CAREER: Domain-aware Statistical Learning

点击翻译按钮获取中文摘要

This Faculty Early Career Development Program (CAREER) grant will contribute to the advancement of national competitiveness by transforming how governing physics and engineering domain knowledge is integrated into data-driven models for high-stakes applications. Applications in domain-knowledge intensive engineering environments such as energy infrastructure, aviation safety, and manufacturing require interpretable models, explainable decisions and actionable insights. In these environments, the old paradiam of “letting the data speak for themselves” is being replaced by the capability of “letting the data speak based on the laws of physics and engineering”. This project will address the development of methods to integrate data with physics-based models in three main use cases, namely environmental processes to enhance resilience of our national utilities during extreme events intensified by climate change; thermal modeling to improve energy efficiency in Data Center operations; and structural dynamics to enhance aviation safety in an increasingly crowded airspace. The accompanying educational plan aims to address the gaps between general-purpose data science education at the school and university level and the specific needs for next-generation engineering students with diverse backgrounds. The educational plan also aims to improve data literacy among the general public by improving awareness of the increasing availability of data and the capability of interpreting those data through local community activities.This research establishes a new Structure-Exploiting-Preserving (SEP) domain-aware statistical learning paradigm that enables the direct embedding of governing physics into data-driven models during model construction. Unlike existing approaches that impose governing physics as auxiliary regularizations or constraints, the SEP framework will enable the embedding of data-driven models into the solution space of governing physics (i.e., governing physics will no longer be used as auxiliary regularizations, but an inherent component that is directly integrated into data-driven models during model construction). This is achieved by exposing the solution structure of governing physics (i.e., structure exploiting), and embedding data-driven models into the solution space of governing equations (i.e., structure preserving). Under the SEP framework and through the collaboration with industry partners, the research will initiate a trajectory that leads to a set of new methodologies for domain-aware statistical modeling, data-driven discovery of governing physics, and dynamic sampling for non-stationary engineering processes.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.