Differentiable Thermodynamic Phase-Equilibria for Machine Learning
This episode discusses the paper "Differentiable Thermodynamic Phase-Equilibria for Machine Learning," which introduces DISCOMAX. This method enables neural networks to predict liquid-liquid equilibrium while strictly adhering to thermodynamic laws. By using a discrete grid and a straight-through gradient estimator, it achieves higher accuracy and physical consistency than previous surrogate-based methods.