Physics-Inspired Machine Learning Framework for Real-Time Optimization of Pulsed Laser Thin Film Growth

Pulsed laser deposition (PLD) produces materials with exceptional properties, but growing demands in process control necessitates data-driven approaches. Machine learning offers real-time adaptability beyond traditional experimental capabilities, enabling exploration of new nonequilibrium growth regimes and accelerated materials discovery. In PLD, film quality critically depends on the flux (Φ) and kinetic energy (K) of arriving species—both sensitive to laser fluence (F) and spot area (A). Accurate models for Φ(F,A) and K(F,A) during film growth are elusive due to experimental variability. We propose Gaussian Process (GP) surrogate modeling with active learning to discover these relationships in real time. Preliminary studies using synthetic data from laser ablation simulations (F=1–10 J/cm², A=0.8–13 mm²) demonstrate the approach's feasibility: 100 model runs initialized with random 5-point seeds achieve mean R²=0.999 for K and R²=0.989 for Φ after 15 iterations. We deploy this framework in PLD experiments using a KrF excimer laser ablation of copper targets under vacuum, with Φ and K extracted from ion probe measurements. The key innovation is incorporating physics-informed structured means into the GP, embedding the physics of the PLD plasma expansion into the model rather than relying on zero-mean priors. The framework enables autonomous plasma synthesis control and extraction of interpretable parameters relating the plasma dynamics to thin film characteristics.