Silicon Wafer EngineeringAI-Driven Disruptions & Innovations

Bio-substrates in silicon wafer engineering: where biology can meet the wafer, and what AI has to carry

Bio-substrates in silicon wafer engineering are materials of biological origin joined to silicon devices — cellulose nanofibril film, hydrogels, capture chemistry, DNA templates, living cell layers — admitted only after the front end is finished. None survives high-temperature processing, so every real flow is a boundary problem, and AI's first job is replacing the measurements that boundary destroys.

Silicon Wafer Engineering · AI-Driven Disruptions & Innovations

Atomic Loops EngineeringIndustrial AI practiceReviewed by Atomic Loops Solutions ArchitecturePublished 20 August 202660 min read

Key takeaways

Bio-substrates never enter the front end. Every publicly reported device — CMOS neural probes, cellulose-nanofibril microwave circuits, silicon microfluidics — puts the biological material after the last high-temperature step, usually by transfer or post-passivation deposition. The engineering question is not 'can we make a biological wafer' but 'where in this flow may this material be admitted'.
Two hard walls set that admission point: thermal budget and contamination. Most biological materials stop being themselves far below the ~400 °C a finished back end can still take, and organics, salts and biologicals are excluded from shared toolsets — so admitting one forks the line into a segregated module with its own rules.
The boundary destroys metrology, and that is where AI earns its place first. Monolayers, hydrated films and living layers cannot be measured by the inline metrology a fab already owns, so the honest first model is a virtual sensor with a stated error bound standing in for the measurement — not materials discovery, and not autonomy.
Incoming variance is biological, which means a substrate lot is a harvest rather than a specification. Lot genealogy — source batch, moisture, molecular weight, storage age — is the dataset that decides whether any model generalises, and it is almost always missing at the point people start modelling.
If the device is medical, the model becomes a design input. Biological evaluation under ISO 10993 and design control under ISO 13485 and the EU Medical Device Regulation mean a model that influences release must be versioned, validated and reconstructable — which is a documentation discipline built during the work, never after it.

Abbreviations used on this page

CMOS: Complementary metal-oxide-semiconductor — the standard silicon device process
FEOL: Front end of line — transistor formation; high temperature, high vacuum, implant and anneal
BEOL: Back end of line — interconnect and passivation; typically capped near 400 °C
MEMS: Micro-electromechanical systems — the post-CMOS module most bio-facing devices use
CNF: Cellulose nanofibril — a wood-derived film used as a flexible, biodegradable substrate
SAM: Self-assembled monolayer — the molecular layer most capture chemistry is anchored on
MEA: Microelectrode array — the silicon device a living cell layer is cultured on
VM: Virtual metrology — predicting a measurement from tool trace instead of measuring it
R2R: Run-to-run control — adjusting the next run's recipe from the last run's result
FDC: Fault detection and classification — the tool-trace layer VM is normally built on
MES: Manufacturing execution system — the fab's system of record for lots and routes
TTV: Total thickness variation — the flatness metric a bonded or transferred stack lives or dies by

What bio-substrates mean in a silicon flow

A definition, the two walls that decide where biology may enter, and the split flow every real device uses.

Bio-substrates in silicon wafer engineering are substrate, interface and template materials of biological origin or biological function that are combined with silicon devices — cellulose nanofibril film, silk fibroin, hydrogels, self-assembled capture chemistry, DNA templates and, at the far end, living cell layers cultured on a microelectrode array. In every publicly reported device, they are joined to the wafer after the front end is complete, because none of them survives what the front end does.

That constraint is not a limitation to be engineered away; it is the shape of the problem. A silicon front end runs implants and anneals well above 1,000 °C, plasma etches, high vacuum and aggressive wet chemistry. A back end still reaches roughly 400 °C. Almost every material on the list above stops being itself somewhere below 200 °C, and a living layer has to stay near 37 °C and wet. So the practical question is never 'can we make a biological wafer'. It is: where in this flow may this material be admitted — and what does admitting it cost the rest of the line?

Bio-substrate as carrier
The biological material replaces the handle the finished device sits on — a wood-derived cellulose nanofibril film instead of a rigid, non-degradable substrate. This is always a transfer operation: devices are made conventionally, thinned, released and bonded onto the new substrate. The published work by Jung and colleagues in Nature Communications (opens in a new tab) is the clearest example — high-performance flexible microwave and digital electronics on biodegradable cellulose nanofibril paper, with fungal biodegradation of the finished device demonstrated.
Bio-substrate as interface
The biological material is a thin functional layer on top of a finished silicon device: a self-assembled monolayer carrying antibodies or aptamers over a CMOS sensor array, a hydrogel over an electrode, a microfluidic channel that brings a sample to it. This is the commercially serious case — pure-play foundries publish silicon-based microfluidics platforms (opens in a new tab) for sequencing, liquid biopsy and microelectrode arrays — and it is where most fab-side engineering effort actually goes.
Bio-substrate as template
The biological molecule is not part of the finished device at all; it is the thing that positions something else. DNA origami binds to lithographically defined sites and places emitters, particles or wires with a precision the pattern alone does not provide. It is a patterning aid layered on top of lithography, not a replacement for it.
Bio-substrate as payload
The silicon is the instrument and the biology is what is being run on it: neurons or organoids cultured on a microelectrode array, kept alive by incubation and media handling. Here the wafer engineering is entirely conventional and the hard problems are sterility, life support and interpretation — which is why this class belongs in a different register from the other three.

Where biology is admitted to a silicon flow

The split flow every publicly reported bio-facing device uses. The front-end lane never sees biological material. The admission point is a decision, not an accident, and everything after it runs under different cleanliness rules with different metrology — which is why the model lane exists at all.

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

The process, in words

The front-end lane is closed to biology by physics. Implant, anneal, plasma and vacuum destroy every material on the bio-substrate list, so a wafer travels through FEOL, BEOL and passivation as ordinary silicon regardless of what it will eventually become.
The admission point is the design decision. It pairs a material's real ceiling — decomposition temperature, glass transition, solvent and plasma tolerance, whether it can be pumped down at all — with the earliest step in the route that sits below that ceiling. Everything else in the flow follows from where that line is drawn.
The bio module runs under different rules: segregated tools, aqueous and organic chemistry, humidity and temperature control, and a destructive off-line assay that returns the only real answer days after the wafer has moved on. This is where value leaks, because the module is making decisions in the dark.
The model lane exists to close that gap. Tool trace and incoming lot features feed a virtual sensor that predicts the assay result now rather than next week; the prediction is written into the MES as a hold or a bounded correction with an engineer in the loop; and every prediction, offset and approval lands in the lot record that a regulator, customer or auditor will eventually read.

Step-by-step insights

The front end is closed — and that is the good news: Practitioners new to this topic often read 'biology cannot enter the front end' as the constraint that spoils the idea. It is the opposite: it means the hardest, most capital-intensive, most tightly controlled part of the flow does not change at all. The wafers are ordinary, the tools are ordinary, the process control is ordinary, and every bit of existing fab discipline still applies. All the novelty, and therefore all the risk and all the engineering, is concentrated into a bounded module at the end of the route. That is a far more tractable programme than 'a new kind of wafer', and it is why every device that actually ships is built this way.
The admission point is the whole design: Naming the earliest step a material can survive forces four decisions at once: whether the material arrives before or after singulation, which tools are now contaminated for anything else, what metrology remains available, and what the rework path is if the step fails. Teams that skip this and start with chemistry discover the decisions later, individually, each as a surprise. Writing an admission rule per material — with numbers, reviewed like any other process specification — takes an afternoon and changes the shape of the programme, because it converts a research question into a routing question the MES can actually enforce.
Transfer is the real technology, not the substrate: When the biological material is the carrier, the interesting engineering is bond, thin, release and handle: temporary carrier selection, adhesive that survives the subsequent steps without outgassing, release chemistry that does not attack the device, and a total-thickness-variation budget that suddenly becomes the critical dimension of the module. The published cellulose-nanofibril work is a transfer story exactly like this. Any programme that describes itself as 'developing a bio-substrate' but has no bond-and-release engineer is describing a materials project, not a device flow.
What the boundary breaks: the metrology, not the process: This is the part fabs consistently underestimate. A capture monolayer is roughly a nanometre thick, functionally wet, and destroyed by the vacuum an electron microscope needs. A hydrogel changes thickness with humidity. A cellulose film is rough and translucent, so ellipsometry and pattern-recognition autofocus both struggle. A living culture cannot be measured at all in any fab sense. The tools a fab already owns for film thickness, overlay and defect inspection — the mainstay of every process-control estate — simply stop returning meaningful numbers, and the module falls back to a destructive assay that reports days later.
Virtual metrology as a measurement substitute, with an error bound: The honest framing of AI's first job here is narrow: predict, from data the tools already produce, the number the destroyed metrology would have given you, and state how wrong that prediction can be. The published benchmarks set expectations usefully low — a cross-benchmark of machine-learning virtual metrology in mass-produced CVD reported best accuracy around 0.70, with data imputation required because roughly 30% of the inputs were missing. A model at that level is still transformative when the alternative is an eight-day wait, provided the hold rules are designed around the measured error rather than the hoped-for one.
Evidence: the moment the model becomes a design input: If the finished device touches a patient or a diagnostic result, the model that held or released a wafer is part of the device's design history, not part of the analytics stack. That means a version, an intended-use statement, a validation record against held-out material, defined limits, and a lot record that can be reconstructed years later. Building this while the pipeline is being written costs very little; retrofitting it after two years of production is a project. Rung 4 on this page's ladder is essentially the rung at which a team has accepted that and built accordingly.

