Research
My research develops models that combine engineering knowledge with data, and the software that makes them usable. Each theme links to its publications.
Hybrid models with physical and qualitative constraints
Hybrid models combine first-principles relations with learned components. My work places engineering knowledge into the training of such models: bounds on the output, the direction in which it changes with an input, and simplified balances. The aim is a model whose behavior outside its data stays close to its behavior within them. Examples include surrogate models and transfer learning for the optimization of high-temperature polymer electrolyte membrane (HT-PEM) fuel cells, hybrid models for ionic separations, and the incorporation of qualitative knowledge in hybrid modeling.
Testing data-driven methods for chemical plants
Learned models and controllers need evidence before they operate a plant. I develop frameworks for testing data-driven and statistical methods on simulated chemical plants with a verified ground truth. The tests run over long horizons and grids of initial states. They report safety and convergence separately. This line builds on earlier work on process monitoring, data mining, and dimensionality reduction for industrial data.
Scientific machine learning software
In Data and Intelligent Systems at Princeton University, I build machine learning methods and research software with research groups across disciplines. Much of this work concerns training and evaluation workflows that run at scale on GPU clusters, with reproducible environments and batch pipelines. It also concerns scientific software that is reusable, tested, and documented, so that a group can keep building on it. HElioS, an open-source Python framework for modeling and optimizing electrochemical systems, is an earlier example.
Electrochemical systems and separations
My doctoral work at Louisiana State University applied machine learning to electrochemical devices and separations. It includes models that guide HT-PEM fuel cells toward greater power density and transfer learning across membrane electrochemical systems. Related work combined molecular dynamics with machine learning to estimate ion activity coefficients in ion-exchange membranes.