CAREER: Accelerating Scientific Discovery via Deep Learning with Strong Physics Inductive Biases

科学发现--从对复杂物理过程的经验观察中获得关于自然世界的新知识的过程--长期以来一直是研究的主题。虽然历史方法往往需要有才华的学者终身收集数据和分析，但现代计算技术已经彻底改变了科学发现的方法。如今，先进的计算机器和机器学习（ML）促进了大规模复杂数据的快速分析。然而，目前的方法通常依赖于不透明的“黑盒”ML模型。它们缺乏物理上的一致性、可概括性和/或可解释性。为了解决这些限制，这个职业生涯项目的重点是建立一个协同的研究和教育计划，以推进科学发现的方法。该项目的核心研究思想是将物理领域知识嵌入深度学习（DL）模型的设计中。通过使用精心制定的假设和大量数据收集来增强强有力的基于物理的模型设计，这种方法将产生新的DL模型，这些模型强制执行物理原理，以严格限制模型结果的搜索空间，使其在物理上一致。这种创新的方法有望大大加快科学发现的进程，并提高其效率，准确性和可解释性。该项目的预期成果也广泛适用于热力学和分子动力学等相关研究领域。该项目研究数据驱动的复杂时空物理过程的科学发现。它旨在实现以下具体目标：（1）物理一致的动力学模型，通过对神经网络的设计施加强物理偏差来精确地保留重要的物理特性;（2）基于DL的符号回归算法，以推断目标动力学模型的可解释的数学表达式，其中模型结构被设计为符合物理原理;以及（3）用于科学应用的多模态DL算法，其利用以不同模态存在的数据。此外，该项目旨在通过利用上述任务的发现，为数据科学下游任务开发（4）物理启发的DL模型，为ML研究做出贡献。将根据项目成果开发一门跨学科课程，以教育研究人员、各级学生和更广泛的社区。最后，该项目为学生提供研究机会，优先从代表性不足的群体中招募。该奖项反映了NSF的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

Scientific discovery--the process of gaining new knowledge about the natural world from empirical observations of complex physical processes--has long been a topic of research. While historical approaches often necessitated the lifelong collection of data and analysis of talented scholars, modern computing technologies have revolutionized methods of scientific discovery. Today, advanced computing machines and machine learning (ML) facilitate the rapid analysis of large-scale complex data. However, current approaches often rely on opaque "black-box" ML models. They suffer from a lack of physically-consistency, generalizability, and/or interpretability. To address these limitations, this CAREER project focuses on establishing a synergistic research and education program to advance methods of scientific discovery. The core research idea of the project is to embed physics domain knowledge into the design of deep-learning (DL) models. By augmenting strong physics-based model designs with carefully formulated hypotheses and massive data collections, this approach will produce novel DL models that enforce the principles of physics to strongly restrict search spaces of model outcomes to be physically consistent. This innovative approach is expected to greatly accelerate the process of scientific discovery and enhance its efficiency, accuracy, and interpretability. The expected outcome of this project is also widely applicable to relevant research areas such as Thermodynamics and Molecular dynamics.This project investigates data-driven scientific discovery of complex spatiotemporal physical processes. It aims to achieve the following specific goals: to develop (1) physically-consistent dynamics models, which exactly preserve important physical characteristics by imposing strong physics biases to the design of neural networks; (2) DL-based symbolic regression algorithms to infer interpretable mathematical expressions of target dynamics models, where the model structures are designed to conform to principles of physics; and, (3) multimodal DL algorithms for scientific applications that leverage data that exist in different modalities. In addition, this project seeks to develop (4) physics-inspired DL models for data science downstream tasks by leveraging findings from the above tasks, contributing to ML research. An interdisciplinary course will be developed based on the project outcomes to educate researchers, students at all levels, and the broader communities. Lastly, this project offers research opportunities for students, prioritizing recruitment from underrepresented groups.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.