Dataset and Experimental Design

Source

The MVP uses the WM-811K wafer map dataset — 811,457 real wafer maps collected from semiconductor manufacturing, released alongside Wu, Jang and Chen, "Wafer Map Failure Pattern Recognition and Similarity Ranking for Large-Scale Data Sets", IEEE Transactions on Semiconductor Manufacturing, 2015.

Canonical distribution	MIR Lab public datasets (LSWMD.pkl)
Commonly used mirror	Kaggle: WM-811K wafer map
Terms of use	Set by the distributor, not by this project. Check the licence on the page you download from before redistributing.
Included in this repository	The raw distribution is never committed here. The dataset is downloaded locally; what is versioned is a derived subset index, plus the twelve evaluation wafers embedded in the playground so it can run without a server — die coordinates and pass/fail labels for 18,241 dies, about 0.002% of the 811,457 maps.

Implementation status

The subset builder is python -m nano.subset and the harness is python -m nano.benchmark. Any field below written as recorded at run time is deliberately empty on this page: it is filled from the artifacts a run produces, not typed in by hand. The published run used the twelve wafers named in data/subsets/wm811k_eval.json, drawn from a local LSWMD.pkl by that builder.

Without a local copy of the dataset the pipeline still runs, on generated stand-in wafers (--synthetic). Those are not WM-811K: the results file records dataset.name as synthetic-wafers with a note, every command prints a warning, and the results table on this site prints the wafer source. See Honest Scope.

Wafer map semantics

Each wafer map is a 2-D integer array on the die grid:

Value	Meaning	Used as
0	Outside the wafer / no die at this grid position	Masked out, never predicted, never measured
1	Die present, pass	Reality value 0.0
2	Die present, fail	Reality value 1.0

So the quantity NANO reconstructs is a binary per-die failure indicator on the in-wafer die grid. Reconstruction error is measured only over in-wafer positions.

Preprocessing rules

1. Drop wafers whose die grid is smaller than the minimum usable size, because a sparse budget is meaningless on a handful of dies (--min-dies, default 400).
2. Keep only wafers with a labelled failure pattern, so results can be broken down by pattern type. The evaluation set is then sampled round-robin across the surviving pattern classes, so no one class dominates.
3. Build the in-wafer mask from value != 0 and treat it as known geometry — the agent is allowed to know which grid positions hold a die.
4. Map {1 → 0.0, 2 → 1.0} to get the reality field.
5. Do not resize or interpolate wafers. Spatial structure is the signal being studied.

The core experimental idea

We treat the complete WM-811K wafer map as hidden reality and expose only sparse observations to the agent. A deliberately biased prior emulates simulation mismatch, allowing the Sim2Real correction process to be evaluated against known ground truth.

This is what makes the experiment scoreable. Real metrology never has the full wafer to compare against; here it does, and it is withheld from the agent so the comparison stays fair.

A WM-811K wafer map becomes the hidden reality field. From it, a biased prior generator produces the prior the agent sees everywhere, and an observation mask produces the few values it may read. Both feed the agent, which outputs prediction and uncertainty. Reality and prediction meet only inside the evaluation harness.

Building the evaluation index

python -m nano.subset --raw data/raw/LSWMD.pkl --wafers 12 --seed 0

This writes data/subsets/wm811k_eval.json: the wafer indices, their pattern labels, grid shapes and die counts, plus the filters and seed that selected them. The index is versioned; the dataset it was derived from is not. python -m nano.benchmark reconstructs the same evaluation set from the index plus a local LSWMD.pkl.

Building the biased prior

The prior must be wrong in a structured way, because that is how simulation is wrong. Random noise added to ground truth would be a much easier problem and would not test anything interesting. The prior is generated from the reality field by composing:

• Spatial smoothing — a mask-aware Gaussian blur (smoothing_sigma), so fine structure present in reality is absent from the prior. Positions outside the wafer are excluded from the blur rather than treated as zeros, which would drag the edge down for the wrong reason.
• Radial bias — failure is scaled down by 1 − radial_strength × radius^radial_power, systematically under-predicting towards the wafer edge: the classic signature of a model calibrated on centre measurements.
• Global offset and gain — a calibration error applied to the whole wafer.

The bias parameters live in nano/prior.py as BiasParams, are deterministic, are recorded in the results file under experiment.prior_bias, and are identical across all strategy arms.

Drawing the initial observation mask

The initial mask is drawn centre-biased, not uniformly, to reproduce the sampling bias described on the problem page. All strategies — Random, Grid and NANO — start from the same initial mask for a given (wafer, seed) pair. Only the additional measurements differ.

The continuous target

WM-811K labels are binary, and real metrology readings are not. So the stand-in source can also generate a continuous target — smooth surfaces (radial, tilt, saddle, ring, spot, stripe) plus per-die noise, the shape a CD, thickness or overlay map has:

python -m nano.benchmark --synthetic --target continuous --seeds 0 1 2 3 4

Nothing else in the pipeline changes: the same prior generator, the same estimator, the same selection rules, the same budget. What changes is what the metrics can say — the class-balanced metrics have no classes to balance, so a continuous run leads with plain MAE and reports RMSE beside it, and the constant reference predicts the prior's mean level instead of "no die fails".

This is the only source here that can ask whether the same selection rule works on the kind of data metrology actually produces. It answers with a stand-in, not with a fab, and Honest Scope records what that run found — including one result that contradicts the binary run.

Evaluation protocol

Setting	Value
Evaluation wafers	Sampled by seed from the filtered subset, stratified across failure patterns · recorded at run time
Seeds	recorded at run time
Initial measurements	recorded at run time
Additional measurement budget	recorded at run time
Metric	Reconstruction error over in-wafer dies, direction recorded in the results file
Ground-truth access	Evaluation harness only Hidden from agent

There is no train/test split in the usual sense. Nothing is fitted across wafers: each wafer is an independent active-measurement episode, and the reported metric is the distribution over wafers × seeds. That is a deliberate choice — it isolates the selection rule from any benefit a cross-wafer model might bring.

Honesty about the proxy

This setup is a proxy for Sim2Real, not Sim2Real itself:

• The prior is a synthetic distortion of ground truth, not the output of a TCAD, CAE or RCWA simulator. It has no physics in it.
• WM-811K labels are binary pass/fail, not continuous metrology readings such as CD, thickness or overlay. Real metrology values are continuous and noisy, and the estimator would face a different problem.
• Every die is assumed equally expensive to measure, which is not true of real tool scheduling.

What the proxy does support is the question this project is actually asking: given a fixed budget and a biased prior, does choosing measurements by acquisition score beat choosing them by convention? Full detail in Honest Scope.

Next: the results →