Discovery of classical gas-solid flow correlations using a reinforcement learning-based symbolic regression framework

Abstract

Empirical correlations are extensively applied to predict minimum fluidization velocity in gas–solid flow systems; however, their generalizability and physical consistency remain limited under complex conditions. This work integrates a deep reinforcement learning-based symbolic regression framework, Physical Symbolic Optimization (PhySO), to rediscover and improve classical correlations. A systematic benchmark comprising eleven representative expressions across four correlation categories was established, incorporating domain knowledge such as dimensionless numbers and dimensional homogeneity to constrain the search space.The proposed framework yielded compact and physically consistent expressions that retained key features of classical models. For complex formulations, such as the Ergun correlation, expression complexity was reduced from 24 to 14 while maintaining high predictive accuracy (R² = 0.9930). Performance across multiple random seeds confirmed the model’s stability, and tests under noise levels up to 10 % demonstrated strong robustness. These findings demonstrate the potential of physics-informed symbolic regression as an interpretable and reliable alternative to conventional or black-box models for multiphase flow prediction.