PCINN neural network predicts atomic layer deposition 50,000x faster

Spatial atomic layer deposition (SALD) is a high-throughput, atmospheric-pressure method for growing precise thin films in semiconductor manufacturing. But optimizing its operating window has been stuck between two extremes: high-fidelity CFD simulations that are too slow for scanning many conditions, and analytic models that miss key transport effects like gas curtains. To solve this, Ning Hu and colleagues developed PCINN (Physics-Chemistry-Informed Neural Network), a hybrid surrogate that combines a small neural network with a hard-coded, trainable chemistry layer integrated along the substrate trajectory.

The results are striking: PCINN answers a coverage query in about 7 milliseconds—roughly 5×10^4 times faster than a CFD solve—while matching CFD accuracy (test R²_log = 0.998, leave-one-out R²_raw = 0.974) using only 30 training cases spanning four orders of magnitude in coverage. The architecture's single-scalar bottleneck prevents overfitting, keeps the model interpretable, and allows reliable inversion of kinetic parameters. The authors also performed a full identifiability analysis using Fisher information and profile likelihood, showing that adsorption energy and desorption rate are robustly identifiable, while the adsorption rate constant is not separately identifiable at a single temperature. They even derived an analytical degeneracy valley across temperatures and turned it into a diagnostic for unmodeled site heterogeneity, making the model not just fast, but trustworthy for process control and materials discovery.