The five rungs in detail

For each rung: what it looks like on the floor, the signals a reviewer can check in an afternoon, the anti-pattern that traps teams there, and what leaving costs.

The ladder below measures a module, not an organisation. A fab can be excellent at CMOS process control and sit at rung 1 on its first bio-facing module, because the ladder is about a specific boundary: whether the material entering the flow is legible, whether the admission point is written down and enforced, whether the measurement the boundary destroyed has been replaced, and whether any of it would survive being moved or audited.

What each rung actually releases

Value stays close to flat through rungs 1 and 2 — where most bio-facing work sits — and inflects at rung 3, when a substitute measurement starts holding wafers before they consume the expensive step. Rungs 4 and 5 are where the flow becomes certifiable and then portable, which is where commercial volume lives.

Module capability released by stage

Stage 1 · Coupon — 46% of operators. Biology meets silicon only on hand-run coupons — pieces on a carrier, one operator, results in a notebook.
Stage 2 · Characterised — 31% of operators. The boundary is written down and the incoming material is measured, but nothing in the line acts on either.
Stage 3 · Bridged — 16% of operators. A split flow runs on full wafers, and a model stands in for the metrology the boundary destroyed, writing into the MES.
Stage 4 · Qualified — 6% of operators. The bio module is under statistical control with bounded model-based correction, and the model is a versioned design input inside the quality system.
Stage 5 · Transferable — 1% of operators. Recipe, model, control policy and evidence pack move to a second line or supplier without re-deriving the process window.

Curve shape: logistic, plotted from the stage data above. Distribution: Illustrative shape; rung definitions anchored to published device and foundry material.

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

Coupon

46% of operators sit here

Biology meets silicon only on hand-run coupons — pieces on a carrier, one operator, results in a notebook.

Rung 1 is not ignorance. The chemistry usually works, often beautifully, and the people running it know more about the material than anyone else in the building. What is missing is a record that pairs what was on the coupon with what was in the beaker, the room and the schedule — so every result is a story about one afternoon rather than a fact about a process.

The tell is that the coupon has no identity. Ask which substrate lot produced the best binding signal and the answer is a month; ask which reagent lot, and the answer is 'the good batch'. The variables that actually moved the result — ambient humidity in the bay, the gap between plasma activation and immobilisation, how long the reagent had been out of the fridge — were never variables at all, because nobody wrote them down.

This is a cheap rung to leave and an expensive one to occupy. The cost is not the failed experiments; it is that nothing amortises. The tenth experiment costs what the first did, and the moment somebody asks for a full wafer, none of the coupon evidence transfers — edge exclusion, handling, thermal uniformity across 200 mm and vacuum chucking were all outside the scope of every run so far.

In practice

The coupon that worked in March

A biosensor group demonstrated a three-fold improvement in capture-layer binding signal on quarter-wafer coupons and wrote it up. Six months later, with a customer waiting, nobody could reproduce it. The notebook recorded spin speed and cure time faithfully. It did not record the reagent lot, the bay humidity that week, or the fact that the March coupons sat for twenty minutes between activation and immobilisation because the technician was at lunch — which, when the group finally instrumented the bench, turned out to be the dominant term.

What it looks like

Work runs on quarter-wafer coupons or split pieces, not full wafers
Substrate and reagent lots are identified by date, not by a record a system can query
No written ceiling — thermal, solvent, plasma — for any of the materials in use
The assay result lives in a spreadsheet that nothing else joins to

Diagnostic signals you can check this week

Ask which substrate or reagent lot produced the best result. If the answer is a date, you are here
Look for a written thermal, solvent and plasma ceiling per material — at rung 1 it lives in one person's head
Check whether coupon identifiers join to anything queryable: a lot record, an assay database, a trace file
Ask what the edge exclusion will be on a full wafer. Silence means the flow has never been scoped beyond the bench

Anti-pattern · Buying a materials-informatics platform

The instinctive move at rung 1 is to reach for machine-learning materials discovery, because the problem presents itself as 'we do not understand this material'. But you have thirty rows of data, half of them unlabelled, and the variable that is actually moving the result is a handling variable, not a chemistry one. Instrument the bench first — lot identifiers, ambient capture, timestamps between steps, an automatic join to the assay result. Three months of that beats any screening platform, and it is the dataset a screening platform would need anyway.

What holds you here

Nothing joins. The coupon, the material lot and the assay result live in three places, so no result is reproducible and no model has a training set.

Highest-leverage next move

Give every coupon, substrate lot and reagent lot an identity a system can query, and capture ambient conditions and inter-step timing automatically.

Cost of leaving

Effort: 2–4 months
Team: One process engineer, one data engineer part-time, the bench technician who actually runs it
Risk: Low — the work is additive and nothing in production depends on it yet
To next stage: 2–4 months

If this is you, the next step is

A two-week engagement: define the lot record, the trace capture and the assay join.

Instrument one bio bench

Stage 2

Characterised

31% of operators sit here

The boundary is written down and the incoming material is measured, but nothing in the line acts on either.

Rung 2 is where the programme stops being folklore. Somebody has written the ceilings down — this film decomposes above roughly this temperature, this chemistry does not survive re-exposure to plasma, this material may not enter before passivation — and somebody else has made the incoming material legible, so a lot is a row of numbers rather than a delivery note.

The characteristic artefact of this rung is a designed experiment whose model surprises people. Almost every bio-facing module has at least one dominant term that was never a fab variable: bay humidity, the queue time between activation and coating, the age of a reagent since it left cold storage. Discovering it is genuinely valuable, and it is also the moment the rung's limitation becomes obvious — you now know what drives the result and the line still cannot act on it before the assay comes back.

The risk at rung 2 is that knowledge substitutes for control. A team with a good DOE model and a confident explanation of last quarter's excursion feels advanced, and the module is still running to a fixed time-based recipe on wafers whose incoming material varies by harvest. Value released is close to zero until something between the model and the wafer changes.

In practice

The DOE that found the humidity

A team ran a three-factor design on plasma activation power, activation time and queue time before immobilisation, and fitted the assay result against tool trace. The largest term was none of the three: it was bay relative humidity, which nobody had controlled because humidity had never been a variable in the CMOS line the module was bolted onto. They now had an explanation for eighteen months of unexplained lot-to-lot scatter — and the module continued to run the same fixed recipe for another year, because knowing was not wired to anything.

What it looks like

A written admission rule per material: what it survives, and the earliest step it may enter
Substrate and reagent lot genealogy captured as data and joined to wafer identifiers
Designed experiments analysed with real models rather than one-factor-at-a-time
Still no statistical control on the bio module — the recipe is time-based and open-loop

Diagnostic signals you can check this week

Ask to see the admission rule for one material. At rung 2 it exists and has numbers in it
Check whether incoming certificates of analysis are captured as data or filed as PDFs
Ask what fraction of assay results can be joined automatically to trace and material lot — if a person does the join, it is not a pipeline
Ask what the bio module does differently when a new substrate lot arrives. Usually nothing

Anti-pattern · Going straight from a DOE to an automatic recipe

A model fitted on forty runs with no held-out lots looks strong enough to drive the tool, and the proposal to close a loop on it arrives about a week after the DOE readout. It is the wrong next step in the wrong order. A model that has never been tested against material it did not see will confidently correct on the first lot from a new harvest, and the correction will be wrong in a direction nobody is monitoring. Build the prediction as a prediction first — validated on held-out lots, reported with an error bound, changing nothing — and earn the loop afterwards.

What holds you here

You can explain variance after the fact, but nothing in the line acts on it before the off-line assay returns days later.

Highest-leverage next move

Pick the single measurement the boundary destroyed and build a virtual sensor for it — validated on held-out lots, with a stated error bound, before any control loop.

Cost of leaving

Effort: 4–8 months
Team: Process engineer, data engineer, an equipment engineer for trace access
Risk: Low to medium — the work is analytical, but the first trace extraction touches the tool network
To next stage: 4–8 months

If this is you, the next step is

We take your existing experiments and tell you honestly whether they can support a predictive model yet.

Turn the DOE into a virtual sensor

Stage 3

Bridged

16% of operators sit here

A split flow runs on full wafers, and a model stands in for the metrology the boundary destroyed, writing into the MES.

Rung 3 is the first rung where the boundary becomes an engineered object rather than a fact of life. The route is split on purpose: everything that needs heat, vacuum, implant or plasma happens while the wafer is still purely silicon; the biological material is admitted at a named step; and the wafers that have crossed that step are restricted to a defined toolset for the rest of their lives.

The change that releases value is not the segregation, though — it is the substitute measurement. A capture monolayer is a nanometre or so thick, wet for most of its life and destroyed by the vacuum a scanning electron microscope needs, so the fab's inline metrology simply has nothing to say about it. A virtual sensor built from activation dose, chamber humidity, queue time, reagent age and incoming lot signature does have something to say, and if its error bound is honest it can hold a wafer before that wafer consumes the most expensive step in the module.

The discipline at rung 3 is closer to site reliability engineering than to data science. What is the error budget for this prediction? What happens when the trace feed stops? Which recipe do we revert to, and who has actually pressed that button in the last six months? Programmes that arrive here with a good model and none of those answers spend their first quarter discovering them the expensive way.

