(149a) Exploiting Knowledge When Learning Reaction Rates with Neural ODEs from Experimental Observations

In this work, we investigate the effects of incorporating structural knowledge on the accuracy of a trained NODE (Neural Ordinary Differential Equations). Using a simulation case study, we investigated this effect when perfect knowledge is available, and found that such information does appear to improve the final model. We further investigated this phenomena on Grignard data set, and found incorporating structural assumptions based off the expert models reduced the final model’s performance. Thus, incorporating structural information appears to reduce the expressiveness of a NODE. When the information accurately reflects the underlying mechanics, the reduction in expressibility converges toward the truth. However, inaccurate information may reduce expressibility such that an accurate model cannot be expressed. Future work may investigate combining an interpretable model formulation, such as SISSO into the NODE-style framework. This can be accomplished by training an interpretable model via the adjoint sensitive method used in the NODE framework.