare realised benefit against the forecast and feed the variance back into the cost model.

Our plant is regulated. Does that rule out AI layout optimisation?

No, but it changes what the optimiser is allowed to consider. In pharmaceutical manufacture, cleanroom grade boundaries and qualified equipment positions constrain which moves are feasible and impose requalification costs that frequently exceed the physical move; in food, high-care and low-care segregation and allergen separation do the same. Encode those boundaries as hard filters and carry requalification in the move economics. What optimisation then does very well in regulated plants is find the valuable changes that stay entirely inside a zone — staging, sequencing, buffer sizing and transport routing — which are numerous and usually under-exploited.

How much of a plant's layout can realistically be automated?

The reversible layer, and only inside written bounds. Storage assignment, staging allocation, kitting points, buffer sizes and vehicle traffic rules can re-optimise on a schedule with a one-instruction reversion, and mature plants cap how much may change per run so operators are not relearning the floor weekly. Anything bolted to the floor, connected to a service or inside a regulated boundary goes through certification, engineering and EHS review and a shutdown window. That division is not a technology limitation to be overcome; it is the mechanism that makes the whole capability approvable.

What does a layout optimisation programme cost in team terms?

For the first hall, roughly one industrial engineer and one simulation or data engineer for a quarter, plus a day of surveying per hall and meaningful time from a plant engineering owner and an EHS reviewer for the constraint workshops. The dominant cost is rarely the software. It is the reconciliation of the spatial model and the mapping of MES work centres and WMS locations onto coordinates, which is unglamorous, hard to parallelise, and the reason picking a hall whose data you already control is the single biggest lever on the timeline.

How often should a plant re-run its layout optimisation?

The soft layer on a weekly or monthly cycle, bounded by a churn cap; the cell layer whenever a shutdown window is being planned or the product mix has materially shifted; the structural layer only inside a capital planning cycle. The trigger that matters more than any calendar is mix change: new SKU families, a line reallocation or a volume shift between product families invalidates a from-to matrix quickly, and a plant running last year's matrix is optimising a factory it no longer operates.

Where does layout work sit relative to our MES and ERP roadmap?

Alongside it, not behind it. Layout optimisation consumes data those systems already produce — transactions, work-centre definitions, storage locations, maintenance records — and writes back into the WMS and the fleet manager rather than into the MES core, so it rarely competes for the same change windows. The one dependency worth sequencing deliberately is the identifier mapping: work-centre and location identifiers must be stable enough to join to coordinates, so a major MES or WMS re-implementation is a reason to agree the mapping early, not a reason to wait.

About the author

Atomic Loops Engineering

Industrial AI practice

Atomic Loops builds production AI systems for manufacturing, logistics and energy operators — forecasting, scheduling, optimisation and vision inspection running against live plant data and integrated into the MES, WMS, CMMS and fleet-control layer rather than delivered as dashboards or slide decks.

· Optimisation and simulation work delivered into live plant estates
· Layout and flow studies run jointly with plant engineering and EHS teams
· Integration-first delivery: MES/WMS write-back, monitoring, reversion plans
· 16 cited sources on this page

Sources

Readiness and transformation roadmap for manufacturing AIThe phased factory-digitisation sequencing that decides whether the data behind a layout study exists before you need it.AI adoption and the maturity curve in manufacturingWhat separates leading plants from lagging ones across the whole AI portfolio, not just the layout capability.Manufacturing industry practiceProduction AI systems Atomic Loops has built for process, discrete and consumer-goods manufacturers.

Find out what your floor is actually costing you

We sample your spatial model against the floor, extract a real from-to matrix from your MES and WMS, price a metre of travel on your busiest routes, and leave you with a costed 90-day plan for one hall. You keep the plan whether or not we build it.

Book a layout review See our manufacturing work

Published 20 August 2026 · Last updated 20 August 2026

Key takeaways

Abbreviations used on this page

Score your plant on the layout ladder

What AI layout optimisation actually does in a plant

The five rungs of the layout ladder, in detail

Drawn

Measured

Simulated

Optimised

Adaptive

Where manufacturing plants actually sit on the ladder

Sort layout changes by reversibility before you sort them by value

Where layout optimisation lands in a plant

What the transitions look like in public

The reference architecture for a layout capability

Spatial record

Flow record

Cost model

Constraint set

Generation and certification

Execution and reconciliation

The four dimensions that set your rung

A 90-day plan: cut travel and congestion in one packing hall

The order matters

Instrumenting layout KPIs: formula, source, cadence

Failure modes that send a layout programme backwards

The model quietly stops matching the floor

The simulation is calibrated to a mix that has gone

A rule that was never written down gets broken

The move is priced on capex and delivered into disruption

Glossary

Frequently asked questions

Can AI design a factory layout on its own?

What data do we actually need before starting?

How long does it take to move from Measured to Simulated?

Do we need a digital twin of the whole plant?

How do we stop the optimiser proposing something unsafe?

What is the difference between layout optimisation and slotting optimisation?

How do we measure whether a layout change actually worked?

Our plant is regulated. Does that rule out AI layout optimisation?

How much of a plant's layout can realistically be automated?

What does a layout optimisation programme cost in team terms?

How often should a plant re-run its layout optimisation?

Where does layout work sit relative to our MES and ERP roadmap?

About the author

Sources

Related reading

Find out what your floor is actually costing you