In practice

The eight-day truth, made same-shift

A post-CMOS functionalisation module ran open-loop because the only measure of success was a destructive binding assay that reported eight days later, by which time the lot had moved on and the reagent had been consumed. The team fitted predicted binding density from activation trace, bay humidity and reagent lot age, validated it on lots the model had never seen, and wrote the prediction into the MES as a disposition field with a hold at the lower error bound. Nothing about the chemistry changed. Scrap fell because bad wafers now stopped before the expensive step instead of after it.

What it looks like

Standard front end, segregated bio module, and a documented route between them
A virtual metrology model predicts the unmeasurable step result with a stated error bound
The prediction is a field on the lot in the MES with a hold rule — not a dashboard
Off-line assay results flow back automatically and join to wafer, trace and material lot

Diagnostic signals you can check this week

Ask where the prediction appears. If the honest answer is a dashboard, the module is still open-loop
Ask for the model's error on held-out material lots, not on the training set
Check that the fallback recipe exists and find out when it was last exercised deliberately
Check whether a bio-exposed wafer can physically be routed back into a shared tool, or whether the MES stops it

Anti-pattern · Trusting the prediction because it correlates

A correlation coefficient on training lots is not an error bound, and a hold rule set from one is a machine for scrapping good wafers and releasing bad ones. The published benchmarks are sobering on this point: a cross-benchmark of machine-learning virtual metrology in mass-produced CVD reported a best prediction accuracy of about 0.70 after careful imputation and non-linear feature selection. Design the hold rule around the accuracy you actually measured on held-out material, and state what the model may not be used for.

What holds you here

The module is bridged but not qualified: control limits, model version and biological evaluation evidence are not part of the release record.

Highest-leverage next move

Put the model under change control as a design input, set control limits from the holdout rather than from the training set, and generate the evidence pack as a by-product of the run.

Cost of leaving

Effort: 9–18 months
Team: Process engineer as module owner, ML engineer, integration engineer, equipment engineer
Risk: Medium — the first MES write and the first route restriction both need a rollback path
To next stage: 9–18 months

If this is you, the next step is

The rung 2→3 transition is our most common engagement. Typically 90 days on one module.

Bridge one bio module

Stage 4

Qualified

6% of operators sit here

The bio module is under statistical control with bounded model-based correction, and the model is a versioned design input inside the quality system.

Rung 4 is where the module stops being a clever bridge and becomes a process anyone is willing to certify. Correction is bounded: the controller may move activation dose within a stated range and no further, the range came from held-out material rather than from optimism, and an engineer sees and can veto every step outside it. The stability question — how aggressive a correction can be before the loop itself becomes the source of variance — is a well-studied one in run-to-run control, and importing that literature is faster than rediscovering it on product wafers.

The distinguishing artefact is not the controller, it is the record. At rung 4 the answer to 'why was lot X released' is a single document that names the recipe, the model version that produced the prediction, the offsets applied, the control limits in force, the person who approved the exception and the biological evaluation file covering the exposed material set. That document is generated by the flow, not assembled for the audit, which is the whole reason the rung is worth reaching.

The cost of rung 4 is real and it is mostly not engineering. It is the discipline of treating a model as a released artefact under change control — with the awkward consequence that improving it costs paperwork. Teams meet that consequence in one of two ways: they build a lightweight, honest re-validation path, or they freeze the model and let it go quietly out of validity. Only one of those is a rung-4 organisation.

In practice

The audit that took an afternoon

An inspector asked a medical-device manufacturer why a particular lot of functionalised biosensor wafers had been released eleven months earlier. The team pulled one record: incoming material lot and its certificate values, the activation recipe, the model version and its validation summary, the predicted binding density with its error bound, the two bounded offsets the controller had applied, the control limits in force that week, and the engineer who signed the exception on the third wafer. The review closed the same day. Nothing in that record had been prepared for the audit.

What it looks like

Bounded run-to-run correction on the bio step, with limits derived from a holdout
The model has a version, an intended use, a validation record and an owner
Contamination and segregation controls are enforced by equipment automation and monitored
The lot record names recipe, model version, offsets applied and approvals

Diagnostic signals you can check this week

Ask whether the model has a version number that appears in a lot record, not just in a repository
Ask how a model change is released today — if the honest answer is 'we redeploy it', there is no design control
Check whether run-to-run correction has hard bounds, and where those bounds came from
Ask when the biological evaluation file was last reconciled against the actual exposed material set

Anti-pattern · Freezing the model to keep the auditor happy

Re-validating a model that sits in a regulated release path is genuinely expensive, so the pragmatic-looking move is to freeze the version and stop touching it. Two years later the supplier has changed grade, the tool has been rebuilt twice and the frozen model is quietly out of validity — still versioned, still documented, still wrong. The fix is a re-validation path proportionate to the change: a defined holdout protocol, a fixed acceptance criterion and a pre-agreed evidence template, so a refresh is a week of work rather than a project nobody will fund.

What holds you here

The flow is qualified where it was built and nowhere else: a second line or a new substrate supplier means re-deriving the window from scratch.

Highest-leverage next move

Prove transfer on purpose — run the flow on a second chamber, line or supplier lot and measure how much of the model, the window and the evidence actually survives.

Cost of leaving

Effort: 12–24 months
Team: Module owner, ML engineer, quality and regulatory partner, equipment automation engineer
Risk: Higher — the binding constraint moves from engineering to evidence and change control
To next stage: 12–24 months

If this is you, the next step is

What the lot record must contain, and how to generate it from the flow rather than for the audit.

Design the evidence path

Stage 5

Transferable

1% of operators sit here

Recipe, model, control policy and evidence pack move to a second line or supplier without re-deriving the process window.

Rung 5 is narrower than it sounds, and deliberately so. It is not an autonomous bio-fab. It is a specific claim: that this flow, on this material class, can be stood up somewhere else — a second chamber, a second line, a contract manufacturer, a new supplier grade — without going back to the bench, because the things that vary have been made into parameters rather than assumptions.

What makes it possible is that the two sources of difference have been separated. Material difference is carried by lot signature features, so a new grade produces a feed-forward offset rather than a new model. Equipment difference is carried by an explicit matching step, which is exactly the problem domain-adaptation research addresses in fab terms: aligning data from one tool to another so a model trained on the first remains valid on the second.

Rung 5 is also the rung most likely to regress, because portability decays without anyone touching it. Suppliers reformulate, tools get rebuilt, standards revise, and the transfer that took six weeks last year takes nine months this year. Operators who hold the rung treat transfer cost as a standing metric and rehearse a transfer deliberately rather than waiting for a business reason to need one.

In practice

The second supplier that took six weeks

A manufacturer with a single-source biological reagent finally qualified a second supplier. The first time they had done this, years earlier, it consumed nine months of re-derivation because the process window had been discovered empirically and written nowhere. This time the incoming certificate values were already model features; the new grade produced a measurable shift in one feature, the feed-forward offset absorbed it, the holdout confirmed the residual was inside limits, and the evidence pack regenerated from the same pipeline. Six weeks, most of it waiting for material.

What it looks like

Material lot signature is a model feature, so a new supplier grade is an offset rather than a project
Transfer cost — elapsed days to re-qualify elsewhere — is measured and trending down
Equipment differences are handled by an explicit matching or adaptation step, not by refitting
The evidence pack regenerates itself at the receiving site from the same pipeline

Diagnostic signals you can check this week

Ask what a supplier grade change costs today, in weeks. That number is the rung
Check whether the model consumes incoming lot attributes as features or ignores them
Ask whether a second tool runs the same model, and what step makes that valid
Ask when a transfer was last rehearsed rather than forced by a business event

Anti-pattern · Declaring a portable process from one successful transfer

One good transfer is an anecdote, usually explained by the two sites sharing a tool generation, a supplier and half the engineering team. Portability is a claim about the second and third transfers, made with the cost of each measured. Teams that declare victory early stop maintaining the matching step and the lot-signature features, and rediscover at the next transfer that the model had been quietly memorising one site all along.

What holds you here

Portability decays silently as suppliers, tools and standards move, so the transferable state has to be re-earned rather than held.

Highest-leverage next move

Make transfer cost a standing KPI, and rehearse one transfer a year deliberately instead of waiting for a business event to force it.

Cost of leaving

Effort: Continuous
Team: Module owner, platform engineer, quality partner, plus a standing review forum
Risk: Concentrated — low frequency, high consequence, and mostly regulatory in nature

If this is you, the next step is

We run the transfer against a second tool or supplier lot and measure what actually survives.

Stress-test a transfer

Where bio-facing programmes actually sit today

The distribution across the ladder, and why the rung 2 → 3 step is where almost all of the loss happens.

Most bio-facing work sits at rung 1 or rung 2 — on coupons, or characterised but uncontrolled. That is not a criticism of the teams doing it; it reflects where the field genuinely is. The devices that have crossed into qualified, repeatable production are a small and identifiable set, and they share a common trait: they treated the boundary as an engineering artefact early, rather than treating the material as a research subject for years.

Illustrative distribution of bio-facing modules across the five rungs

Rung 2 is where knowledge accumulates without control. The drop from rung 2 to rung 3 is the largest single loss on this ladder, and it is a measurement-substitution problem rather than a materials problem.

