PCINN Predicts SALD Coverage, Inverts Kinetics in Real-Time.

Spatial Atomic Layer Deposition (SALD) is a high-throughput industrial process, but its design and control are hampered by the slow computational cost of predicting surface coverage using traditional methods like Computational Fluid Dynamics (CFD). Researchers have developed a Physics-Chemistry-Informed Neural Network (PCINN) to overcome this limitation. This hybrid surrogate model delivers CFD-level accuracy at real-time speeds, predicting coverage in milliseconds—approximately 50,000 times faster than a CFD simulation. The PCINN architecture is not a black box; a small neural network learns the relationship between operating conditions and near-wall concentration, while known surface kinetics are hard-coded into a trainable chemistry layer. This design ensures accuracy with sparse training data, interpretability, and invertibility. The study also includes a thorough identifiability analysis, confirming robust identification of adsorption energy and desorption rates, and providing diagnostics for unmodeled site heterogeneity.