Harnessing machine learning and optimization for informed chemical engineering decisions: A styrene reactor analysis

Abstract

Machine learning (ML) has become a powerful asset in chemical engineering, offering accurate, data‐driven approaches for modelling and process optimization. While ML enhances the prediction of complex system behaviours beyond traditional empirical models, its integration with optimization frameworks is essential to identify operating conditions that align with multiple industrial goals. This study presents an integrated approach that combines various supervised ML models with multi‐objective optimization (MOO) and multi‐criteria decision‐making (MCDM) techniques, applied to a simulated styrene production reactor. The goal is to train various ML models and use them to simultaneously optimize ethylbenzene conversion and styrene yield, using datasets of varying size and quality. Model performance is evaluated under three data conditions: a complete dataset, a reduced clean dataset, and a reduced noisy dataset. While individual ML models generate different Pareto fronts and optimal solutions, this work demonstrates that a novel consensus‐based strategy that leverages recurring patterns across the full ensemble of ML models can better predict outcomes and produce more robust and reliable recommendations for process conditions than relying on any single model. By aggregating agreement predictions across diverse learners, the consensus approach reduces model‐specific bias and sensitivity to data imperfections, yielding recommendations that are more stable across data quality scenarios and more defensible for decision‐making. By highlighting both the predictive power and optimization potential of ML models across data environments, this research provides a practical framework for integrating ML with MOO/MCDM in chemical process design. The findings emphasize the importance of robustness and interpretability in model‐driven decision‐making.