(340ab) Data-Driven Optimization of Dynamic Hybrid Models

A dramatic increase in the quantity of data has spurred the development of machine learning (ML) tools that can harness the increased volume and velocity of information. Although these tools have proven robust for some applications, purely data-driven approaches are often limited in their interpretability and predictive power. In contrast, mechanistic modeling approaches offer greater interpretability and generalizability. Yet purely mechanistic approaches are viable only with the availability of correct domain knowledge. To overcome the limitations of both options, hybrid modeling approaches have been proposed to accelerate modeling and optimization while constraining the predictions to verified physical laws.[1, 2] This submission shares recent work at leveraging data-driven tools to achieve mechanistic insight in a hybrid fashion. Specifically, this work explores the following:

• The use of Neural Differential Equations to estimate the parameters of dynamic mechanistic models via time-series data.
• The use of surrogate models to accelerate the discrete element modeling (DEM) for the mechanocatalysis of fine powder systems.
• The use of dynamic optimization to discriminate between dynamic hybrid models.

In all these studies, it is shown that machine learning and mechanistic knowledge can be merged to yield new insights, reduce the computational cost and/or more efficiently use available data for optimizing the process. Finally, progress toward developing open-source tools and use cases that implement the above methods will be shared with the aim of democratizing the developed methods for use by practitioners.

Research Interests: My primary interests lie in the design of software tools for (bio)pharmaceutical development and production. Thus far, my research has focused on investigating methods for merging data-driven tools and mechanistic knowledge for modeling dynamic process data. In future, I wish to create a democratized infrastructure for deploying ML + mechanistic models for process automation, design, optimization and control. Such interests include developing new tools for dynamic modeling as well as novel applications for those tools.