Share of modules

46% — 1 · Coupon (the bench)
31% — 2 · Characterised (knowledge, no control)
16% — 3 · Bridged
6% — 4 · Qualified
1% — 5 · Transferable

Source: Illustrative distribution, synthesised from published device work and foundry platform material (imec, X-FAB, Nature-published device papers)

The reason so much of the distribution is stacked at the bottom is that the rung 2 → 3 step demands a discipline the field is not organised around. Rung 2 rewards understanding: a good designed experiment, a plausible mechanism, a paper. Rung 3 rewards something else entirely — a prediction that is wired into the manufacturing execution system (opens in a new tab), with an error bound, a fallback recipe and a named owner. Research groups rarely have a reason to build that; production teams rarely have material stable enough to justify it until the research group is finished. The programmes that cross do it by putting one process engineer in charge of the module and giving them both jobs.

It is worth being precise about what the far side looks like, because the honest version is less dramatic than the pitch deck. It is not a biological fab. It is a conventional CMOS or MEMS line with a bounded, segregated module at the end of the route, a written admission rule, a virtual sensor standing in for one destroyed measurement, and a lot record that can be reconstructed. Institutes that publish in this space — imec's health technologies programme (opens in a new tab) among them — describe exactly that architecture, and the foundries that sell into it describe it too.

The admission ledger: what each material costs the line

One row per material class — the ceiling it lives under, the earliest step it may enter, what it forces on the line, the metrology it breaks and the model that has to stand in.

Every bio-substrate decision reduces to one row of the table below. A material has a real physical ceiling, that ceiling determines the earliest step in the route at which it can be admitted, admission forces a set of consequences on the rest of the line, and one of those consequences is always that a measurement you relied on stops working. The last column is the only one AI can help with directly — which is why this page treats virtual metrology as the first model, not the last.

Material class	Practical ceiling	Earliest admissible step	What it forces on the line	Metrology it breaks	The model that stands in
Cellulose nanofibril (CNF) film	Low hundreds of °C, and dimensionally sensitive to humidity	After the devices exist — the substrate is joined to finished, thinned device layers by transfer	A bond-and-release module, humidity-controlled handling, and a wafer shape the standard chuck was not designed for	Ellipsometry on a rough translucent surface; vacuum chucking; overlay on a dimensionally unstable carrier	Predicted bondline thickness and TTV from bonder trace, film lot moisture and ambient
Silk fibroin film	Roughly 150 °C for a stable film; water-processed, and it will not tolerate a vacuum bake	Post-passivation, often post-dice — as a coating, an encapsulant or a handle	An aqueous processing bay outside the CMOS toolset, and sterility handling if the device is implantable	Any vacuum metrology; contact profilometry on a soft film	Predicted film thickness from spin and dry curve, solids content and ambient humidity
Hydrogel / polymer gel layer	Wet at all times; degrades above roughly 60 °C and changes thickness as it dries	After wafer test, frequently after singulation	A segregated wet bench, temperature and humidity controlled storage, and a short shelf life on finished parts	Ellipsometry and electron microscopy — both need a dry sample in vacuum	Predicted swollen thickness and ligand density from dispense, spin and cure trace
Self-assembled monolayer plus capture chemistry	Roughly 120 °C; destroyed by re-exposure to plasma or aggressive solvent	After passivation and pad opening, before or after singulation	Organics excluded from every shared toolset the wafer might otherwise re-enter; a dedicated activation chamber	Sub-nanometre thickness sits below inline film metrology — only contact angle and a destructive assay remain	Predicted binding density from activation dose, queue time, humidity and reagent lot age
DNA origami template	Room temperature, in a high-salt buffer; nothing about it survives a hot or dry step	After lithographic patterning of binding sites — a solution step on a finished pattern	Salt and buffer handling inside a fab, which is an ionic contamination problem, plus dedicated wet tools	Placement yield and orientation are invisible to inline metrology; AFM or SEM sampling only	Predicted site occupancy from pattern critical dimension, buffer chemistry and incubation conditions
Living cell or organoid layer on a MEA	37 °C, wet, and it has to stay alive	After packaging — the device is finished and the biology is the payload	Incubation, sterility, media handling and a biosafety regime; this is a biology laboratory attached to a device, not a fab step	Everything a fab measures. The substrate is a culture, not a film	Electrophysiological health scores and culture-condition models — not a wafer metric at all

The admission ledger for common bio-substrate material classes. Ceilings are order-of-magnitude engineering ranges for orientation, not specifications: every formulation needs its own measured limit, and the number that matters is the one your own material shows on your own tools.

Two walls run through every row. The first is thermal and chemical: the material's own ceiling against the route's remaining steps. The second is contamination, and it is the one that catches fabs by surprise, because it is not about the wafer being damaged — it is about the wafer damaging the fab. Organics, salts, proteins and biological material are excluded from shared toolsets for the same reason copper was: cross-contamination in a chamber is discovered as a yield excursion on somebody else's product, weeks later, and the investigation is expensive even when the answer turns out to be benign.

The practical consequence is that admitting a material forks the line. After the admission point, a wafer belongs to a restricted toolset, and getting it back into general circulation requires a defined clean and a positive release — or, more commonly, is simply never allowed. That is a routing rule the MES can enforce, and enforcing it there rather than by procedure is one of the cheapest maturity moves available: it converts a discipline problem into a configuration problem. The trade press covering fab manufacturing and process integration (opens in a new tab) treats this segregation question as routine for any non-standard material, and it should be treated that way here too.

Teams tend to read this ledger from the bottom up, because the bottom rows are the exciting ones. The programmes that produce shipped devices read it from row four: capture chemistry over a finished CMOS sensor, admitted after passivation, in a segregated activation bay, with a virtual sensor standing in for a measurement that does not exist. It is the least visionary row on the table and it is where essentially all of the commercial volume in bio-facing silicon currently sits.

Real today, published research, and still speculation

The three registers this topic constantly blends — separated, with what each source actually claims and at what scale.

Almost every confusing claim about biology and silicon comes from blending three registers: what is shipping, what a paper demonstrated once under controlled conditions, and what somebody hopes follows from it. Separating them is not scepticism — it is the only way to plan, because each register implies a different decision. Shipping means you can buy it and qualify it. Published research means you can attempt it with a research budget and an honest failure probability. Speculation means it belongs in a strategy conversation, not a capital plan.

Claim family	Shipping today	What the published research actually claims	Still speculation
Silicon devices that meet living tissue	Yes. Silicon CMOS neural probes and microelectrode arrays are catalogue products, and silicon-based microfluidics is an offered foundry platform	Jun et al. (Nature, 2017) describe the design, fabrication and performance of a fully integrated silicon probe recording well-isolated activity from hundreds of neurons in freely moving animals; X-FAB publishes a microfluidics platform with a customer track record in sequencing, liquid biopsy and microelectrode arrays	That the wafer process changes. It does not — these are ordinary CMOS and MEMS flows with a controlled exposed-material set
Bio-derived, biodegradable substrates	Not in volume silicon. Flexible hybrid electronics ships thinned die bonded onto polymer, not devices grown on biological film	Jung et al. (Nature Communications, 2015) report high-performance flexible microwave and digital electronics on biodegradable cellulose nanofibril paper, made by transferring thin-film devices, and demonstrate fungal biodegradation of the finished electronics	Transistors fabricated directly on a biological substrate at production yields and volumes
DNA-directed placement and self-assembly	Nothing in production. Where it is used at all, lithography defines the sites first	Kershner et al. (Nature Nanotechnology, 2009) report 70–95% of e-beam-patterned sites occupied by individual DNA origami with angular dispersion as low as ±10° on diamond-like carbon and ±20° on silicon dioxide, in roughly 100 mM magnesium chloride buffer; Gopinath et al. (Nature, 2016) scale placement to 65,536 independently programmed photonic-crystal cavities on one chip	DNA replacing lithography. Every published demonstration needs lithography to define where the DNA goes
Transient and resorbable electronics	Niche research and early clinical devices only	Work published in Science in 2012 (Hwang et al.) demonstrated 'physically transient' silicon electronics — thin silicon nanomembrane devices designed to dissolve in biofluid over a controlled period. The device layer is still silicon; the transience is a thickness-and-encapsulation design	Consumer electronics that dissolve, or production flows organised around designed dissolution
Biological computing substrates	Benchtop research units. Neurons cultured on silicon microelectrode arrays, with life support	The organoid-intelligence programme (Smirnova, Hartung et al., Frontiers in Science, 2023) sets out a research agenda and its prerequisites rather than a product; Cortical Labs publishes closed-loop work with neuronal cultures on multi-electrode arrays	Biological compute displacing silicon compute in any manufacturing sense, or a 'grown' processor
AI designing the substrate itself	ML-guided screening is genuinely used in materials R&D, upstream of any fab	The A-Lab (Nature, 2023) reports an autonomous laboratory synthesising novel inorganic compounds from computationally predicted candidates — a closed synthesis-and-characterisation loop, explicitly a research instrument	A model producing a qualified fab process. Qualification is evidence, holdouts and elapsed time, none of which a generative model shortens

The three registers, by claim family. Middle-column claims are stated as the cited work states them, including scale and conditions; nothing has been rounded up.

