A Physics-Chemistry-Informed Neural Network (PCINN) for Real-Time Spatial-ALD Coverage Prediction and Reliable Kinetics Inversion

Spatial atomic layer deposition (SALD) enables thin‑film growth at atmospheric pressure with high throughput, but engineers still struggle to predict how precursor gases will coat a moving substrate. Traditional CFD simulations give detailed surface‑coverage maps but require minutes to hours per case, making rapid design sweeps impractical. Analytic shortcuts run fast but ignore subtle flow features such as the gas curtain that can dramatically alter local chemistry. A new surrogate model bridges this gap by embedding physics and chemistry directly into a neural network, delivering CFD‑level accuracy in milliseconds.

What You Need to Know

The authors introduce a Physics‑Chemistry‑Informed Neural Network (PCINN) that acts as a surrogate for full CFD‑based SALD simulations. Rather than learning a pure black‑box mapping from operating conditions to coverage, the network is constrained to respect known mass‑transfer and reaction‑rate relationships. A small feed‑forward architecture learns only the residual deviation between a physics‑based baseline prediction and the high‑fidelity CFD result, which keeps the model compact and interpretable.

Training the PCINN required only 30 carefully selected cases that span four orders of magnitude in surface coverage, from sub‑monolayer to near‑saturation regimes. Each case includes variations in precursor flow rates, carrier‑gas velocity, temperature, and substrate speed. After training, a single query—providing the instantaneous inlet conditions and local geometry—returns a coverage field in about 7 ms. Compared with a typical CFD solve that takes roughly 0.35 s, the surrogate is ~5 × 10⁴ times faster while maintaining excellent fidelity: test‑set R²_log = 0.998 and leave‑one‑out R²_raw = 0.974.

Because the network learns only the correction term, users can inspect the physics‑based baseline to understand which mechanisms dominate under different conditions. This hybrid approach avoids the opacity of deep‑learning models while still capturing complex transport effects such as the gas curtain that analytical correlations miss.

Why It Matters

Real‑time coverage prediction enables closed‑loop control of SALD tools, allowing operators to adjust precursor pulses or substrate speed on the fly to maintain uniform film thickness across large areas. Faster evaluation also accelerates process‑window exploration, reducing the number of costly experimental runs needed to locate optimal parameters. For manufacturers scaling SALD to roll‑to‑roll or large‑area panel production, the ability to simulate thousands of candidate settings in seconds translates directly into higher yield and lower development cost.

Beyond immediate process control, the PCINN framework provides a template for other gas‑phase deposition techniques where CFD is prohibitive. By embedding known conservation laws and reaction kinetics into the surrogate, developers can trust that extrapolations remain physically plausible, a critical advantage when moving from lab‑scale coupons to industrial‑scale reactors.

Key Details

• Query latency: ~7 ms per coverage field.
• Speed‑up vs. CFD: ≈5 × 10⁴× faster.
• Accuracy: test R²_log = 0.998; leave‑one‑out R²_raw = 0.974.
• Training set: 30 simulations covering 4 orders of magnitude in coverage.
• Inputs: precursor flow, carrier‑gas velocity, temperature, substrate speed, local geometry.
• Architecture: small feed‑forward net learning the residual of a physics‑based baseline.
• Interpretability: baseline term retains explicit mass‑transfer and reaction equations.

What’s Next

Future work will focus on integrating the PCINN into real‑time tool controllers, validating its predictions against in‑situ spectroscopic measurements across varied precursor chemistries, and extending the surrogate to multi‑layer stacks where surface reactions evolve with each cycle. Expanding the training envelope to include plasma‑enhanced ALD regimes and testing the model’s robustness to noisy sensor data will be essential steps toward deploying physics‑informed AI as a standard routine in high‑throughput deposition manufacturing.

📌 Source: Arxiv Ml