Machine-Learning-Enabled Design of Lab-Scale Continuous Distillation Systems for Pharmaceutical Process Development

Abstract

Abstract Predictive tools for continuous distillation design lag behind upstream pharmaceutical synthesis─including AI-enabled digital tooling for retrosynthetic route design─leaving column specification dependent on computationally expensive simulation or extensive experimentation. This work presents a machine-learning surrogate framework trained entirely on first-principles mechanistic simulations, requiring no experimental distillation data, and eliminating the need for commercial tools such as ChemCad or AspenPlus. Three surrogate models for binary tert-butylamine/alcohol systems (methanol, ethanol, isopropanol) predict distillate and bottoms compositions with mean absolute errors below 0.005 mol/mol and R2 > 0.999, completing a single prediction in under 1 ms─a 24-fold speedup. A coupled Pareto optimizer returns capital-versus-operating-cost trade-off curves in under 10 s, with over 93% of designs verified within ±0.010 mol/mol. Applied to a pharmaceutical case study, the framework correctly identifies isopropanol as the optimal solvent without prior thermodynamic knowledge. The system-agnostic architecture could readily integrate with open-source large language model platforms to further lower the barrier to implementation.