Neural surrogates for crystal growth dynamics with variable supersaturation: explicit vs. implicit conditioning
Researchers have developed Convolutional Recurrent Neural Network surrogate models to simulate crystal growth dynamics. These models are trained on data from Allen-Cahn dynamics and can account for variable supersaturation levels. The study compared two architectures: one that implicitly infers supersaturation from a mini-sequence of frames, and another that takes supersaturation as an explicit input. Results indicate that explicit parameter conditioning yields the most accurate predictions, though the implicit method can achieve comparable results with larger training datasets. AI
IMPACT Introduces novel neural network architectures for simulating complex physical processes, potentially accelerating materials science research.