Accelerating operation optimization of complex chemical processes: A novel framework integrating artificial neural network and mixed-integer linear programming

We present an automated framework that integrates rectified linear unit activated artificial neural network (ReLU-ANN) and mixed-integer linear programming (MILP) to enable efficient operational-level optimization of complex chemical processes. Initially, data is generated through rigorous simulations to pre-train surrogate models based on ReLU-ANN (classification and regression), and subsequently, MILP is employed for optimization by linearly formulating these models. This novel framework efficiently handles complex convergence constraints through a classification neural network which will be used for high-throughput screening data for regression, while simultaneously implementing an 'optimizing while learning' strategy. By iteratively updating the neural network based on optimization feedback, our approach streamlines the optimization process and ensure the feasibility of optimum solution. To demonstrate the versatility and robustness of our proposed framework, we examine three representative chemical processes: extractive distillation, organic Rankine cycle, and methanol synthesis. Our results reveal the framework's potential in enhancing optimization effect while concurrently reducing computational time, surpassing the capabilities of typical optimization algorithms. As for the three processes, optimization effectiveness improved by 10.11%, 28.69%, and 5.45%, respectively, while execution time were reduced by 71.71%, 54.49%, and 59.38%. This notable enhancement in optimization efficiency stems from a substantial reduction in costly while ineffective objective function evaluations. By seamless integration of ReLU-ANN and MILP, our proposed framework holds promise for improving the optimization of complex chemical processes, yielding superior results within significantly reduced timeframes compared to traditional approaches.