Team NANO · AI Development Lifecycle Hackathon
An autonomous metrology agent that knows what it does not know. It estimates reality from a biased simulation prior plus sparse measurements, quantifies where it is uncertain, and selects the most valuable die to measure next.
A model can score well on the data you have and still be wrong about the wafer. Full inspection of millions of dies is impossible, so every dataset is a sample, and every sample carries a bias the metric does not show. High accuracy is not the same thing as confidence about reality.
Constraint 01
Only a small fraction of dies can be measured. The rest is inference, whether or not anyone calls it that.
Constraint 02
Simulation covers the whole wafer but is systematically off. Measurement is correct but covers almost nothing.
Constraint 03
SEM and TEM see a very narrow field of view, and sampling plans tend to revisit convenient locations.
Constraint 04
Process and metrology tables arrive with gaps, so the feature space itself is only partially observed.
NANO reframes the AI Development Lifecycle (AI DLC). Instead of training on the data that happens to exist, it actively acquires the data needed to understand reality — under a fixed measurement budget.
Maximize knowledge of reality per measurement.
Schematic only. These fields come from a deterministic formula inside this page, written to explain the loop. They are not WM-811K wafers and not benchmark output.
Observe → Estimate → Select → Measure → back to Observe, once per unit of measurement budget.
Read the sparse set of real measurements available so far, plus the simulation prior for every die.
Fit a spatial model of reality and produce three maps: prediction, uncertainty, and prior-vs-reality gap.
Score every unmeasured die with the acquisition function and pick the argmax as the next measurement.
Reveal the hidden real value at that die, append it to the observation set, and spend one unit of budget.
Repeat until the measurement budget is exhausted. Every iteration is logged, so the decision trace is auditable.
It does not only predict. It quantifies uncertainty and decides what should be measured next. That decision — not the interpolation — is the contribution.
NANO
0.3479failure probability
± 0.0524
↓ lower is better
Final Balanced MAE after the full budget
Random baseline
0.3812failure probability
± 0.0527
↓ lower is better
Same budget, same initial observations
Grid baseline
0.3714failure probability
± 0.0494
↓ lower is better
Spatially uniform selection
Prior error removed
20.5%
Biased simulation prior vs corrected estimate
12 wafers × 5 seeds · measurement budget 60 · Balanced MAE ↓ lower is better · generated from results/benchmark_summary.json. Full results →
Separated from Random, Grid at 95% confidence, over the paired episodes.
Every number on this site is generated from a single results file in the repository — the run against WM-811K, 12 wafers × 5 seeds. Nothing is hand-written into the pages, so a page whose artifact is missing says so instead of inventing one.
Clone the repository, run the benchmark, and regenerate every figure on this site.
Run the experiment yourself