An Automatic Chemical Process Generation Framework for GPU Tensor Parallelism

Abstract This paper presents FlowGen, which is an automatic chemical process generation framework designed for GPU tensor parallelism. Taking feed and product specifications as input, the framework automatically assembles candidate flowsheet topologies via beam search. A surrogate simulator purpose-built for GPU parallelism (incorporating a Spectral Decomposition distillation model, SPD, and a Matrix Exponential heat-exchanger model, ME-NTU, among others) enables full-batch evaluation of 50,000 candidate flowsheets within 1 min. Multi-objective Pareto optimization then recommends the optimal process configurations. Validation across six industrial case studies demonstrates a prediction MAPE below 5%, GPU speedups of 1540–2190× relative to Aspen Plus on a single CPU, total computation time of only 7 min and 43 seconds, and energy savings of 2.2–17.5%.