Autonomous tool fleets in silicon wafer engineering: from chamber matching to a self-tuning fab

Stage 1 is not a data problem in the way people expect. The fab is not short of measurements — it is drowning in them. Every tool emits SECS/GEM events, every carrier is tracked, every lot has a genealogy in the MES, and the metrology systems produce numbers all shift. What is missing is any representation of the fleet as an object. Chamber 3 of etcher B is a row in a maintenance system and a line on a chart; it is not a member of anything.

The consequence is that all comparison is human. A process engineer who has owned a module for four years knows which chamber runs a shade fast after a wet clean and compensates for it, and that knowledge is genuinely expert and completely undocumented. When that engineer moves to another area, the compensation goes with them and the fleet quietly widens. Fabs describe this as “losing tribal knowledge”; structurally it is a fleet with no persistent state.

This stage is cheap to leave and expensive to sit in, because the cost is invisible. It shows up as guardband — margin subtracted from the specification to absorb variation nobody has isolated — and guardband is spent silently. Nobody files a report saying the fab gave away twenty per cent of its process window to chamber-to-chamber spread. It simply appears as a tighter recipe, a slower process, or a layer that will not yield on the older toolset.

Stage 2 is where most fabs are, and it looks far more finished than it is. The dashboards are genuinely good: a process engineer can pull up six chambers on one axis, see the spread, and act. Fault detection catches trace excursions before they reach metrology. From the outside this reads like a controlled fleet. What makes it stage 2 is that every correction is an interrupt on a human being.

The specific failure is latency of decision, not latency of data. The trace is available seconds after the run; the metrology arrives hours later on a sampled basis; the decision to change a chamber's offset happens the next morning, in a meeting, after somebody notices. Between the drift starting and the offset moving, the fleet has produced material at the old setting. On a high-volume layer that gap is measured in hundreds of wafers, and the fab absorbs it as normal variation because it always has.

There is a second, subtler cost. Because correction is manual and effortful, engineers economise on it by restricting the problem: dedicating a chamber to a product, narrowing the qualification matrix so fewer tools can run the difficult layer, or pinning a hot product to the best-behaved chamber. Each of these is a rational local decision and each one removes flexibility from the fleet. Published simulation work on recipe restriction found the effect can be dramatic — more than a 40 per cent increase in fab cycle time from limiting which recipes tools are enabled to run.

Stage 3 is the first rung where the fleet is a real object with real state. A difference score exists per chamber, it is computed on a cadence, it has thresholds, and crossing a threshold produces an action rather than an agenda item. The engineering effort is unglamorous: aligning traces so that the same parameter means the same thing on tools from different vendors, joining them to wafer genealogy from the MES, and threading the controller so that the correction applies where it belongs.

The threading question is where most stage-3 programmes are won or lost. A run-to-run controller keeps a separate estimate — a thread — for each context it believes is distinct: product, layer, tool, chamber, and often reticle or consumable set. Thread too coarsely and you average away the chamber difference you are trying to correct. Thread too finely and each thread sees so few runs that the estimate is noise, which practitioners call thread explosion. Fleet-level models earn their keep here by letting sparse threads borrow strength from the fleet, which is exactly the transfer-learning framing published equipment-matching research has been exploring.

The other stage-3 discipline is knowing what the reference is. Published tool-to-tool matching work is explicit that the traditional approaches depend on static configuration data or on a golden reference that is hard to obtain in a commercial line — and a golden chamber is a moving target, because the reference drifts too. Fabs that reach a durable stage 3 usually stop nominating a golden chamber and start comparing every chamber against the fleet's own distribution, which has no single point of failure.

At stage 4 the fleet stops waiting to be measured. Virtual metrology predicts the controlled parameter from the tool's own trace plus upstream context, so the run-to-run loop has an input for every wafer rather than for the sampled minority. The published case for this is quantitative: a cross-benchmark of machine-learning virtual metrology for chemical vapour deposition reported that the accuracy achieved would correspond to a 70 per cent reduction in CVD processing variation, alongside a reduced need for physical metrology. More recent work has moved to graph-attention architectures that model dependencies between process parameters and film layers directly.

The second half of stage 4 is maintenance. A fleet whose chambers are matched but whose PM windows are fixed by calendar or wafer count is fighting itself: the PM is the single largest scheduled perturbation to a chamber's state, and putting it in the wrong place either wastes chamber life or lets a chamber run past the point where its behaviour leaves the model's experience. Fabs that get this right schedule PM as a joint decision with production, against predicted degradation and against fleet capacity, rather than as a maintenance-department calendar.

There is a discipline here that is easy to skip and expensive to skip: re-planning physical metrology around the model. Virtual metrology is calibrated by measured wafers, and the temptation once predictions look good is to cut sampling to bank the metrology capacity. Cut it in the wrong places and the model loses the very data that keeps it honest. The correct move is to redistribute sampling — fewer routine measurements, more measurements where the prediction is least certain — which requires the uncertainty estimate to be real rather than decorative.

Stage 5 is narrower than the phrase suggests. It is not an unattended fab; it is an enumerated set of fleet decisions — which chamber a lot goes to, how a furnace batch is composed, when a PM window opens, how far an offset may move — executing without a person inside stated bounds, with everything outside those bounds escalating. Layers under automotive or medical qualification are frequently and correctly held at stage 4 permanently, because the change-control obligations are stricter than the engineering risk.

The engineering at this rung is largely a solved problem in the literature. Published work has demonstrated deep reinforcement and self-supervised learning scheduling a modern semiconductor manufacturing model more efficiently than the hierarchical dispatching strategies fabs typically use, reducing order tardiness and time to completion. The hard part is not the optimiser. It is the artefact that says what the optimiser is permitted to do, and proving to a customer or an auditor that it stayed inside it.

That artefact is the qualification envelope, and treating it as configuration is the most common way stage 5 unwinds. Fabs have a strong cultural instinct here worth borrowing rather than fighting: copy exactly, the discipline of freezing a process and replicating it identically, exists precisely because unrecorded change is how yield is lost. A self-tuning fleet does not abandon that instinct — it moves the frozen thing up a level. What is frozen is the envelope, reviewed and versioned like code; what moves is the offset inside it.