Somewhere on one of your tools, a film is growing right now whose composition nobody will know until next week. The run will end, the wafer will join the characterization queue, and the first hard fact about the material will arrive days after every decision about it has already been made.
The industry has spent several years pursuing the promise of AI for materials discovery, and this effort is warranted. Screening and structure prediction now surface promising candidates faster than any lab can grow them. But discovery answers the question "what material?" The question that new technology rests on is "how do we make it, every run?", and every advance on the first question widens the gap to the second.
Materials enablement is our name for the work of closing that gap, the same gap our thesis describes: applying AI to the journey from “works once” to “works every time,” built on the process data a lab or fab already generates.
No such system exists off the shelf, because the two disciplines that might have built it were aimed elsewhere. Statistical process control assumes a mature, settled, high-volume process. Discovery AI points upstream, at the candidate. Between them sits the actual decades of scale-up and uncertainty, still run on trial and error and the memory of whoever has been with the tool longest.
Four needs define the missing system: measurement, understanding, detection, and action. Each one unlocks the next, so this series takes them in order. Measurement comes first because everything else depends on it.
A team cannot learn from a run it never saw, catch drift it never measured, or steer a process it cannot observe. Everything downstream in this series rests on one capability: a trustworthy estimate of the state of the material on every run, arriving fast enough to act on. How fast depends on the process and the signals it emits, and the direction of travel is toward an estimate available while the run is still live.
Today's Ceiling: Late, Partial Signal
The estimate most teams actually have is retrospective. Ex-situ characterization is slow, often destructive, and samples a sliver of the material: a few checkout wafers per campaign, a few sites per wafer, with results hours, days, weeks, or months behind the run. Meanwhile the properties that decide success (composition, thickness, stress, defect density) have no instrument that reads them mid-growth.
The chamber does stream varied data while the film grows. Diffraction, spectroscopy, reflectometry, pyrometry, dozens of tool-trace channels. But these are proxies, raw signals correlated with the properties you want in ways that nobody has fully written down.
Teams bridge the gap themselves where they can. A staff engineer fits a calibration tying one proxy to one property, and it holds near the center of the process window where it was built, then degrades toward the edges, which is exactly where new materials development happens. The obvious upgrade, a proper machine-learning model, runs into arithmetic: classical ML wants hundreds or thousands of labeled runs, and a team scaling a new material has a few dozen.
None of these constraints comes from a shortage of data. A materials process tool produces more data each month than its team can fully analyze; what the data lacks is interpretation. And until it gets interpreted, the characterization queue sets the speed limit of the entire program. Experiments get batched to fit the queue. Decisions wait on it. If structural data takes two weeks to come back, then two weeks is the floor under your learning rate, whatever pace the tool itself could sustain.
The Solution: Building a Live Estimate of Material State
The way out is to change the job ex-situ metrology does. Atomscale’s physics-informed models can fuse the sensor streams a tool already emits into a live estimate of material state. The concrete forms are virtual and computational metrology and learned composition metrics: models that infer the properties you care about from the signals you already collect, continuously, while the run is in progress.
Physics makes the approach workable from the earliest stages of development when volumes are low. Because the physical constraints are built into Atomscale’s model, it needs only reference runs to calibrate, and a handful will do. That one fact moves the technique out of the expensive high-volume labeling regime where classical ML lives and into the regime of real-world early process development: low volume, high mix, processes still being learned.
Ex-situ characterization keeps its place in this picture, with a different job. It becomes the anchor, the ground truth that validates and periodically re-grounds a model running underneath every run.
The Unlock: Faster Cycle Times, Faster Scale-Up
Measurement becomes continuous: every run, every wafer, every second, with material state available as a live signal during growth. That signal is the core input to every later pillar in this series.
The first return shows up in cycle time. Development stops waiting on the queue, because the queue no longer holds the only copy of the truth. The second return is larger: measurements that never existed become routine. One compound semiconductor team we work with needed within-run wafer composition data they could not get at any turnaround time. A learned metric, calibrated on a few reference runs, now delivers it live, with a tenfold accuracy gain over their previous best estimate (RMSE near 1.0 down to about 0.1) against the same ex-situ reference.
That second category is the one worth a leader's attention, because the two returns land on different lines of the business. Continuous versions of yesterday's measurements make a program faster: fewer months to qualification, fewer wafers spent learning, a development budget that buys more experiments. That is a cost line. Measurements that never existed change the unit process itself, on every run, and that reaches the revenue line: higher yield when the process moves to volume, a material held to spec closely enough to sell, a product window a competitor working at queue speed will not make. The first makes the program you already have cheaper to run. The second makes viable a program that was not.
Through the rest of our series on materials enablement, we’ll see how teams can practically use Atomscale to recover how processes actually behave using past runs, detect crucial moments when runs change, and eventually steer runs as they grow.