Physics-Based Machine Learning for Artificial Intelligence

Machine learning models have already found considerable success for many commercial applications, including scientific problems that involve processes that are not entirely understood or where mechanistic models are impractical because of the computational resources required. Such applications are especially suited for deep neural networks (DNN) because of their performance on spatiotemporal data. However, DNN’s success is limited in many applications due to the need for large data sets, poor extrapolation accuracy, and lack of interpretability. This focus area seeks technologies that integrate physics-based insight to help interpret the DNN solution and constrain the search space to help alleviate training and validation data requirements. Of particular interest are technologies that combine physics and DNN, to enable fast models for predicting complex system behavior in training, optimizing, and testing autonomous systems.