Our Head of Product on why sequential, multi-modal, uncertainty-aware models — not LLMs — are the right substrate for SMTO.
The first wave of foundation models was built for language. That made sense. Language is the interface through which people document work, retrieve knowledge, explain decisions, and coordinate action. Large language models have become powerful reasoning surfaces for enterprise software because so much enterprise work is represented in text.
A fab is different.
Semiconductor manufacturing technology optimization does not live primarily in language. It lives in sequences, traces, images, wafer maps, recipes, metrology, tool states, process windows, equipment histories, product context, engineering intent, and lagging yield signals. It lives across thousands of coupled steps where the important question is rarely “what does this document say?” and more often “what changed, where did it propagate, how confident are we, and what should we run next?”
That is why foundation models for fabs cannot simply be LLMs pointed at manufacturing data. They require a different substrate: sequential, multi-modal, uncertainty-aware models built for physical systems.
The distinction matters. An LLM can help an engineer ask better questions, summarize a process note, retrieve a relevant procedure, or explain a hypothesis in natural language. Those are valuable capabilities. But they are not sufficient for SMTO. A fab foundation model has to understand that a process flow is not a bag of tokens. It is an ordered, constrained, path-dependent system. What happens at deposition can matter several steps later at etch. A small chamber drift can appear downstream as a yield signature. A tool event, a metrology outlier, and a spatial wafer pattern may only become meaningful when interpreted together.
That requires sequential modeling. The model has to reason over process history, not isolated data points. It must represent the manufacturing route, the timing of measurements, the dependency between steps, and the fact that cause and effect in a fab are often separated by time, tools, and layers of abstraction. Without sequence, the model sees snapshots. Engineers need trajectories.
It also requires true multi-modality. Fabs do not produce one kind of data. They produce images from inspection tools, time-series traces from equipment, spatial wafer maps, recipe parameters, process metadata, engineering annotations, simulation outputs, and yield outcomes. Each modality carries partial information. None is complete on its own. A useful fab foundation model must learn across these modalities at the same time, aligning them into a shared representation of the product, process, tool, and outcome.
Most importantly, it requires uncertainty awareness. In consumer AI, a confident wrong answer is annoying. In the fab, it is operationally dangerous. SMTO decisions affect wafers, cycle time, yield, engineering bandwidth, and customer commitments. A model must know what it knows, what it does not know, and which experiment would reduce uncertainty fastest. That is why Bayesian methods, calibrated digital twins, and active learning are not optional add-ons. They are core product requirements.
This is the architectural gap between LLMs and fab foundation models.
LLMs are excellent interfaces. They are not the industrial reasoning substrate. The right architecture uses language models where they are strongest: as the interaction layer that lets engineers express intent, interrogate results, and navigate complex workflows. Beneath that surface, the system needs models that are grounded in semiconductor process data, physics, uncertainty, and sequence. It needs digital twins that can simulate where data is sparse. It needs reusable workflows that encode how classes of fab problems are solved. It needs agents that can assemble the right path from question to analysis to recommendation.
For SMTO, the foundation model is not one model. It is an integrated stack.
At the bottom is the physics and data foundation: curated, contextualized, governed fab data connected to process knowledge and simulation assets. Above that are sequential and multi-modal models that learn representations across wafers, lots, tools, steps, images, traces, and outcomes. Around those models are uncertainty quantification and active-learning loops that decide where the next measurement, simulation, or split is most valuable. On top are digital twins and master workflows that convert raw capability into repeatable engineering applications. The LLM sits at the surface, giving engineers a natural-language way to use the system without forcing every interaction through code or dashboards.
This architecture changes the product experience.
Instead of asking an engineer to choose from a fixed menu of analytics tools, the system can accept intent: investigate this excursion, adapt this workflow to a new node, compare these process windows, explain why this wafer map pattern emerged, recommend the next experiment. The agent translates that intent into a sequence of executable steps: find the relevant data, align the modalities, run the right model, quantify uncertainty, test hypotheses against the digital twin, and return a decision-ready answer.
That is the difference between AI that comments on fab work and AI that participates in fab work.
The industry does not need another chatbot for manufacturing. It needs a foundation model architecture that respects the structure of manufacturing itself: sequential because process history matters, multi-modal because no single signal is enough, uncertainty-aware because decisions must be trusted, and physics-grounded because fabs operate in the real world.
LLMs will be part of that architecture. They will make it usable. But they will not be the architecture.
The future of SMTO belongs to systems that combine language, physics, data, uncertainty, and action into one deployable intelligence layer. That is the substrate fabs need if AI is going to move from isolated use cases to compounding process learning across products, nodes, tools, and generations.