DNA origami are synthesized in solution and uncontrolled deposition results in random arrangements; this makes it difficult to measure the properties of attached nanodevices or to integrate them with conventionally fabricated microcircuitry.
Kershner et al., Placement and orientation of individual DNA shapes on lithographically patterned surfaces, Nature Nanotechnology

That sentence is the whole discipline in one line. The interesting property of biological self-assembly — that it happens by itself, in solution — is also precisely what makes it unmanufacturable without conventional patterning to constrain it. The published solution is not less lithography but more: electron-beam-defined binding sites, dry oxidative etch, controlled buffer chemistry. Read against this page's ladder, the state of the art in directed self-assembly is a very sophisticated rung 2 — deeply characterised, not under control in any production sense.

The register that most often gets over-read is the last row. Autonomous laboratories are a real and impressive development, and the A-Lab work published in Nature (opens in a new tab) is worth reading in the original rather than in summary, because the paper is careful about what it did and did not establish. What it demonstrates is a closed loop for synthesis and characterisation of candidate compounds. What it does not do — and does not claim to do — is produce a process window, a control strategy or an evidence pack. Those are the artefacts a fab actually needs, and they are earned on tools with product wafers, not in a screening loop.

The practical conclusion is deflationary, and it is the most useful thing on this page: future-readiness for bio-substrates is almost entirely present-readiness. The capabilities that let a fab exploit any of the middle column when it matures — lot-level provenance data, a written and enforced admission boundary, validated virtual metrology with honest error bounds, bounded correction, and a reconstructable evidence trail — are the same capabilities that pay for themselves today on the least exotic row of the ledger. There is no version of the exciting future that a fab reaches without them, and no version of the boring present that does not need them anyway.

Where AI lands in a bio-facing silicon flow

Six domains, the decision worth wiring in each, the system that owns it, the KPI it moves, and the rung at which it pays.

AI value in a bio-facing flow concentrates in six domains, and none of them is materials discovery. A decision is a good first candidate when three things are true: the data already exists in a system you control, the decision cycle is short enough to measure inside a quarter, and the KPI it moves is one a plant manager already reports. The map below is how we scope first and second use cases with fabs and MEMS foundries running bio-facing modules.

Domain	The decision worth wiring	System of record	KPI it moves	Pays at
Incoming material control	Accept, reject or bin an incoming bio-material lot, and set the recipe offset that lot needs	MES lot genealogy plus the laboratory system holding certificates of analysis	Scrap rate by material lot, recipe rework	Rung 2–3
Transfer and bonding	Carrier and bond recipe per lot; predicted total thickness variation and voiding before the stack is committed	MES plus bonder fault-detection trace	Post-bond TTV, void count, module yield	Rung 3
Activation and functionalisation	Activation dose per wafer, and hold-or-release before the wafer consumes the expensive reagent step	MES disposition field plus the run-to-run controller	Assay coefficient of variation, functional yield, scrap	Rung 3–4
Post-CMOS wet processing	Bath life and replacement timing, and the cross-contamination risk of the next wafer in	Fault detection and classification plus chemical management	Particle adders, excursion count, bath cost	Rung 3–4
Segregation and contamination control	Which toolset a bio-exposed wafer may re-enter, and after which clean	Equipment automation rules inside the MES	Cross-contamination excursions, unplanned tool downtime	Rung 2 onwards
Qualification and release	What evidence releases a lot: model version, limits in force, biological evaluation status	Quality management system and the design history file	Time to release, audit findings, re-validation effort	Rung 4–5

The bio-facing decision landscape. 'Pays at' is the rung on this page's ladder at which the decision typically starts earning; wiring a rung-4 decision from a rung-1 data foundation is the flying-blind quadrant described below.

Activation and functionalisation is where most fabs should begin, for the same reasons warehouse decisions come first in logistics: the system of record is yours, the feedback loop is measured in shifts rather than quarters, and scrap is a number nobody disputes. Transfer and bonding carries larger absolute value where a transferred substrate is in play, but it touches the device stack and therefore a longer approval path. Qualification and release has the highest strategic value and the lowest early tractability — it is a rung-4 decision that cannot be wired before the evidence pipeline exists.

Choosing the control strategy for a bio step

Plot the incoming material's variance against what you can measure inline on the step. The quadrant names the next investment — and in three of the four, that investment is not a better recipe.

Flying blind

Variance you cannot see until the assay returns
The most common and most expensive position
Fix: build the virtual sensor before any control loop

Feed-forward

Variance is visible in time to do something about it
Fix: feed the lot signature forward into the recipe
Run-to-run correction on the residual only

Stable but unprovable

Runs acceptably open-loop, and you cannot demonstrate why
Fine until volume or an auditor arrives
Fix: a sampling plan and an evidence trail before scaling

Ordinary process control

This is now a normal fab problem
Fix: statistical process control and periodic model checks
Resist over-engineering — the exotic material is behaving

Incoming material variance — top: Biological, lot-to-lot, harvest-driven, bottom: Tight, synthetic, specified

Inline measurement on the bio step — left: None — destructive off-line assay only, right: Inline and fast enough to act on

Most bio-facing modules start in the top-left quadrant and try to leave it by improving the recipe, which cannot work: you cannot tune what you cannot observe. The move that changes the quadrant is a substitute measurement, and the move that changes the axis is provenance data — making the incoming variance visible before the wafer is committed rather than after the assay. Those two investments, in that order, are the entire content of the rung 2 → 3 transition.

What silicon-and-biology has actually demonstrated in public

Three figures from peer-reviewed work: the scale directed self-assembly has reached, the placement yield it achieves on a lithographically patterned surface, and the accuracy a virtual-metrology model reported in mass production. Every figure is externally sourced and listed in the references at the end of this page.

65,536

Independently programmed photonic-crystal cavities built on one chip by placing DNA origami on lithographically defined sites

Gopinath et al., Nature

70–95%

Of patterned binding sites occupied by a correctly oriented DNA origami, at angular dispersion as low as ±10° on diamond-like carbon

Kershner et al., Nature Nanotechnology

0.70

Best prediction accuracy reported for a machine-learning CVD virtual-metrology model in mass production — the realistic ceiling to design hold rules around

Xie and Stearrett, arXiv

What the boundary looks like in public

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

The clearest evidence for the boundary thesis is in what the successful programmes chose to build. In every case below the biological material was kept out of the front end entirely, the hard engineering went into the admission point and the module after it, and the thing that made the result durable was not the chemistry but the fact that somebody else can now run it.

Three programmes read against the ladder

Outcomes as reported by the organisations themselves; verify figures against the linked source before reusing them. The card images are generated industry scenes from our library — none depicts a named organisation's facility, and none implies endorsement.

Illustrative cleanroom scene: a wafer on an inspection stage beneath a multi-lens metrology head, with an analysis display alongside

imec and the Neuropixels programmeNanoelectronics research institute · Leuven · CMOS neural probes25

Challenge: Record from hundreds of neurons at once, inside living tissue, on a probe cheap and reliable enough that laboratories worldwide could standardise on the same part rather than each building their own.
Approach: Do everything difficult in ordinary silicon. The amplifiers, multiplexers and digitisers sit on the same CMOS die as the electrode array, the shank is defined by conventional lithography and MEMS release, and the only genuinely biological consideration is the exposed material set that meets tissue. Nothing in the front end is bespoke to biology.
Reported outcome: As published in Nature by Jun and colleagues, the probe records well-isolated activity from hundreds of neurons simultaneously in awake, freely moving animals; the design is now distributed as a catalogue product through imec, with published specifications, calibration files and user documentation.; What it shows about the curveThe rung that mattered was rung 5. A research prototype that only its inventors can run is a rung-2 artefact; a part with documentation, calibration data and a supply route is a transferable process, and that is what turned a probe into a standard.

Jun et al., Nature — fully integrated silicon probes (opens in a new tab)

Illustrative laboratory scene: a stack of wafers on a stage beneath a robotic head, with signal-trace displays behind

University of Wisconsin–Madison (Ma group)University research group · flexible and biodegradable electronics12

Challenge: Replace the non-renewable, non-biodegradable substrate under a high-performance microwave device with a wood-derived one, without giving up device performance.
Approach: Fabricate the active devices conventionally, then transfer thin-film device layers onto cellulose nanofibril paper. The substrate swap happens after every high-temperature and vacuum step, which is what makes it possible at all; the engineering effort goes into transfer, adhesion and handling of a humidity-sensitive film.
Reported outcome: Reported in Nature Communications: high-performance flexible microwave and digital electronics on biobased, biodegradable cellulose nanofibril paper, including gallium arsenide microwave devices in transferrable thin-film form, with fungal biodegradation of the finished electronics demonstrated.; What it shows about the curve'Bio-substrate' almost always means 'transfer'. Read against the ladder this is an excellent rung 2 — deeply characterised, published, and not a controlled process: there is no lot genealogy, no substitute metrology and no evidence pack, because a research result does not need them.

Jung et al., Nature Communications — cellulose nanofibril electronics (opens in a new tab)

Illustrative cleanroom scene: an inspection head scanning wafers on a track while an operator reviews a wafer map

X-FABPure-play analogue/mixed-signal and MEMS foundry · 200 mm lines35

Challenge: Make the bio-facing boundary a catalogue capability rather than a bespoke project, so customers building sequencing, liquid-biopsy and microelectrode-array devices can buy a qualified route instead of inventing one.
Approach: Combine established silicon technologies — photolithography, etching, bonding — with glass and polymer, and integrate microfluidic structures onto silicon ASICs as an on-chip combination, offered as a documented platform alongside the foundry's CMOS, SOI and MEMS processes.
Reported outcome: X-FAB publishes silicon-based microfluidics as a technology offering with a stated customer track record across next-generation DNA sequencing, liquid biopsy and microelectrode arrays for drug development and food-safety testing, and reports improved system efficiency and reproducibility from the higher level of integration.; What it shows about the curveThis is what rung 5 looks like commercially: the flow is a product. A customer can order it, a second customer can order it, and neither has to re-derive the window — which is only possible because the admission point, the segregation rules and the qualification path were engineered once and written down.

