AI-Driven Process Optimization for Green Petrochemical Synthesis

Thesis Abstract

In the face of rising energy demand and mounting environmental concerns, the petrochemical sector confronts inefficiencies and greenhouse gas emissions arising from conventional synthesis pathways, prompting a shift toward greener, more adaptable processes. This study addresses the imperative to minimize resource use, enhance process intensification, and reduce emissions in petrochemical synthesis through an AI-driven optimization framework that integrates real-time process data, predictive modeling, and control strategies. The aim is to develop and validate a multi-objective optimization platform that simultaneously improves yield, energy efficiency, and lifecycle environmental Impact for selected green synthesis routes, with an emphasis on integrating novel catalysts, solvent-free or low-solvent operations, and process intensification techniques. Specific objectives include (1) to construct a digital twin of a representative green petrochemical process chain (e.g., ethylene oxide to ethylene glycol or renewable feedstock upgrading) comprising reactors, separators, heat exchangers, and catalysts; (2) to develop machine learning surrogate models (Gaussian process regression, gradient boosting, and neural networks) for rapid prediction of reaction kinetics, heat/mass transfer, and catalyst performance; (3) to formulate and solve a multi-objective optimization problem using Pareto optimization and reinforcement learning-based control to minimize energy consumption and CO2 emissions while maintaining target product purity and yield; (4) to validate the framework against pilot-scale data (n = 6 experiments) and scale-up scenarios, and (5) to perform a sensitivity and uncertainty analysis to assess model robustness under feedstock variability and catalyst aging. Methodologically, the study adopts a mixed-methods design anchored in systems engineering and chemical process theory. The population comprises industrially relevant green petrochemical processes and laboratory-scale pilot plants operating under validated process conditions. A purposive sample of six pilot-scale experiments and three industrial case studies will be analyzed. Data collection will utilize inline process sensors, spectroscopic data (NIR, FTIR), chromatographic analyses (GC-MS, HPLC), and catalyst performance metrics, complemented by operator logs and process diaries. Instrumented data will be integrated with process simulations (Aspen Plus) to generate a digital twin framework, and historical process data (5 years, averaged 12 months per case) will be used to train surrogate models. The data analysis will combine (i) regression analysis and Gaussian process regression for kinetic and transport predictions, (ii) ANOVA and multivariate regression for design-of-experiment insight, (iii) multi-objective optimization via Pareto-front generation using NSGA-II and Bayesian optimization, and (iv) reinforcement learning for adaptive control of reaction parameters and separation stages. Model validation will be performed using holdout testing (20% of data) and concurrency checks against pilot-scale runs. The conceptual framework will be grounded in the Theory of Constraints and the Green Chemistry Principles, with supporting theories including Process Systems Engineering and Dynamic Energy Optimization. Expected findings indicate that the integrated AI-driven platform will produce a Pareto-optimal set of process conditions that reduce overall energy consumption by 15–25%, lower life-cycle CO2 equivalent emissions by 20–35%, and maintain product yields within ±2% of target levels, under varied feed compositions. Surrogate models are anticipated to achieve prediction errors below 5% for key variables across operating envelopes, enabling real-time decision support and control actions with latency under 1 s. The study is expected to identify critical factors such as catalyst deactivation rates, heat integration effectiveness, and solvent recycle efficiency that most influence sustainability gains, along with robust control policies resilient to feedstock variability. Theoretical contributions include an integrated digital twin architecture for green petrochemical synthesis combining data-driven models with first-principles kinetics, advancing knowledge on how AI-driven optimization can harmonize process performance with environmental objectives. Practically, the work will provide a tested methodology and open-access software modules for industrial deployment, including a modular AI core, surrogate modeling templates, and a multi-objective optimization workflow compatible with existing process simulators. Recommendations emphasize (i) phased deployment of digital twin-enabled optimization in refinery clusters, (ii) investment in high-fidelity sensors and catalyst monitoring, and (iii) development of standard data governance for cross-site data sharing to maximize learning and continuous improvement.