Can AI explain a transistor it has only simulated?

Five thousand simulated cases cover six design inputs

The dataset comprised 5,000 simulated devices. Inputs spanned gate lengths from 5 to 30 nanometres, oxide thicknesses from 0.50 to 3 nanometres, silicon-pillar radii from 10 to 20 nanometres, doping concentrations from 10^15 to 10^19 per cubic centimetre, metal work functions from 4.2 to 5.2 electron volts and drain biases from 0.05 to 1 volt. The six predicted outputs were threshold voltage, on-state current, off-state current, subthreshold swing, transconductance and drain-induced barrier lowering.

The researchers randomly assigned 80% of the simulations to training and 20% to testing, with five-fold cross-validation used during model development. The comparison for threshold voltage included random-forest regression, an artificial neural network and XGBoost. XGBoost used 300 trees, a learning rate of 0.05, maximum depth six and subsampling of 0.8. A random split is a reasonable first check of interpolation across the sampled parameter combinations, but it does not show what happens beyond those ranges or under a deliberately difficult holdout such as an unseen geometry regime.

The paper also compares a simulated transfer curve with one previously reported experimental curve as a calibration reference. That is more informative than leaving the simulator unanchored, yet it is not an external test of the learned model across a population of independently fabricated devices. One reference curve cannot reproduce wafer-to-wafer variability, process defects, measurement noise, temperature effects or the many material interfaces encountered in manufacturing.[1]

Accuracy is high on the held-out simulations

For threshold voltage, XGBoost recorded a mean absolute error of 0.004176 volts, root-mean-square error of 0.005281 volts, R² of 0.995003 and mean absolute percentage error of 1.638%. The random forest reached an R² of 0.988079 and the neural network 0.977915. Across the six predicted characteristics, the paper reports R² values of roughly 0.9795 to 0.9989 and percentage errors below 5%. Those figures show that the boosted trees reproduced the simulator's outputs well for randomly held-out points drawn from the same design-space distribution.

They should not be read as a 99.5% success rate for chip design. R² describes how much variation in one numerical target is accounted for under a particular dataset and split; it is not yield, reliability or the probability that a physical device works. Because neighbouring parameter combinations can be similar, a random split may also be easier than testing a model on a fully withheld corner of the design space. Reporting repeated split variability and an untouched out-of-distribution set would make the generalisation claim clearer.[1]

Explanations identify sensitivities, with a physical caveat

SHAP ranked gate length as the strongest influence on threshold-voltage predictions, with a reported mean absolute contribution of about 0.10, followed by metal work function at 0.065, drain bias at 0.05, pillar radius at 0.041, doping at 0.026 and oxide thickness at 0.015. LIME supplied local explanations for individual cases. The ranking broadly matches the expectation that channel geometry and gate material strongly affect electrostatics, offering a useful plausibility check rather than only a model score.

Yet explanation methods describe how this fitted model responds to its inputs. Correlated variables and simulator assumptions can shape those attributions, and neither SHAP nor LIME proves causality. A designer should compare the explanation with device physics and run targeted simulations or experiments before treating it as an engineering rule. Stability checks across seeds, alternative models and deliberately perturbed inputs would also show whether the rankings survive modest changes in training data.

The largest physical limitation appears at the shortest scale. The TCAD calculations use a drift-diffusion transport framework. The authors acknowledge that at 5 nanometres carrier transport may enter a quasi-ballistic regime, meaning the simulated lower bound may not fully represent the relevant physics. Quantum confinement, contact resistance, self-heating and variability can also become decisive in advanced nodes. A machine-learning model cannot correct omitted physics unless suitable evidence is present in its training targets.[1]

What the result could change for engineering teams

A well-validated surrogate could reduce the number of expensive simulations required during early exploration. Engineers might use it to screen parameter combinations, identify influential variables and reserve higher-fidelity calculations for promising or uncertain regions. Explainability can help reviewers catch implausible sensitivities before a model becomes part of a workflow. That is a productivity opportunity rather than evidence that AI can replace device simulation, process engineers or experimental characterisation.

The immediate public effect is indirect. Faster exploration could eventually influence chip cost, performance and energy efficiency, but this study measures none of those outcomes. It does not estimate design-cycle time saved, fabrication cost, yield improvement or energy consumption. Nor does it address whether a foundry can share process data safely enough to validate the model. Teams should therefore treat it as a research prototype for simulation acceleration, with human review and conventional verification still in control.[1]

What would move the evidence beyond simulation

A stronger assessment would begin with a locked model tested on unseen TCAD regimes, including geometry combinations and transport settings outside the random training distribution. Comparisons should include physics-informed surrogates, uncertainty estimates and active-learning strategies that decide when a full simulation is still necessary. Results should be repeated across seeds and reported separately for interpolation and extrapolation.

The decisive step is independent physical validation: multiple fabricated devices across wafers and process corners, measured at different temperatures and operating conditions, with the model evaluated before those results are revealed. High-fidelity quantum or quasi-ballistic simulations would be important near the 5 nm boundary. The authors report no external funding and declare no competing interests. Until such evidence arrives, the defensible conclusion is narrow: explainable boosted trees reproduced six outputs of one TCAD workflow very accurately, not that AI has validated a manufacturable 5 nm transistor.[1]