X-FAB — silicon-based microfluidics (opens in a new tab)

Read together, the three cases make one argument. The programmes that reached production did not solve a materials problem more cleverly than anyone else; they industrialised a boundary. The research result that stayed a research result is not worse science — it simply never needed lot genealogy, a substitute measurement or an evidence trail, because nobody was going to ship it. If you intend to ship, those three artefacts are the work.

The reference architecture for a bio-facing module

What actually has to exist at each rung — and which layer you can honestly defer.

A bridged module needs five layers, and the order in which they are built decides whether the programme compounds or stalls. The architecture below is deliberately unfashionable: nothing in it is specific to a vendor, and every layer is defined by what it must guarantee rather than by what product provides it. The layer people skip is the assay and label pipeline, and skipping it is why so many virtual-metrology attempts in this space never get a second training set.

Layers required by rung

Each layer is annotated with the rung that first requires it. A module trying to reach rung 3 without the label pipeline is building a rung-2 experiment with extra infrastructure.

Substrate and reagent supply
Stage 1+
- Lot genealogyHarvest or batch, supplier certificate values, storage age
- Shelf life enforcementTemperature, humidity and expiry enforced by the MES, not by memory
- Incoming characterisationThe two or three cheap measurements that predict behaviour
→
The admission boundary
Stage 2+
- Material ceiling registerDecomposition, glass transition, solvent and plasma limits per material
- Route and segregation rulesWhich tools a bio-exposed wafer may re-enter, and after which clean
- Transfer and bond moduleCarrier, release chemistry and an explicit TTV budget
→
Bio module process control
Stage 3+
- Tool trace captureDispense, spin, dose, chamber humidity and inter-step timing
- Virtual metrologyThe measurement the boundary destroyed, with a stated error bound
- Bounded run-to-run correctionLimited offsets, engineer approval, fallback recipe one switch away
→
Assay and label pipeline
Stage 3+
- Automatic label joinAssay results joined back to wafer, trace and material lot without a person
- Label latency managementTraining on labels that arrive after the lot has shipped
- Holdout designA chamber or lot stream deliberately left uncorrected
→
Evidence and change control
Stage 4+
- Model as design inputVersion, intended use, validation record, owner
- Reconstructable lot recordRecipe, model version, offsets, limits, approvals
- Biological evaluation fileThe ISO 10993 test plan mapped to the actual exposed material set

Pipeline described

Substrate and reagent supply (stage 1+) — Lot genealogy: Harvest or batch, supplier certificate values, storage age; Shelf life enforcement: Temperature, humidity and expiry enforced by the MES, not by memory; Incoming characterisation: The two or three cheap measurements that predict behaviour
The admission boundary (stage 2+) — Material ceiling register: Decomposition, glass transition, solvent and plasma limits per material; Route and segregation rules: Which tools a bio-exposed wafer may re-enter, and after which clean; Transfer and bond module: Carrier, release chemistry and an explicit TTV budget
Bio module process control (stage 3+) — Tool trace capture: Dispense, spin, dose, chamber humidity and inter-step timing; Virtual metrology: The measurement the boundary destroyed, with a stated error bound; Bounded run-to-run correction: Limited offsets, engineer approval, fallback recipe one switch away
Assay and label pipeline (stage 3+) — Automatic label join: Assay results joined back to wafer, trace and material lot without a person; Label latency management: Training on labels that arrive after the lot has shipped; Holdout design: A chamber or lot stream deliberately left uncorrected
Evidence and change control (stage 4+) — Model as design input: Version, intended use, validation record, owner; Reconstructable lot record: Recipe, model version, offsets, limits, approvals; Biological evaluation file: The ISO 10993 test plan mapped to the actual exposed material set

Step-by-step insights

Substrate and reagent supply — the layer that decides whether models generalise: A biologically derived input is a harvest with a certificate, not a specification with a tolerance, and the variance between lots is frequently larger than anything the process contributes. Capturing certificate values as data rather than documents, joining them to wafer identifiers, and adding two or three cheap incoming measurements of your own is the highest-return work available at rung 1 — because without it, a model trained on this quarter's material is memorising this quarter's supplier. Every later layer inherits whatever legibility this one establishes.
The admission boundary — a routing rule, not a policy document: The material ceiling register and the segregation rules only mean something when the MES enforces them. Written procedures degrade under schedule pressure, and the failure mode is not a spoiled wafer but a contaminated shared chamber discovered weeks later on somebody else's product. Encoding the boundary as route restrictions with a positive-release step converts an ongoing discipline problem into a one-off configuration problem, and produces an exception log that is itself a useful signal: a rising exception rate means the boundary and the reality have started to diverge.
Bio module process control — prediction first, correction second: The sequence matters more than the technology. Build the virtual sensor as a read-only prediction, validate it on material it has never seen, publish the error bound, and let it hold wafers before it moves anything. Only then close a bounded loop, and bound it deliberately: the run-to-run control literature is explicit that an over-aggressive controller becomes the dominant source of variance rather than its remedy. Keep the fallback recipe one switch away and exercise it on purpose, because the first unannounced trace outage is a bad time to find out nobody knows how to revert.
Assay and label pipeline — the layer everyone skips: In a bio module the ground truth is destructive, expensive and days late, which makes labels the scarcest resource in the system. If joining an assay result back to the wafer, the trace and the material lot requires a person with a spreadsheet, the join will happen for the first training set and never again, and the model will age out silently. Automating the join is unglamorous integration work and it is what converts a one-off model into a capability. The holdout belongs in this layer too: a chamber or lot stream deliberately left uncorrected is the only way to say later what the correction was worth.
Evidence and change control — build it during, not after: For a regulated device, a model that holds or releases wafers sits inside the design history, so it needs a version, an intended-use statement, a validation record and a lot trail that can be reconstructed years later. Generated as a by-product of the pipeline this costs almost nothing; assembled retrospectively it is a project with an unbounded scope. The same discipline pays outside regulated work, because the question 'why did we hold that lot in March' has exactly one good answer, and it is a record rather than a recollection.

The layer most often deferred is the fallback recipe, and it is the one that decides whether anybody will let the model run at all. A prediction that can be switched off in one action, back to the current time-based recipe, is an operational change a process owner will approve. A prediction without that switch is a change request that sits in a queue for two quarters — and the team spends those quarters improving a model nobody is allowed to use.

A 90-day plan: giving a functionalisation module a measurement

The rung 2 → 3 transition made concrete on one problem — a post-CMOS capture-chemistry module whose only truth is a destructive assay that reports eight days late. Contains no chemistry development.

Moving one rung takes about 90 days when it is scoped to a single module, and several years when it is scoped to a technology. To make that concrete, the plan below runs the transition on a specific and extremely common problem: a CMOS biosensor line where post-passivation surface activation and capture-chemistry immobilisation run open-loop, because nothing inline can measure a monolayer, and the only ground truth is a destructive binding assay that reports eight days after the lot has moved on. There is no chemistry work in the quarter at all — the chemistry already works. The quarter builds the measurement it never had.

Rung 2 → rung 3 on one functionalisation module, in one quarter

One module, one material set, one owner. If a phase needs longer than its window, narrow the scope — one chamber, one product — rather than extending the plan.

Days 1–15
Pin the boundary and baseline the loss
Write the admission rule for every material in the module: what it survives, the earliest step it may enter, which tools it contaminates. Pull twelve months of activation tool trace, material and reagent lot records, and assay results, and join them. Baseline functional yield, scrap and assay coefficient of variation by material lot. Name the module owner — a process engineer whose numbers these are, not a data scientist.
A written admission rule and a scrap baseline by material lot
Days 16–45
Build the virtual sensor, and change nothing
Fit predicted binding density from activation dose, chamber humidity, queue time between activation and immobilisation, reagent age since cold storage and incoming certificate values. Validate on material lots the model has never seen, not on a random split. Publish an error bound and a written statement of what the model may not be used for. Deliberately do not close any loop this month.
A prediction with a measured error bound on held-out lots
Days 46–70
Write the prediction into the MES as a disposition
The prediction becomes a field on the lot with a hold rule set at the measured lower bound, so weak wafers stop before they consume the expensive reagent step. The current time-based recipe stays live as the fallback, one switch away, and the switch gets exercised once on purpose. Every accept, hold and override is logged with its context — this log is the training set for any later correction.
Predicted hold live, override log accumulating
Days 71–90
Close one bounded loop and attribute the result
Allow the controller to adjust activation dose within a stated range on one chamber only, hold a second chamber uncorrected as a holdout, and report the difference in scrap, assay coefficient of variation and functional yield. Write the first evidence record — model version, limits, approvals, offsets applied — and check that somebody outside the team can read it.
A yield and scrap delta against a holdout, plus a lot record an auditor can read

The order matters

Boundary before model
Writing the admission rule first is not bureaucracy — it determines which data even exists. A material admitted after singulation produces a different trace, a different lot structure and a different failure mode than the same material admitted before it. Teams that model first frequently discover in month three that half their training data came from a route that is about to be changed.
Prediction before correction
A read-only prediction can be wrong in public for a month at no cost, which is exactly what you want while the error bound is being established on real material. A correction that is wrong is a scrapped lot and a suspicious process owner. The published run-to-run control literature is unambiguous that an over-aggressive loop becomes the dominant variance source — see the stability analysis of EWMA run-to-run controllers (opens in a new tab) for the formal version of the argument.
One chamber before the module
A holdout chamber is the only way to say afterwards what the correction was worth, and it is far easier to establish before anyone has seen a result than after. It also limits the blast radius: if the loop misbehaves, half the module is still running the process that worked last quarter.
Evidence during, not after
The lot record — model version, limits in force, offsets applied, who approved what — costs an afternoon of pipeline work in week seven and an unbounded project in year two. If the device is or may become a regulated one, this is the single decision on the whole plan that is genuinely hard to reverse.

