Estimation of particle growth kinetics via physics-informed neural networks

Artificial intelligence is increasingly important in materials science for accelerated discovery and analysis. This work tackles the challenge of accurately extracting kinetic parameters, crucial for modeling chemical transformation rates like nucleation and growth. While classical models such as the Avrami equation provide a useful phenomenological description, conventional AI-based solutions often lack physical constraints and require extensive datasets, which are not always available. On the other hand, traditional curve-fitting methods can be sensitive to noise and data limitations. To address these issues, we introduce a Physics-Informed Neural Network (PINN) designed to infer Avrami parameters with improved accuracy and robustness to noisy and sparse data while ensuring physical plausibility. A novel automated adaptive loss scaling strategy is proposed to address common PINN training challenges, dynamically adjusting loss weights without manual tuning, thus enhancing convergence stability. Our PINN framework outperforms standard fitting techniques by providing reliable parameter estimates from limited, noisy data, effectively integrating data-driven adaptability with the constraints of physical modeling for robust kinetic characterization.