EAGER: XAISE: Explainable Artificial Intelligence for Science and Engineering

前三种科学范式(实验、理论和模拟)数据的日益可获得性，以及人工智能和机器学习(AI/ML)的进步，为使用数据驱动的科学加速科学发现提供了前所未有的机会。尤其是深度学习已经成为一种从材料科学、生命科学、药物设计等科学领域的海量数据中获得见解的变革性技术。然而，深度学习模型的可解释性和可解释性仍然是一个主要问题和悬而未决的问题。对可解释人工智能的需求在科学和工程中往往是至关重要的，具有国家重要性的应用程序，如材料设计、建筑、运输、健康科学、能源储存等，在这些领域，错误决策的成本可能是灾难性的巨大，这使得确保模型不仅在数量上准确，而且实际上是从正确的特征中学习，并以可理解的方式学习有意义的东西是至关重要的。但是，对可解释性的抽象看法是极其困难的，因为解释还需要应用程序域中的上下文。该项目寻求通过结合和利用来自科学应用领域的上下文以及探索传统的机器学习技术来开发在数字图书馆的使用中的地址解释性。本项目旨在探索和研究一种ML-DL集成的方法，以实现符合NIST(国家标准与技术研究所)四项原则的可解释人工智能。该项目的具体目标是：设计、开发和实施XAISE-一个框架，以最小的影响提高人工智能模型在科学和工程应用中的可解释性；使XAISE适用于不同的数据类型，例如数字、图像等；将XAISE扩展到能够处理大型、多维数据；以及评估XAISE在至少两个应用领域的适用性，包括材料科学和纳米技术。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

The increasing availability of data from the first three paradigms of science (experiments, theory, and simulations), along with advances in artificial intelligence and machine learning (AI/ML) has offered unprecedented opportunities for accelerating scientific discoveries using data-driven science. In particular Deep Learning (DL) has emerged as a transformative technology for deriving insights from massive datasets in many scientific domains such as material science, life-science, drug design etc. However, interpretability and explainability of DL models remains a major issue and an open problem. The need for explainable AI is often crucial in science and engineering, with applications of national importance such as materials design, construction, transportation, health-sciences, energy storage, etc., where the cost of wrong decisions can be catastrophically large, making it critical to ensure that the model is not just quantitatively accurate but is in fact learning from the correct features, and learning things that make sense in an understandable manner. But an abstract view of explanability is extremely difficult, because explanation also requires context within the application domain. This project seeks to develop addresses explanability within the use of DL by incorporating and utilizing context from scientific application domains and by exploring traditional machine learning techniques. This project seeks to explore and investigate an approach of ML-DL integration to realize explainable AI in terms of the four NIST (National Institute of Standards and Technology) principles. The specific goals of this project are: to design, develop, and implement XAISE – a framework to enhance the explainability of AI models for science and engineering applications with minimal impact on accuracy; to adapt XAISE for heterogenous data types, e.g., numerical, images, etc.; to scale XAISE to be able to handle large, multi-dimensional data; and evaluate the applicability of XAISE for at least two application domains, including materials science and nanotechnology.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.