高级检索

深度学习与第一性原理计算

Deep learning and first-principles calculations

  • 摘要: 第一性原理计算基于量子力学基本原理,通过求解复杂的多电子相互作用问题实现高精度材料计算预测,已成为现代物理学、化学、材料科学等诸多领域中不可或缺的研究手段。然而,高昂的计算成本限制了第一性原理计算的广泛应用,使得大尺度材料模拟和材料大数据构建等重要领域的发展面临重大挑战。近年来,AlphaGo、AlphaFold、ChatGPT等突破性工作的涌现宣示了人工智能新时代的来临,第一性原理计算领域也迎来了变革性转变的历史机遇。深度学习为第一性原理计算提供了新的研究范式,通过精确建模和高效预测,有望突破传统方法的瓶颈问题。文章介绍了一类基于深度学习的第一性原理计算方法,利用神经网络对密度泛函理论中的核心物理量——密度泛函理论哈密顿量进行建模,并设计出满足局域性原理、协变性原理等关键物理先验的先进神经网络架构,实现了高效精确的深度学习电子结构计算。该方法已成功应用于转角范德瓦耳斯材料等体系的大尺度材料模拟、基于材料大数据的通用材料模型构建等极具挑战性的任务中,为发展材料大模型、推动人工智能驱动的材料发现提供了新的机遇。

     

    Abstract: First-principles methods have become indispensable research tools in modern physics, chemistry, materials science and other fields. Based on the fundamental principles of quantum mechanics, first-principles calculations can achieve highly accurate material property predictions by solving the complicated problem of interacting electrons. However, their widespread applications are limited by the high computational cost, posing significant challenges for large-scale materials simulations and the construction of materials big data. In recent years, the emergence of groundbreaking works, such as AlphaGo, AlphaFold, and ChatGPT, heralds the advent of a new era of artificial intelligence (AI), bringing transformative opportunities to first-principles calculations. Deep learning provides a novel research paradigm for the field, enabling us to overcome the bottlenecks of traditional methods through precise modeling and efficient prediction. In this article, we introduce a series of deep-learning based first-principles computation methods. By leveraging the deep-learning modeling of a fundamental quantity in density functional theory (DFT), the DFT Hamiltonian, we propose a neural network framework designed to satisfy the key prior knowledge of the Hamiltonian, including the principles of nearsightedness and equivariance. This method has been successfully applied to large-scale simulations of twisted van der Waals materials, as well as the construction of universal models based on materials big data. These advancements offer new opportunities for developing large materials models and advancing AI-driven materials discovery.

     

/

返回文章
返回