Proving the module is actually bridged: formula, source, cadence

Where each metric comes from — the formula, the system that produces it, how often to read it, and the rung at which it starts meaning something.

A metric you cannot name a source system for is an opinion, and bio-facing modules attract opinions. Every measurement below reduces to timestamps, counts and values that the MES, the tool trace layer, the laboratory system or the model-serving log already record — the work is joining them, not creating them. The table is a build sheet: formula, source, cadence, and the rung at which the number first measures something real.

Metric	Formula / read	Source	Cadence	Honest from
Admission compliance	Lots whose bio step occurred at or after the admission point ÷ all lots in the module	MES route history	Per lot	Rung 2
Provenance coverage	Incoming bio-material lots with a complete, joined genealogy record ÷ all incoming lots	MES plus the laboratory certificate system	Weekly	Rung 2
Lot-to-lot variance	Between-lot variance ÷ total variance on the assay result	Assay database joined to lot records	Monthly	Rung 2
Virtual metrology error	Mean absolute error against the off-line assay, computed on held-out material lots only	Assay database plus model serving log	Per assay batch	Rung 3
Prediction latency	Prediction timestamp − process-step timestamp	Serving log plus tool trace	Per wafer	Rung 3
Label join rate	Assay results joined automatically to wafer, trace and lot ÷ all assay results	Label pipeline	Weekly	Rung 3
Correction acceptance	Offsets accepted by the engineer ÷ offsets proposed by the controller	Run-to-run controller log	Weekly	Rung 4
Cross-contamination excursions	Tool excursions traced to a bio-exposed wafer, per quarter	Fault detection and excursion records	Monthly	Rung 2
Evidence completeness	Released lots whose record names model version, limits and approvals ÷ all released lots	Quality system and MES	Per release	Rung 4
Transfer cost	Elapsed days from decision to qualified running at a second tool, line or supplier grade	Project record	Per transfer	Rung 5

Instrumentation build sheet for a bio-facing module. 'Honest from' is the rung at which the metric starts describing a controlled process rather than an aspiration.

The rung transitions themselves are verified by the four measurements below. Each has a threshold that separates the rung beneath from the rung above, and each is readable from telemetry rather than from a self-assessment.

Metric	Rung 2	Rung 3	Rung 4	How to read it
Time to a usable result	Days — the off-line assay	Same shift, predicted	Same shift, and it moves the recipe	Timestamp gap between the process step and a number someone can act on
Where the number lands	A report	A field on the lot in the MES	A bounded controller offset	Which system holds the value the operator actually sees
Model error basis	Training fit	Held-out material lots	Held-out lots plus a live holdout chamber	Ask what material the reported error was computed on
Release evidence	None	Model named	Version, limits, approvals	Open one lot record and read it

Verification metrics for each rung transition. All four are readable from system telemetry, which is the point — a rung claim that depends on someone's recollection is not a rung claim.

Bridged-rung readiness checklist

If you cannot tick all eight, the module is still at rung 2 regardless of how well the chemistry performs. Tick as you go — this list works without JavaScript.

A written admission rule exists per material, with numbers in itCeiling, earliest admissible step, and which toolset the wafer may re-enter afterwards.
Every incoming bio-material lot has a genealogy record joined to wafer identifiersCertificate values as data, not as filed documents.
The measurement the boundary destroyed has a model standing in for itWith an error bound computed on material the model never saw.
The prediction is written into the MES, not shown on a dashboard
Off-line assay results are joined back automatically to wafer, trace and lot
Segregation is enforced by equipment automation, not by a sign on the doorRoute restrictions with a positive-release clean, and an exception log.
There is a fallback recipe one switch away, and it has been exercised deliberately
A holdout chamber or lot stream stays uncorrectedOtherwise no later claim about the correction can be attributed.

0 of 8 ticked

Nothing ticked — and the first move is not a model

A blank list almost always means the module is still at the bench, and that is a normal place to be. Do not start with modelling: start by writing one admission rule and giving the incoming material an identity a system can query. Both are afternoon-scale tasks and everything else on this list becomes possible once they exist.

Instrument one bio bench with us

Failure modes that send a bio-facing module backwards

Maturity here is not monotonic. Four regressions account for almost all of it, and three are silent.

Modules regress, usually without anyone noticing, because the conditions that sustained a rung quietly stopped holding. In bio-facing flows the regressions have a particular character: the inputs are alive, or were, and they change on a schedule nobody in the fab controls. Four patterns account for almost all of the loss.

Likelihood: highImpact: high

The material lot changes and the model does not

A biologically derived input is a harvest. A new supplier grade, a new growth batch or a reformulation shifts the input distribution in a way no calendar-based retrain schedule anticipates, and the virtual sensor degrades over weeks while everyone attributes the drift to the process.

PreventionTie the retrain trigger to supplier lot changes and certificate deltas, not to the calendar, and alert on lot-signature features moving outside their trained range.

Likelihood: lowImpact: high

A bio-exposed wafer re-enters a shared toolset

One routing exception during a busy shift, and organic or ionic contamination reaches a chamber that other products depend on. The excursion is discovered weeks later on somebody else's yield, and the investigation costs more than the bio programme's entire quarter — often ending in the module being suspended.

PreventionEnforce segregation in the equipment automation with a positive-release clean, and monitor the exception log as a leading indicator rather than an archive.

Likelihood: mediumImpact: high

The model becomes an unversioned design input

The prediction quietly starts influencing release decisions before anyone has decided it is part of the device's design history. Two redeployments later, nobody can say which model version held which lot, and the first serious audit question has no answer that a record can support.

PreventionTreat the model version as a released document from the first day it touches a disposition, with an intended-use statement and a validation record attached.

Likelihood: highImpact: medium

Coupon results do not survive the full wafer

A process characterised on quarter-wafer coupons meets edge exclusion, vacuum chucking, radial uniformity and handling for the first time on 200 mm, and the window that looked wide on a coupon turns out to be narrow and off-centre. The programme reads as a chemistry failure when it is a scale-up that was never budgeted.

PreventionBudget an explicit edge and uniformity study before scaling, and treat the first full-wafer run as a characterisation exercise rather than a demonstration.

Three of the four are silent by construction: the output keeps arriving, the wafers keep moving and the records keep filling. That is why the verification metrics in the previous section are worth reading on a cadence rather than at a review — a rising exception rate, a lot-signature feature drifting outside its trained range, or a falling label-join rate each announce one of these regressions weeks before it becomes an incident.

Glossary

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

Admission point: The earliest step in a silicon route at which a given non-standard material can be introduced without being destroyed by the process or contaminating the line. Naming it per material is the founding decision of any bio-facing flow.
Thermal budget: The cumulative temperature-time exposure a wafer receives from a given point onwards. Front-end steps exceed 1,000 °C and back-end interconnect is typically capped near 400 °C, which is why almost all biological material is admitted after passivation.
Coupon: A piece of wafer — often a quarter or a 25 mm square — used for bench experiments. Coupons carry no edge exclusion, no vacuum chucking and no radial uniformity, so coupon results routinely fail to transfer to a full wafer.
Layer transfer: Fabricating devices conventionally, thinning them, releasing them from their original substrate and bonding them to a new one. Every reported bio-substrate carrier device is a transfer story, which makes bond, release and handling the real technology.
Total thickness variation (TTV): The difference between the thickest and thinnest points of a wafer or bonded stack. In a transfer module TTV becomes the critical dimension, because downstream lithography, handling and yield all key off it.
Virtual metrology: Predicting a measurement from equipment trace data instead of measuring it. In a bio module it is not an optimisation — it is the only available substitute for metrology the material has made impossible, and it is only usable with a stated error bound.
Run-to-run control: Adjusting the next run's recipe from the last run's measured or predicted result. Correction must be bounded: the control literature is explicit that an over-aggressive loop becomes the dominant source of variance rather than its remedy.
Lot genealogy: The record linking an incoming material lot — harvest or batch, certificate values, storage history, age — to the wafers it touched. For biologically derived inputs it is the dataset that decides whether a model generalises past this quarter's supplier.
Label latency: The delay between a process step and the arrival of the ground-truth result that grades it. In a bio module the label is destructive, expensive and days late, which makes labels the scarcest resource in the system.
Biological evaluation: The structured assessment of a device's biological safety against the exposed material set, carried out under the ISO 10993 series. It is not a chemistry review of the whole flow — it is anchored to what actually contacts tissue or sample.
Design input: An element of a medical device's specification and design history, governed by design control under ISO 13485 and the EU Medical Device Regulation. A model that holds or releases wafers becomes one, with the versioning and validation duties that follow.
Holdout: A chamber, lot stream or product line deliberately left on the previous process so an improvement can be attributed rather than assumed. In bio modules it is the only defence against crediting a correction for a change in material.

