(661f) Towards Domain-Informed Machine Learned Models from High Throughput Experimental Data

High throughput experimental (HTE) data are increasingly becoming available heterogeneous catalysis, whereby a large space of materials and reaction conditions can be sampled to identify suitable catalyst candidates for a target reaction. We posit that such data intrinsically contain kinetic information and, therefore, developing data-driven kinetic models can allow for understanding the underlying chemistry and catalyst trends as well as discovering optimal catalysts and operating conditions. We further posit that incorporating available domain information (such as mass balance, plausible reactions, equilibrium/non-equilibrium thermodynamics, etc.) will improve the robustness and generalizability of such models.

In this talk, we will present how such data-driven kinetic models can be developed from HTE data with an illustrative example of oxidative coupling of methane. In particular, we will discuss three approaches. First, we will show that simple neural network models can be trained but keeping in mind plausible reactions of the system. Second, we will show how reactor design equations can be approximated by training residual deep neural networks. Third, we will demonstrate how treating the packed bed reactor system of the experimental set up using neural ordinary differential equations allows for building rate models that are both chemistry cognizant and thermodynamically consistent. This work, thereby, extends the general idea of physics-informed machine learning to building reaction models.