Why Does Half the Gain Variation in Phototransistors Come From Between Runs?

A silicon bipolar phototransistor spends months in a cleanroom before anyone can measure its gain. The wait makes any model that predicts gain from process parameters valuable even when imperfect. The new study treats this as a small-sample, hierarchically structured problem rather than a large-corpus one.

Thirteen to fourteen runs. One device. That is the entire dataset.

Conventional virtual metrology assumes thousands of wafers. Here the hierarchy — batch, wafer, die — carries more signal than any flat table could. The nested structure is not a nuisance. It is the only lever available.

Where The Variation Actually Lives

Decomposing the variance of device gain revealed a split. Roughly half of it lies between process runs rather than within them. Recipe-only prediction is bounded by construction because the run itself carries information no recipe parameter captures.

The team built a multi-level data-quality assessment tailored to those nested physical entities. Batch, wafer, die — each level gets an explicit cross-level linkage score. The score quantifies how much information at one level explains variation at another.

A Predictor That Knows Its Own Uncertainty

The forward gain predictor outputs a relative, uncertainty-aware signal. It does not merely guess a number. It attaches a calibrated confidence interval that reflects the hierarchical noise structure.

Uncertainty here is not a post-hoc bootstrap. It propagates from the cross-level linkage scores through the predictor. A run with weak wafer-to-die linkage produces a wider interval. A run with strong linkage produces a tighter one.

Searching Backward From A Target Gain

The inverse search returns recipes for a target gain. It does not invert a deterministic function. It samples from the posterior predictive distribution conditioned on the desired gain and the observed hierarchy.

Because the forward model is uncertainty-aware, the inverse search naturally avoids regions where the predictor is unreliable. It prefers recipes that land in well-calibrated zones of the process space. The method does not guarantee a feasible recipe exists for every target.

What The Released Code Contains

Forty-three pages and fifteen figures document the approach. The normalized dataset and analysis code are released for full reproducibility. The release includes the hierarchical data-quality assessment, the forward predictor with its uncertainty calibration, and the inverse search routine.

No benchmark against large-corpus virtual metrology appears in the paper. The authors do not claim superiority over methods designed for thousands of runs. They demonstrate that for thirteen to fourteen runs of one device, hierarchy-aware modeling beats ignoring the hierarchy.

The cleanroom still takes months. The model now tells you what to expect before you commit the time.