Frequently asked questions

The questions process, equipment and quality engineers ask most often when a biological material is proposed for a silicon flow.

What are bio-substrates in silicon wafer engineering?

They are substrate, interface and template materials of biological origin or biological function that are combined with silicon devices: cellulose nanofibril films, silk fibroin, hydrogels, self-assembled capture chemistry, DNA templates and living cell layers on microelectrode arrays. In every publicly reported device they are joined to the wafer after the front end is complete, because none survives implant, anneal, plasma or high vacuum. The engineering question is therefore not how to build a biological wafer, but where in an ordinary silicon route a given material may safely be admitted.

Can a wafer fab actually process biological materials?

Yes, but only in a segregated module after the last high-temperature step, and admitting one forks the line. Organics, salts and biological material are excluded from shared toolsets because cross-contamination in a chamber surfaces later as a yield excursion on somebody else's product. In practice a bio-exposed wafer is restricted to a defined toolset for the rest of its life, and returning it to general circulation requires a defined clean and a positive release — or is simply not allowed. Enforcing that as an MES routing rule rather than a written procedure is one of the cheapest maturity moves available.

What temperature limit decides where biology can enter the flow?

There is no single number, which is why the admission ledger on this page lists a practical ceiling per material class rather than one figure. As orientation: front-end steps exceed 1,000 °C, back-end interconnect is typically capped near 400 °C, most polymer and protein films stop being themselves somewhere below 200 °C, capture chemistry sits nearer 120 °C, and a living layer has to stay at 37 °C and wet. Every formulation needs its own measured limit on your own tools; the ledger is for orientation and route planning, not for specification.

Where does AI actually help with bio-substrates?

First and most usefully, it replaces measurements the boundary destroyed. A capture monolayer, a hydrated gel or a rough biodegradable film cannot be measured by the inline metrology a fab already owns, so a virtual sensor built from tool trace, ambient conditions and incoming lot attributes stands in for the number the destroyed metrology would have given you. After that it feeds material variance forward into the recipe, drives bounded run-to-run correction, and carries process knowledge across a transfer to a second tool or supplier. Materials discovery is the last of these, not the first.

Is DNA self-assembly going to replace lithography?

No, and the published work is explicit about why. DNA origami assembles in solution, so uncontrolled deposition produces random arrangements — the property that makes it interesting is the same property that makes it unmanufacturable alone. Every demonstration to date uses lithography first: electron-beam-patterned binding sites, dry oxidative etch, controlled buffer chemistry. Kershner and colleagues reported 70–95% of patterned sites occupied with orientation dispersion as low as ±10°, and Gopinath and colleagues scaled placement to 65,536 cavities on a chip. That is directed self-assembly layered on top of lithography, not a replacement for it.

Are biodegradable substrates for chips real?

They are real as published research and not as production silicon. The clearest demonstration is Jung and colleagues in Nature Communications, who reported high-performance flexible microwave and digital electronics on biodegradable cellulose nanofibril paper — including gallium arsenide microwave devices in transferrable thin-film form — and demonstrated fungal biodegradation of the finished electronics. Note the mechanism: the devices were fabricated conventionally and transferred. Nothing was grown on the biological substrate. Volume flexible electronics today bonds thinned die onto polymer, which is the same architecture with a less interesting substrate.

What about biological computing — organoids and neurons on chips?

It exists as research instruments and benchtop units, and it changes nothing about wafer manufacture. The silicon in these systems is a conventional CMOS microelectrode array; the biology is the payload, kept alive by incubation, media handling and sterility control. The organoid-intelligence programme published in Frontiers in Science sets out a research agenda and its prerequisites rather than a product, and companies working in the field publish closed-loop culture experiments. Treat it as a fascinating adjacent field rather than a semiconductor roadmap item — the hard engineering is biological, not lithographic.

How do medical-device rules change an AI control loop?

They turn the model into a design input. If a prediction holds or releases wafers for a device that touches a patient or produces a diagnostic result, that model sits inside the design history: it needs a version, an intended-use statement, a validation record against held-out material, defined limits, and a lot trail reconstructable years later. Biological evaluation under the ISO 10993 series is anchored to the exposed material set, and design control under ISO 13485 and the EU Medical Device Regulation governs how changes are released. Built into the pipeline this costs little; retrofitted it is a project.

How much data do you need before a virtual metrology model is trustworthy here?

Fewer rows than people expect and more material lots than they have. The binding constraint is not sample count but coverage: a model fitted on three supplier lots will look excellent and fail on the fourth, because between-lot variance usually dominates. A workable starting point is twelve months of joined trace, lot and assay records spanning at least six or seven distinct material lots, with the error reported on lots the model never saw. Published benchmarks are a useful reality check — a mass-production CVD virtual-metrology study reported best accuracy around 0.70.

What does it cost to move one rung on this ladder?

The rung 2 to rung 3 move is roughly a quarter: one process engineer owning the module, one machine-learning engineer, one integration engineer and equipment-engineering support for trace access. The dominant cost is rarely the modelling — it is trace access and the assay join, because those touch systems owned by other teams. Rung 3 to rung 4 is longer, typically twelve to twenty-four months, and most of that time is evidence and change control rather than engineering. Rung 5 is not a project at all; it is a standing discipline measured by transfer cost.

How is this different from ordinary fab AI work?

Four things differ. Labels are destructive, expensive and days late, so the label pipeline matters more than the model. Incoming variance is biological, so a lot is a harvest rather than a specification and provenance features are mandatory. The boundary forks the line, so routing and segregation are part of the solution rather than the environment. And the output may be a regulated device, so a model in the release path is a design input. Everything else — trace capture, virtual metrology, bounded run-to-run control, holdouts — is ordinary fab practice applied to an unusual module.

Should we start with materials discovery or process control?

Process control, almost always. Materials screening answers a question you probably are not blocked on, and it needs exactly the data infrastructure you do not yet have. The blocked question in nearly every bio-facing module is that the step cannot be observed between the recipe and the assay, and that is a virtual-metrology problem solvable in a quarter with data the tools already produce. Build that, and the provenance and label pipelines it forces into existence are the same ones any later screening or discovery work would have required anyway.

About the author

Atomic Loops Engineering

Industrial AI practice

Atomic Loops builds production AI systems for wafer fabs and process manufacturers — trace pipelines, virtual metrology, drift and matching models, and run-to-run control integrated into the MES and the equipment layer rather than delivered as dashboards.

· Trace-to-metrology pipelines built against live equipment data
· Virtual metrology and bounded run-to-run control in production modules
· Integration-first delivery: MES write-back, holdouts, fallback recipes, evidence records
· 22 cited sources on this page

Sources

AI innovation and autonomous tool fleets in wafer fabsThe sibling problem on standard materials: why nominally identical process tools diverge, and the control loops that close each mismatch.AI adoption and the maturity curve in silicon wafer engineeringHow fab AI programmes progress from isolated experiments to production capability, and the KPIs that prove it at each stage.Silicon wafer engineering practiceProduction AI systems Atomic Loops builds for wafer fabs: trace pipelines, virtual metrology, drift and matching models, and control-loop integration.

Find out exactly where one bio-facing module stands — then what to do about it

We run the assessment with your process, equipment and quality engineers, check the answers against your real trace, lot and assay data, and leave you with a costed 90-day plan for the weakest dimension. You keep the analysis and the plan whether or not we build anything.

Book a module review See our silicon wafer engineering work

Published 20 August 2026 · Last updated 20 August 2026

Key takeaways

Abbreviations used on this page

Score one bio-facing module

What bio-substrates mean in a silicon flow

The five rungs in detail

Coupon

Characterised

Bridged

Qualified

Transferable

Where bio-facing programmes actually sit today

The admission ledger: what each material costs the line

Real today, published research, and still speculation

Where AI lands in a bio-facing silicon flow

What the boundary looks like in public

The reference architecture for a bio-facing module

Substrate and reagent supply

The admission boundary

Bio module process control

Assay and label pipeline

Evidence and change control

A 90-day plan: giving a functionalisation module a measurement

The order matters

Proving the module is actually bridged: formula, source, cadence

Failure modes that send a bio-facing module backwards

The material lot changes and the model does not

A bio-exposed wafer re-enters a shared toolset

The model becomes an unversioned design input

Coupon results do not survive the full wafer

Glossary

Frequently asked questions

What are bio-substrates in silicon wafer engineering?

Can a wafer fab actually process biological materials?

What temperature limit decides where biology can enter the flow?

Where does AI actually help with bio-substrates?

Is DNA self-assembly going to replace lithography?

Are biodegradable substrates for chips real?

What about biological computing — organoids and neurons on chips?

How do medical-device rules change an AI control loop?

How much data do you need before a virtual metrology model is trustworthy here?

What does it cost to move one rung on this ladder?

How is this different from ordinary fab AI work?

Should we start with materials discovery or process control?

About the author

Sources

Related reading

Find out exactly where one bio-facing module stands — then what to do about it