Novel feasible set learning and process flexibility analysis method using deep neural networks

Introduction:

Chemical processes are designed under uncertainty: feed quality, reaction rates, temperatures, and other conditions may all shift during operation. Traditional flexibility analysis often asks how large a box or ellipsoid can be drawn around a nominal condition, but this can be misleading when the true safe region is irregular, nonconvex, disconnected, or high-dimensional. This work instead tries to learn the entire feasible set. A physics-informed neural network first acts as a fast surrogate for the process equations, replacing repeated expensive solves of mass-balance, energy-balance, and other equality constraints. A second neural network then maps uncertain operating conditions into a feature space where feasible points are pulled near a hypersphere center and infeasible points are pushed away. With this learned map, Monte Carlo sampling can estimate the volumetric flexibility index: the fraction of the uncertainty space where the process can still be operated safely. The method is demonstrated on heat-exchanger and reactor-cooler systems, including disconnected and five-dimensional feasible regions.