Machine Learning and Multicriteria Optimization for the Optimal Design of Chemical Processes

Abstract

Chemical engineering systems are increasingly characterized by nonlinear behaviour, strong interactions among variables, and growing volumes of experimental, simulated, and literature-derived data. At the same time, many engineering problems involve competing technical objectives and require not only accurate prediction, but also systematic support for analysis, design, and decision-making. In this context, this thesis investigates data-driven frameworks for a range of chemical engineering applications, with emphasis on building predictive models that can be integrated into broader workflows for process analysis, optimization, and engineering decision support.

The thesis is organized around two main groups of studies. The first group focuses on process-level applications, using the styrene reactor as a representative case study. In this group, multiple supervised machine learning (ML) models are developed to predict ethylbenzene conversion and styrene yield, and the trained models are integrated with multi-objective optimization (MOO) and multi-criteria decision-making (MCDM) to identify favourable reactor operating conditions under competing objectives. This work demonstrates how model choice, data quality, and uncertainty influence the reliability of Pareto-optimal recommendations and introduces a consensus-based strategy to identify more robust operating conditions across different surrogate models and data scenarios. A related study addresses missing-data imputation in a styrene production dataset containing process variables using an autoassociative neural network (AANN). The results show that missing process variables can be reconstructed with acceptable accuracy, even when relatively small or noisy datasets are used, although the accuracy depends strongly on the correlation between the missing variable and the remaining process variables. Together, these studies demonstrate how data-driven methods can support both reactor optimization and data reconstruction, strengthening process understanding while improving the data foundation required for reliable downstream modelling and decision support.

The second group focuses on catalytic and reaction-oriented systems. One set of studies examines the catalytic conversion of sulfur dioxide to sulfur trioxide, where predictive models are developed to support catalyst and process design using literature-derived experimental data. In this work, different preprocessing strategies and feature representations are evaluated to determine how data preparation choices influence model accuracy, interpretability, and the reliability of engineering conclusions. The resulting models are integrated with optimization to jointly consider conversion, productivity, and cost, and the study demonstrates that combining information obtained from different preprocessing strategies after optimization can reduce dependence on a single data representation and lead to less biased engineering recommendations. Another study addresses chemoselectivity in hydrogenation reactions involving competing organic substrates. In this work, ML models are developed to predict conversion and yield across different substrate-catalyst pairs and operating conditions, and the trained models are integrated with optimization to recommend catalysts and conditions that favour the desired transformation while limiting undesired conversion. This framework provides directional catalyst-recommendation maps and identifies favourable operating windows for selective hydrogenation. Across these studies, the emphasis is placed not only on predictive accuracy, but also on translating model outputs into practically useful guidance for catalyst selection, operating-condition screening, and reaction planning.

Overall, this thesis demonstrates that data-driven approaches can contribute far beyond prediction alone. Across process modelling, data reconstruction, catalytic process design, and reaction planning, the work shows how predictive frameworks can be used to support more informed, efficient, and transparent engineering decisions. The thesis, therefore, presents a coherent perspective on the role of data-driven methods in chemical engineering, showing their value as tools for transforming diverse data sources into actionable knowledge for analysis, design, and decision support.