State Estimation Using Physics Constrained Neural Networks

The ability of Neural Networks (NN) to exploit the input-output relationship given by the data is limited by the quality and availability of the data. The unavailability of complete and accurate mechanistic models to represent knowledge inferred by the first principles renders them unsuitable for many applications. To overcome this issue, an approach to train the NN model by capturing the information from the available data, as well as the model, has been evaluated in this paper. The approach is also known as Physics Informed Neural Network-PINNs (Raissi et al. 2019) and has been applied to estimate the measured outputs and internal states of a continuous stirred tank reactor system. It is observed that this approach can give more accurate predictions in estimating the outputs and states.