CAREER: Composite Physics-Informed Learning of Dynamic Systems

信息物理系统（cps）是许多现代工程系统的核心技术，从汽车、机器人、医疗设备、建筑到电网和先进制造系统。随着这些系统数据的广泛可用性，机器学习（ML）和人工智能（AI）在许多CPS应用中取得了巨大成功。然而，它们目前面临的根本挑战是，它们通常需要大数据，可能违反导致性能不佳甚至失败的基本物理原理，并且不能健壮地处理来自现实系统的混乱数据。该项目在网络基础设施（CI）中创建了新的方法、算法和软件，将ML/AI与传统物理知识无缝地协同集成在所谓的物理知情机器学习（PIML）模型中，可以克服这些挑战。该模型建立在统一的PIML理论基础、将异构模型组合成复合PIML模型的框架和软件以及提高模型效率和准确性的新方法的基础上。开发的技术将推动机器学习/人工智能在cps中的前沿，为克服固有挑战，提高人工智能驱动的cps的性能和安全性开辟新的令人兴奋的途径，从而扩大其实际应用。该项目将研究活动与教育活动深度结合，激发和培养本科生和研究生在计算机科学与工程领域的体验式学习和研究经验，并通过与当地学校和公共项目的合作，促进弱势群体参与STEM，丰富公众对STEM的理解。该项目服务于国家利益，正如NSF的使命所述，通过促进科学的进步，促进国家的健康、繁荣和福利。该项目的总体目标是将机器学习和物理集成到一个全面、灵活和协同的CI中，用于复合PIML和动态系统的主动学习。为此，其目标是：(1)建立统一的PIML框架的理论基础；(2)组合模型和复合模型物理性质的理论框架和软件；(3)基于物理的主动学习方法，直接整合物理，获取最符合物理的信息数据，提高学习的样本效率和准确性。本研究提出了PIML方法的统一、PIML的收益和成本、如何在异构PIML模型中有效和高效地组合模型和物理性质以及如何将物理性质集成到主动学习中的知识状况。它还创建了方法和软件，可以快速开发和探索动态系统的新型数据驱动建模方法，突破极限，增强机器学习在cps中的适用性和性能。通过为集成物理和机器学习建立坚实的基础，以产生准确、可解释、健壮和物理一致的模型，CI将促进cps的高性能数据驱动预测、仿真、优化和控制方法，从而受益于广泛的科学和工程应用。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Cyber-physical systems (CPSs) are core technologies in many modern engineering systems, spanning from automobiles, robots, medical devices, buildings, to power grids and advanced manufacturing systems. With the wide availability of data from these systems, machine learning (ML) and artificial intelligence (AI) have found great success in many CPS applications. However, their current fundamental challenges are that they often require big data, may violate basic physical principles leading to underperformance or even failures, and do not robustly handle messy data from real-life systems. This project creates new methods, algorithms, and software in a cyberinfrastructure (CI) that seamlessly and synergistically integrate ML/AI with traditional physical knowledge in so-called physics-informed machine learning (PIML) models that can overcome these challenges. The CI is built upon a unified theoretical foundation of PIML, a framework and software for composing heterogeneous models into composite PIML models, and novel methods for improving their efficiency and accuracy. The developed technologies will push forward the frontiers of ML/AI in CPSs to open up new exciting pathways for overcoming the inherent challenges and enhancing the performance and safety of AI-driven CPSs, thus broadening their real-life applications. This project deeply integrates research activities with education activities to excite and foster experiential learning and research experience in computer science and engineering at the undergraduate and graduate levels, and to promote STEM participation among underrepresented groups and enrich public understanding through collaboration with local schools and public programs. The project serves the national interest, as stated by NSF's mission, by promoting the progress of science, and to advance the national health, prosperity, and welfare.The overarching goal of this project is to integrate ML and physics within a comprehensive, flexible, and synergistic CI for composite PIML and active learning of dynamic systems. To this end, its objectives are to develop (1) a theoretical foundation of unified PIML frameworks; (2) a theoretical framework and software for composing models and physical properties in composite PIML models; and (3) physics-informed active learning methods which directly integrate physics to obtain the most informative data consistent with physics for improving the sample efficiency and accuracy of learning. This research advances the state of knowledge regarding unification of PIML methods, the benefits and costs of PIML, how to effectively and efficiently compose models and physical properties in a heterogeneous PIML model, and how to integrate physical properties into active learning. It also creates methodologies and software that enable rapid development and exploration of novel data-driven modeling methods for dynamic systems, pushing the limits and enhancing the applicability and performance of ML in CPSs. By building a solid foundation for integrating physics and ML to yield accurate, interpretable, robust, and physically consistent models, the CI will facilitate high-performance data-driven prediction, simulation, optimization, and control methods for CPSs, benefiting a broad range of scientific and engineering applications.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.