Abstract
The chemical and process industries face increasing pressure to deliver more efficient, sustainable, and resilient operations, driven by decarbonisation targets and the need for adaptive systems that operate reliably amid uncertainty and equipment degradation. Data-driven approaches offer powerful alternatives to computationally intractable first-principles models by capturing complex nonlinear relationships directly from process data. When combined with mechanistic knowledge in hybrid architectures, these models additionally gain physical consistency and improved generalisation beyond the training domain. This thesis presents a systematic investigation of data-driven, hybrid, and digital-twin modelling frameworks for the optimisation, control, and monitoring of complex chemical engineering systems, with primary applications in CO2 capture via pressure swing adsorption and oil production via electric submersible pump and gas lift systems. Deep neural network surrogate models are developed for PSA-based CO2 capture, enabling feasible operating maps and computationally tractable optimisation, complemented by a novel theorem for certifying the optimality of surrogate-based solutions. Multi-objective optimisation frameworks combining neural network and symbolic regression surrogates are demonstrated for both PSA and gas lift systems, with the symbolic regression framework producing a Pareto front approximately 8.5 times denser and 34 times faster than the DNN-based approach. Hybrid scientific machine learning approaches correct mechanistic model mismatches in ESP systems, and three complementary control formulations are developed: a nominal NMPC and two robust schemes based on symbolic regression surrogates, achieving per-iteration solve times approximately one order of magnitude lower than mechanistic implementations whilst maintaining offset-free tracking and constraint satisfaction. Digital twin frameworks integrating Bayesian uncertainty quantification, online learning, and cognitive drift detection consistently recovered over 98\% of prediction accuracy under equipment degradation and process drift, and a generalised digital twin architecture is formalised as a transferable framework for chemical engineering. As a cross-cutting contribution, a hybrid graph neural network for thermodynamically consistent viscosity prediction embeds established correlations as inductive biases, reducing extrapolation errors by orders of magnitude relative to unconstrained baselines. Collectively, these contributions address interconnected gaps in the literature concerning the reliable, physically consistent, and computationally efficient deployment of data-driven and hybrid models across the full decision-making hierarchy of chemical engineering systems.