SPLIT-PINN: Separable Probability Learning Technique via Physics-Informed Neural Networks for High-Dimensional Probabilistic Modeling
Researchers have developed SPLIT-PINN, a novel technique using physics-informed neural networks to model material behavior in high-dimensional probabilistic settings. This method represents material states as probability density functions and infers a probabilistic transport model directly from data. SPLIT-PINN incorporates a marginal-correction drift decomposition and orthogonality constraints to ensure accuracy, stability, and physical consistency without restrictive parametric assumptions. The framework has been validated and applied to predict the evolution of microstructural states in polycrystalline materials, demonstrating robust generalization across unseen datasets. AI
IMPACT Introduces a new method for probabilistic modeling in materials science, potentially improving simulations of material behavior.