Machine Learning Based Applications for Data Visualization, Modeling, Control, and Optimization for Chemical and Biological Systems

Abstract

This dissertation report covers Yan Ma’s Ph.D. research with applicational studies of machine learning in manufacturing and biological systems. The research work mainly focuses on reaction modeling, optimization, and control using a deep learning-based approaches, and the work mainly concentrates on deep reinforcement learning (DRL). Yan Ma’s research also involves with data mining with bioinformatics. Large-scale data obtained in RNA-seq is analyzed using non-linear dimensionality reduction with Principal Component Analysis (PCA), t-Distributed Stochastic Neighbor Embedding (t-SNE), and Uniform Manifold Approximation and Projection (UMAP), followed by clustering analysis using k-Means and Hierarchical Density-Based Spatial Clustering with Noise (HDBSCAN). This report focuses on 3 case studies with DRL optimization control including a polymerization reaction control with deep reinforcement learning, a bioreactor optimization, and a fed-batch reaction optimization from a reactor at Dow Inc.. In the first study, a data-driven controller based on DRL is developed for a fed-batch polymerization reaction with multiple continuous manipulative variables with continuous control. The second case study is the modeling and optimization of a bioreactor. In this study, a data-driven reaction model is developed using Artificial Neural Network (ANN) to simulate the growth curve and bio-product accumulation of cyanobacteria Plectonema. Then a DRL control agent that optimizes the daily nutrient input is applied to maximize the yield of valuable bio-product C-phycocyanin. C-phycocyanin yield is increased by 52.1% compared to a control group with the same total nutrient content in experimental validation. The third case study is employing the data-driven control scheme for optimization of a reactor from Dow Inc, where a DRL-based optimization framework is established for the optimization of the Multi-Input, Multi-Output (MIMO) reaction system with reaction surrogate modeling. Yan Ma’s research overall shows promising directions for employing the emerging technologies of data-driven methods and deep learning in the field of manufacturing and biological systems. It is demonstrated that DRL is an efficient algorithm in the study of three different reaction systems with both stochastic and deterministic policies. Also, the use of data-driven models in reaction simulation also shows promising results with the non-linear nature and fast computational speed of the neural network models.