Efficient process optimization for a circular economy by full flowsheet neural networks

• 50,000x speed-up for a waste-to-methanol process simulation by a neural network. • Neural network captures effects of both feedstock composition and process parameters. • Efficient optimization for a plastic and biomass feed for varying economic scenarios. • Explainable AI reveals influence of feed and process parameters on process response. The transition of the chemical industry to a circular economy brings challenges, including fluctuating feedstock compositions and volatile energy prices, which necessitate efficient process optimization. Process flowsheet simulations, invaluable in process design, struggle with high-dimensional optimization due to computational cost and lack of gradient information. We address this by combining neural network surrogate models with gradient-based optimization and explainable AI analysis to enable high-throughput flowsheet optimization. Using a polymer waste-to-methanol process as a case study, neural network inference of the process response is 50,000 times faster than conventional flowsheet simulations. Shapley Additive exPlanations (SHAP) provide insights into the contributions of feedstock compositions and process parameters on process responses and gradient-based optimization optimizes flowsheets within 0.35 s. The optimized process aligns with industry heuristics, achieving syngas ratios near the theoretical optimum. This approach accelerates optimization, provides valuable insights into high-dimensional flowsheets, and advances efficient process design in a circular economy.