Machine learning-enhanced reduced order modeling for pharmaceutical particulate processes
Loading...
Files
Date
2025
Authors
Jiang, Yu
Journal Title
Journal ISSN
Volume Title
Publisher
University College Cork
Published Version
Abstract
A thorough understanding of particulate processes is paramount for the design, optimization, and control of operations across diverse industries, including pharmaceuticals. However, high-fidelity, physics-based simulation methods such as Computational Fluid Dynamics (CFD) and the Discrete Element Method (DEM) are burdened by prohibitive computational costs. This expense severely limits their application for industrial-scale systems, particularly in tasks requiring extensive querying, such as design-space exploration, real-time control, and optimization. To address this computational bottleneck, this thesis aims to develop, implement, and rigorously evaluate a series of data-driven reduced-order models (ROMs) enhanced by machine learning (ML) to significantly accelerate the simulation of complex particulate and multiphase flows while maintaining acceptable predictive accuracy.
The research establishes a methodological pathway that evolves from classical linear methods to advanced, non-linear deep learning frameworks. The work commences in Chapter 3 with a classical ROM based on Singular Value Decomposition (SVD), which successfully modeled steady-state solid-liquid mixing in a stirred tank and achieved a three-order-of-magnitude speedup. However, it also revealed the limitations of linear methods in capturing highly non-linear turbulent effects. To address non-linearity and transient dynamics, Chapter 4 introduces a Graph Neural Network (GNN)-based surrogate to model transient granular behavior. A key innovation was coupling this fast surrogate with a gradient-based optimizer to perform efficient inverse design for DEM parameter calibration.
To tackle the core challenge of high-dimensional, transient 3D multiphase flows, a novel hybrid ROM coupling Dynamic Principal Component Analysis (DPCA) with a transformer network was introduced in Chapter 5. The DPCA-Transformer model demonstrated superior reconstruction and prediction accuracy for transient solid-liquid mixing compared to benchmarks, while also achieving a three-order-of-magnitude speedup. Finally, to demonstrate industrial applicability and scalability, Chapter 6 applies this advanced framework to a complex, three-phase (gas-liquid-solid) centrifuge dewatering process, achieving a remarkable runtime acceleration of over 5000x compared to the full CFD simulation.
In summary, this thesis not only provides efficient simulation tools for specific pharmaceutical processes but, more importantly, establishes and validates a clear progression of ML-enhanced ROM development strategies. This work contributes an enabling technology stack that accelerates the transition toward data-centric, model-driven pharmaceutical manufacturing, facilitating rapid design exploration, real-time monitoring, and closed-loop optimization control.
Description
Keywords
Particulate process , Machine learning , Reduced order model , Data-driven
Citation
Jiang, Y. 2025. Machine learning-enhanced reduced order modeling for pharmaceutical particulate processes. PhD Thesis, University College Cork.