Physics informed machine learning based optimization for chemical engineering applications

Neural networks offer a promising alternative to first-principles models in complex industrial systems, but their effectiveness is often limited by noisy data, scarce samples, and the computational demands of training and optimization. This work addresses these challenges by integrating physics-informed structures, transfer learning, and advanced training strategies for process modeling and optimization. Physics-informed neural networks were employed in surrogate optimization problems, where case studies—from a simple blending process to a crude oil distillation unit—showed consistent improvements in solution quality and optimization reliability. The integration of physics-informed training with piecewise linear approximations reduced CPU times, while predominantly achieving global optima. PINN training was also extended to recurrent neural networks for modeling dynamic parameters in a refinery wastewater treatment plant. Physics-informed LSTM and GRU models improved prediction accuracy, with the LSTM reducing MSE by over 20% and the GRU cutting COD prediction errors by 11% compared to standard models, while eliminating the need for fine-tuning during online validation. Transfer learning techniques were applied to address noisy and limited data conditions, where a hybrid physics-informed transfer learning model outperformed baseline models, reducing test and validation MSE by up to 27% and 59%, respectively. Finally, the semi-continuous specially ordered set based training (SOSX) algorithm was adapted for neural network training, achieving near-zero training error across various network configurations while significantly reducing CPU time compared to traditional MIP-based methods. These results highlight SOSX's scalability and applicability to large neural network configurations. Collectively, these approaches demonstrate how combining physics-based knowledge, architectural innovations, and optimization techniques can enhance the predictive performance and computational efficiency of neural networks in industrial and environmental processes.