(509cy) Chemically-Informed Data-Driven Optimization (ChIDDO): Leveraging Physical Models and Bayesian Learning to Accelerate Chemical Research

Edisonian search approaches are widely used in the chemical sciences to discover reactions, process conditions, material compositions, or product formulations with optimal performance for their intended application. These experimental design methods rely on the generation of grids of variables where experimentally accessible conditions are systematically varied. While these methods are simple to implement, they can result in the oversampling of suboptimal conditions and in information gaps between the selected design parameters. These shortcomings represent significant impediments for expensive experimental campaigns (e.g., during process scale-up, in fine chemicals or pharmaceuticals) or those with large design spaces that can only afford the implementation of coarse experimental grids, underscoring the need for more efficient experimental optimization methods.

To accelerate experimental optimization, Bayesian optimization (BO) has been widely implemented in many different fields of research. Data-driven optimization methods such as BO learn and evolve with new experimental data, but they lack a priori knowledge of the physical laws that dictate the behavior of the chemical system under study. This can result in the need for large experimental campaigns to accurately optimize the design space. On the other hand, optimization methods based on physical models (e.g., density functional theory, molecular dynamics, continuum models, etc.) could be used to identify optima without the need to perform experimental searches, but they often lack the accuracy to effectively capture the complexity of real systems. Given the advantages and shortcomings of both approaches, there is an opportunity to leverage a priori chemical knowledge in data-driven optimization to reduce the data needs and allow for a more efficient identification of the optimum.

In this presentation, we introduce a chemically-informed data-driven optimization (ChIDDO) approach, which is a type of multi-information source optimization (MISO), where inexpensive and low-fidelity information obtained from physical models of chemical processes are combined with expensive and high-fidelity experimental data to optimize a common objective function. While MISO algorithms have been previously implemented to improve BO in computational problems, the implementation of ChIDDO can extend these advantages to chemical experimentation. We will show how the ChIDDO algorithm compares to traditional BO approaches using benchmark objective functions. We will then show how ChIDDO can be used to optimize an electrochemical reaction engineering problem more efficiently than BO.