CAREER: Combining Machine Learning and Physics-based Modeling Approaches for Accelerating Scientific Discovery

点击翻译按钮获取中文摘要

Recent advances in machine learning make it possible to capture spatial and temporal dependencies from complex data, leading to tremendous success in several commercial applications where large-scale data are available. Given their success in commercial domains, machine-learning models are beginning to play an important role in advancing scientific discovery in diverse scientific domains, which are traditionally dominated by physics-based models. The use of machine-learning models is especially promising when relevant physical processes are not completely understood by our current body of knowledge due to the inherent complexity of the underlying phenomenon. However, the direct application of machine-learning models has met with limited success in real-world scientific applications, given that the data sets available for many scientific problems are far smaller than what is needed to effectively train advanced machine-learning models. Recently, there is a growing interest in combining machine learning and physical knowledge for studying scientific problems. Existing methods on this topic have shown promising results in isolated experiments, but they remain limited in real-world scientific applications with scarce training data, data variability, and incomplete or approximate physical information. The goal of this project is to systematically explore ways to combine machine-learning models and existing physics-based modeling approaches in a synergistic manner to model complex, non-stationary, and spatio-temporal processes for scientific problems. This project aims to explore a deep coupling of machine learning and physics-based models for modeling physical systems via four innovations. First, a new knowledge-guided machine-learning architecture will be built for capturing complex dynamics and interactions amongst physical variables, which helps improve the model prediction and generalizability. Second, new data-driven inverse models will be investigated for the discovery of physical parameters, which aids in improving the performance of physics-based models. Third, new model pre-training strategies will be developed to enable knowledge-guided machine-learning algorithms to work effectively even with a small number of observations. These pre-training strategies will leverage simulated data sets generated by physics-based models. New techniques will also be proposed to guide the configuration of physics-based models with the aim to create new simulated data sets for improving the pre-training effectiveness. Finally, new meta-learning approaches will be created for adapting the knowledge-guided machine-learning model over space and time. The proposed methods will leverage physical knowledge to estimate the similarity amongst different physical systems and facilitate model adaptation.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.