An Application of Gaussian Processes for Analysis in Chemical Engineering

Industry 4.0 is transforming the chemical engineering industry. With it, machine learning (ML) is exploding, and a large variety of complex algorithms are being developed. One particularly popular ML algorithm is the Gaussian Process (GP), which is a full probabilistic, non-parametric, Bayesian model. As a blackbox function, the GP encapsulates an engineering system in a cheaper framework known as a surrogate model. GP surrogate models can be confidently used to investigate chemical engineering scenarios. The research conducted in this thesis explores the application of GPs to case studies in chemical engineering. In many chemical engineering scenarios, it is critical to understand how input uncertainty impacts an important output. A sensitivity analysis does this by characterising the input-output relationship of a system. ML encapsulates a large system into a cheaper framework, enabling a Global Sensitivity Analysis (GSA) to be conducted. The GSA considers the model behaviour over the entire range of inputs and outputs. The Sobol’ indices are recognised as the benchmark GSA method. To achieve a satisfactory precision level, the variance-based decomposition method requires a significant computational burden. Thus, one exciting application of GPs is to reduce the number of model evaluations required and efficiently calculate the Sobol’ indices for large GSA studies. The first three case studies used GPs to perform GSA’s in chemical engineering. The first examined the effects of thermal runaway (TR) abuse on lithium-ion batteries. To calculate time-dependent Sobol’ indices, this study created an accurate surrogate model by training individual GPs at each time step. This work used GPs to help develop a complex chemical engineering simulation model. The second GSA calibrated a high-shear wet granulation model using experimental data. This work developed a methodology, linking two GSA studies, to substantially reduce the experimental effort required for model-driven design and scale-up of model processes. This enabled the creation of a targeted experimental design that reduced the experimental effort by 42%. The third case study developed a novel reduced order model (ROM) for predicting gaseous uptake of metal-organic framework (MOF) structures using GPs. Based on previous GSA research, the Active Subspaces were located using the Sobol’ indices of each pore property for the MOF structures. The novel ROM was shown to be a viable tool to identify the top-performing MOF structures showing its potential to be a universal MOF exploration model. The final two case studies applied GPs as a tool in novel techniques that combined ML algorithms. First, GPs are seldom used for mid-term electricity price forecasting because of their inaccuracy when extrapolating data. This research aimed to improve GP prediction accuracy by developing a GP-based clustering hybridisation method. The proposed hybridisation method outperformed other GP-based forecasting techniques, as demonstrated by the Diebold-Mariano hypothesis test. In the final case study, ML models were used to develop an effective maintenance strategy. The work compares ML algorithms for predictive maintenance and maintenance time estimation on a cyber-physical process plant to find the best for the maintenance workflow. The best algorithms for this case study were the Quadratic Discriminant Analysis model and the GP. The overall plant maintenance costs were found to be reduced by combining predictive maintenance with maintenance time estimation into a workflow. This research could help improve maintenance tasks in Industry 4.0. This thesis focused on using GPs to enhance collaborative efforts and demonstrate the enormous impact that ML can have in both research and industry. By proposing several novel ideas and applications, it is shown that GPs can be an efficient and effective tool for the analysis of chemical engineering systems.