(364k) Machine Learning-Enhanced Tools for Bioprocess Modelling

During my academic years, I have focused on developing and applying computational Process Systems Engineering (PSE) and machine learning methods to solve a range of chemical engineering problems.

My research interests involve creating computational tools for modelling, control, and optimization in chemical and bioengineering. Early in my studies, I focused on creating mechanistic models for the two-stage paracetamol crystallization process using process simulation software gPROMs. The successful application of these models motivated me to pursue advancements in model construction frameworks for process modelling. During my PhD, I developed systematic hybrid modelling frameworks that integrate high-fidelity mechanistic models with advanced machine learning techniques to uncover the underlying dynamics of biosystems. This approach allowed us to statistically identify the optimal hybrid model structure, effectively balancing mechanistic knowledge with machine learning techniques. As a result, our models excel in extrapolation due to mechanistic insights and in simulating complex or unknown process dynamics, influenced by unknown reaction kinetics and varying operating conditions. To further enhance model performance, we applied hybrid model-based Design of Experiment (DoE) for data augmentation, collecting informative process data for hybrid model training. Additionally, I collaborated with BASF to test the proposed framework to industrial bioprocess simulations. After training the hybrid model, I benchmarked both Frequentist and Bayesian inference-based uncertainty analyses. By quantifying parameter uncertainty, we can identify the confidence intervals of hybrid model predictions. Currently, I am working on both hybrid model-based and model-free control of bioprocesses. For model-free control, I am developing a novel offline reinforcement learning (RL) framework to train an RL agent that optimizes bioprocess control strategies based on historical process data. This work utilizes the decision transformer algorithm to train the agent, whose performance is benchmarked against hybrid model-based model predictive control.

Over the years, I have developed a robust international network within both academia and industry, enabling me to establish significant cross-departmental and cross-organizational collaborations that will support and advance my career path. I am looking for an industry position where I can apply my technical skills and collaborate with a team to solve real-world problems. Particularly, I’m enthusiastic about the data science and machine learning positions.