Time crystals and visionary AI in silicon wafer engineering: separating physics from speculation

Rung 1 is not a fab without process control. It is a fab whose process control is blind to time. Statistical process control is running, monitor wafers are being measured, charts are being reviewed — and every one of those instruments treats a wafer processed two hours after a wet clean and a wafer processed two hundred wafers later as samples from the same population. They are not, and every process engineer in the building knows they are not. The knowledge simply lives in people rather than in data.

The cost of that blindness is paid in three currencies and it is rarely totalled. The first is monitor wafers and qualification time, spent generously because nobody can justify spending them precisely. The second is availability: PM intervals are set conservatively, because the alternative to a conservative interval is an excursion nobody can predict. The third and largest is argument. When yield moves, the conversation about whether it was the clean, the new focus ring, the material lot or the ambient is conducted in anecdote, and it is settled by seniority rather than by evidence.

This is a cheap rung to leave and an expensive rung to stay on. Leaving it requires no new tools, no new sensors and no new control loop — only a join. Take one module, take six months of metrology results, and plot them against the cycle position of the wafer that produced them. That single chart typically ends a running argument in the process area within a week, and it is the entire prerequisite for everything on rungs 2 to 5.

Rung 2 changes the conversation before it changes the process. Once the sawtooth is on a chart with a cycle axis, arguments that ran for years end in an afternoon — the film does climb across the clean cycle, the removal rate does decay with pad hours, the implant dose does walk with source life. Process engineers usually find that their intuitions were right in direction and wrong in magnitude, which is exactly the useful discovery, because magnitude is what a control loop needs.

What rung 2 does not yet have is authority. The correction still travels through a human: an engineer notices the drift, decides on an offset, files a change, waits for approval, and applies it at some point that has more to do with the meeting calendar than with the process. The lag between observing the drift and acting on it is typically measured in shifts or days, while the drift itself moves in wafers. This is the defining gap of the rung — the fab has become an excellent observer of a phenomenon it still cannot track in real time.

There is also a characteristic rung-2 temptation that deserves naming early. Having proven that drift correlates with cycle position, the obvious move looks like shortening the cycle: clean more often, PM more often, hold the process closer to centre. It works, and it is expensive. Availability falls, monitor-wafer consumption rises, and the fab has bought uniformity with capacity at an exchange rate nobody calculated. The alternative — teaching the control loop about the cycle instead of shrinking the cycle — costs engineering time rather than capacity, and it is what rung 3 is.

Rung 3 is the first rung where the fab holds phase rather than merely watching it. The mechanics are unglamorous and well documented. Virtual metrology gets cycle position as a feature, which typically removes a large, systematic component from its error and — more usefully — makes its uncertainty honest, because the model can now express that it knows less about a wafer three hours after a clean than about one mid-cycle. Published work on virtual metrology in mass production and on uncertainty-aware estimation from small samples both point the same way: the value is less in the point prediction than in a calibrated confidence a controller can act on.

The run-to-run controller is the second half. A standard exponentially weighted moving average filter is designed to track a slow drift, and a wet clean is not a slow drift — it is a step. Feeding a step into a filter tuned for drift produces exactly the pathology the control literature describes: a correction that overshoots, then a correction of the correction, then a lot or two of ringing before the loop settles. The fix is not more aggressive tuning; the stability analysis of EWMA run-to-run control makes clear that raising the gain to chase a disturbance moves the loop toward oscillation rather than toward centre. The fix is to tell the controller that a reset happened, and to give it a prior for what usually follows.

The organisational shift at rung 3 is that cadence stops being a maintenance decision and becomes a control decision. When the trigger for a clean is a predicted state rather than a date, the maintenance schedule, the dispatch policy and the process model are all reading the same variable. That is a genuine coordination change and it needs a genuine owner, because the first time a condition-based trigger asks for a clean on a Friday afternoon during a hot-lot push, somebody has to decide — and that decision should be a written policy, not a corridor conversation.

Rung 4 is where the unit of management changes from the deviation to the recovery. A rung-3 fab can tell you its process capability; a rung-4 fab can tell you that a focus-ring change on a given etch module costs, on average, a specific number of wafers before critical dimension is back inside its normal band, that the number has fallen over the last two quarters, and what the intervention was that moved it. That is a different kind of claim, and it is the one that matters for anything time-critical, because a fab's real variability is dominated by its perturbations rather than by its steady state.

The second rung-4 capability is propagation. A disturbance at deposition does not stay at deposition; it arrives at etch as an incoming variation, and whether the etch loop attenuates it or amplifies it is a property of how the two loops are tuned relative to each other. Most fabs discover coupled-loop amplification the same way — as a yield signature that no single module owns, three weeks after the fact. Making propagation a monitored quantity turns that from a forensic exercise into an alarm, and it is the point at which cross-module control stops being a slogan.

This is also the rung where certification stops being background noise and becomes design input. Semiconductor manufacturing runs inside a dense frame of equipment, data and quality standards — the SEMI standards estate for equipment reliability and data acquisition, automotive quality management for parts that go into vehicles, and customer-specific qualification for anything that ships to a named account. Every one of those frames asks the same underlying question of an automated recovery: who authorised it, on what evidence, and can you reconstruct the decision later. A rung-4 fab answers by producing the evidence as a by-product of the recovery procedure. A fab that automates recovery first and documents it afterwards discovers during an audit that it has automated something it cannot defend.

Rung 5 is narrow, and its narrowness is the point. Roughly three per cent of lines are anywhere near it, and for most fabs that is the correct answer rather than a gap to close. What makes the rung worth describing is that its requirements are not exotic in kind — they are the rung-1-to-4 disciplines carried further than a logic or memory product would ever justify. The published 300 mm silicon spin-qubit work is unambiguous on this: the devices were made with immersion and extreme-ultraviolet lithography, standard high-volume process modules and production-level process control, on lines that were built for transistors.

The specific tolerances are instructive because they are numbers rather than adjectives. A 300 mm silicon MOS qubit platform reported uniform critical dimensions down to 30 nm with a standard deviation of about 1.6 nm. An isotopically purified silicon-28 epitaxial layer was grown on 300 mm substrates at better than 99.992 per cent purity, to the same quality requirements as the natural-abundance epilayers used in the same pre-industrial CMOS facility. A foundry-fabricated two-qubit unit cell has been reported above 99 per cent fidelity in all operations. None of these is a physics breakthrough in the manufacturing sense; each is process control held to a tolerance a conventional product does not need.

The honest caution at rung 5 is about demand, not physics. The market that would consume coherence-grade capacity is small, research-led and moving on its own schedule, and a fab that builds capacity against it is taking a market bet dressed as a technology bet. The defensible posture is optionality: keep rungs 1 to 4 paying on the products you actually ship, hold a research partnership or a shuttle arrangement so the capability is reachable, and require any proposal to state its demand bound as explicitly as its technical one.