Recurrent neural networks (RNNs) are powerful tools for forecasting sequences of variable lengths. We apply this technique to predict changes in the concentrations of multiple species resulting from a series of aqueous inorganic iodine reactions. We train two RNNs in the case that 1) concentrations over a number of time steps are known or 2) only the initial conditions are known. The first model provides good accuracy but has poor interpretability due to the large number of parameters internal to the model. The second model allows for only the initial concentrations to be known and provides better interpretability – the model weights encode the trajectories – but the capacity of the model is restricted. Furthermore, a physical constraint is applied to the loss function to force the system to conform to physical laws, which leads to more physically relevant predictions compared with the same network without the constraint. As a proof of concept of our approach, a test dataset, generated by solving the kinetic equations of the studied reactions using rate constants found in the literature, was applied to the models, and the results were compared. In addition to comparing the results of the two models, we examine the minimum dataset necessary to obtain accurate predictions, as well as the effects of using noisy data to train the model.

Using domain-aware machine learning to forecast aqueous iodine reactions

2019 Fall National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.