Instrumentalism versus realism in semiconductor manufacturing

The objection that should win

Start with the strongest version of the case against this whole page. A fab does not run on vibes about electrons. It runs on process models: SPICE parameters for transistor behaviour, deposition-rate equations for chemical vapour deposition, etch-rate models for plasma chemistry. These models capture structure — physical relationships between voltage, current, temperature, chamber pressure — that persists across nodes, across tool generations, across fabs on different continents. A yield engineer who has just qualified a new lot at 40 nanometre gate length is relying on the same drift-diffusion equations that were reliable at 65 nanometre. If that structure is real and stable, then a model built on it does not need to keep listening. It needs to be correct once. This is structural realism, applied to a clean room: the claim that what a good process model captures is not surface correlation but the actual relational skeleton of the physics, and that skeleton does not decay just because someone stops measuring it. On this view, continuous intake — the fab's endless stream of inline metrology, yield telemetry, equipment logs, materials-lot records — is operationally convenient but not epistemically necessary. The physics was always going to hold. You are just checking your arithmetic.

That is a serious position, argued by serious people, and it deserves to be taken at full strength before any rebuttal.

Where the objection is right

It is right about the existence of durable structure. Fick's laws of diffusion do not get renegotiated because a new photoresist ships. The relationship between overlay error and edge-placement error does not change character between tool vendors. A yield engineer who distrusted every physical relationship on principle would be paralysed, rebuilding Ohm's law from scratch every shift. Structural realism earns its keep here: some of what a fab's models represent is not going to be falsified by next month's wafer.

Where it fails: "in its regime" is not a free clause

The failure sits in three quiet words: "in its regime". Every structural claim in semiconductor physics comes attached to boundary conditions — a doping range, a temperature window, a film-thickness band outside which the equation stops being the equation and becomes an approximation quietly going wrong. Knowing whether today's lot is inside that boundary is not something the structure tells you. It is something only measurement tells you.

This is precisely the shape of the fab's characteristic failure. A lot excursion — a batch of wafers drifting out of process specification — is caught at final test, dozens of steps and often days after the step that actually caused it. The physical model governing, say, chemical-mechanical polishing removal rate was not wrong in principle. It was applied to a chamber whose consumable pad had aged past the interval the model was calibrated against, or a slurry lot with a particle-size distribution shifted outside the qualified range. The structure held. The regime had moved. Nothing in the frozen model could say so, because the frozen model has no channel back to the chamber's present state — only to the state it was fitted on.

The parallel case from physics is exact and instructive. Newtonian mechanics did not fail because its structure was fictional. It failed to predict Mercury's perihelion precession because Mercury's orbit sat at the edge of a regime nobody had checked closely enough, and it took a century of continued astronomical observation — not renewed confidence in the theory — to reveal the anomaly and eventually license general relativity. Confirmation was never the resting state. Monitoring was the only way anyone found out the resting state had ended. A fab's process model is in the same position at wafer-lot scale: durable structure, undisclosed boundary, and no way to know the boundary has been crossed except by comparing what the model predicts against what the inline metrology is reporting right now.

The second objection, and its narrower truth

There is a fallback position, weaker but more practical, that concedes drift is real and answers with schedule rather than structure: recalibrate periodically. Metrology tools themselves are certified this way — a critical-dimension scanning electron microscope is trued against a reference standard on a fixed interval, not watched continuously. Extend the same logic to process models: retrain the etch-rate model quarterly, requalify the CVD recipe against a golden wafer once a month. Continuous intake becomes an engineering parameter, not a philosophical necessity. Choose a tight enough interval and you have covered the same ground as "everything, continuously", just in instalments.

This works exactly where drift is known and bounded. A reference standard's material properties do not misbehave; a scheduled certification interval is safe because someone already measured the drift rate and it sat still. Semiconductor process drift does not sit still. A single tool's chamber can shift behaviour after a preventive-maintenance event, after a consumable-lot change, after a facilities excursion in cooling-water temperature that nobody logs as a process variable. These drift sources are not periodic; they are event-driven and their magnitude is not known in advance. Under those conditions the honest recalibration schedule is not "monthly" but "whenever the residual between predicted and measured output moves past a threshold" — and detecting that requires watching the residual all the time, which is continuous intake by another name. Scheduled recalibration is not a rival to continuous monitoring. It is what continuous monitoring degenerates into once someone has already measured the drift rate and found it slow and stable — a finding itself only available from a period of continuous watching.

Where this lands: the yield engineer's actual epistemic position

Put the two survivals together and the residual claim is narrow but load-bearing. Structural realism is correct that some equations are permanently true of the physics. Instrumentalism is correct that no frozen application of those equations to a live fab carries determinate content, because "this reading is valid" is itself an empirical claim, checkable only against present observation, not against the equation's derivation.

This is the actual daily condition of a yield engineer. Their job is not to defend the drift-diffusion equation. It is to hold, at any moment, hundreds of live beliefs of the form "die at this reticle field on this lot are within spec because the inline metrology at steps 14, 22 and 31 confirm it" — beliefs that must carry their own provenance, because when a lot excursion surfaces at final test, the first question is which upstream reading licensed the assumption that turned out false. A model that only stores the physics, without the chain of measurements that says the physics still applies to this chamber on this shift, gives the engineer nothing to trace back through. That tracing is the entire diagnostic task: not "is the equation true" but "at which step did the observed world stop matching what the equation, applied here, implied."

The narrow claim that holds

None of this makes the Large Universe Model a shipping tool on any fab floor; it names a standard, not a product. What it names, on this axis, is the terminus already implicit in the fab's own best practice: a Large Language Model corresponds to the process equation fitted once and deployed indefinitely, calibrated on a distribution that recedes with every wafer run afterward. A Large World Model corresponds to a within-lot check — metrology consulted at the moment of processing, trustworthy for that episode, reverting to the frozen assumption the instant the lot clears the tool. A Large Universe Model is what you get if you take the fab's inline metrology, yield telemetry, equipment logs and materials-lot records and refuse to let any of them go stale — held as revisable beliefs, each tagged to the observation that licensed it, each subject to revision the moment its residual moves. That is not an upgrade layered onto sound process physics. Given that no frozen application of physics to a live tool is self-certifying, it is the only version of the argument that does not eventually get caught by final test.