Prediction of CFD Simulation Results Using Machine-Learning Models and Process Designs Based on Direct Inverse Analysis of the Models

This study proposes machine-learning methods for predicting and inverse analysis of mathematical models in computational fluid dynamics (CFD) simulations to implement process optimization. In conventional pseudo-inverse analysis based on forward analysis, inefficient designing is required with a machine-learning model between process conditions x and the resulting physical properties y, and it is difficult to comprehensively analyze all conditions in the multidimensional space of x. Therefore, a direct inverse analysis method for models that directly predict x values from the target y values was developed, and the design of the experiments was applied to optimize CFD simulation results. The effectiveness of the proposed inverse analysis method was verified through multiple case studies, and the possibility of process improvements based on the visualization and analysis of the simulation results was demonstrated.