Physics-Guided CNN Predicts Phase Separation with 100% Composition Accuracy
Attention-based model matches Lifshitz-Slyozov law for binary mixtures.
Predicting the spatiotemporal evolution of phase separation in binary mixtures traditionally requires solving the Cahn-Hilliard equation—a nonlinear PDE that is computationally expensive. In a new paper on arXiv, Vijay Yadav and colleagues introduce a physics-guided convolutional neural network (CNN) that acts as a surrogate model, learning the dynamics directly from data. Their architecture incorporates attention mechanisms and physical constraints to ensure conservation of the mixture composition over time. The model achieves stable, accurate predictions for both critical and off-critical mixtures, even during long rollout simulations.
Validation shows the trained CNN correctly captures the domain growth rate, consistent with the theoretical Lifshitz-Slyozov law. This hybrid approach—combining deep learning with physics guidance—offers a path to accelerate simulations of conserved kinetic systems. While the current study focuses on binary phase separation, the framework is designed to generalize to other complex dynamical systems, from biological pattern formation to materials science. The paper is posted on arXiv under ID 2606.26128.
- Attention-based physics-guided CNN predicts Cahn-Hilliard dynamics for binary mixtures.
- Model preserves mixture composition exactly and matches Lifshitz-Slyozov domain growth law.
- Achieves stable long-time predictions for both critical and off-critical mixtures.
Why It Matters
Speeds up materials and biological simulations by replacing costly PDE solvers with accurate AI.