Physics-informed neural networks for optimization of polymer reactor design

Computational fluid dynamics (CFD) is utilized for chemical reactor design/analysis and enables accurate predictions by solving numerical methods. However, real-time prediction is currently infeasible due to its high computational cost. As a result, physics-informed neural networks (PINNs), which integrate data-based machine learning and physical principles, have recently emerged as a surrogate model of CFD (Raissi, M., et al. (2019)). But previous PINNs can only predict properties in terms of location or time. Here, we offer novel PINNs that can also handle operating conditions as input variables to evaluate the impact of various design variables for reactor design optimization (Raissi, M., et al. (2020)). The proposed model is applicated to an autoclave reactor for free radical polymerization of ethylene and demonstrated to be capable of interpolating/extrapolating solutions accurately under various settings.