人工智能首次攻克量子物理学难题 DeepMind 精确计算量子激发态 登上《科学》杂志
北京时间 2024 年 8 月 23 日 – 人工智能领域取得重大突破！来自 Google DeepMind 的研究人员开发了一种全新的方法，成功地利用人工智能精确计算了量子激发态，这一成果被认为是深度学习首次准确解决量子物理学中一些最难的问题，相关研究论文已发表在顶级学术期刊《科学》杂志上。
此前，DeepMind 研发的费米子神经网络 (FermiNet) 在对大量电子的量子基态进行建模方面表现出色，但其最初主要应用于分子的基态。而当分子和材料受到光或高温等能量刺激时，电子会跃迁到更高的能量状态，即激发态。激发态在物理学和化学等领域至关重要，但从第一原理出发对激发态特性进行可扩展、准确且稳健的计算一直是科研人员面临的重大挑战。
DeepMind 研究人员此次开发的全新方法，突破了传统方法的局限性，能够应用于任何类型的数学模型，包括 FermiNet 和其他神经网络。该方法通过变分蒙特卡罗估计量子系统激发态，无需自由参数或正交化，而是将问题转化为寻找扩展系统基态的问题。研究人员将其命名为自然激发态 VMC (NES-VMC)。
NES-VMC 方法能够准确地恢复一系列分子的激发能量和振荡器强度，在从单个原子到苯大小的分子基准系统上都取得了令人瞩目的成果。研究结果表明，该方法在第一行原子上的准确性与实验结果高度吻合。
论文第一作者兼通讯作者 David Pfau 兴奋地表示：“这是深度学习首次准确解决量子物理学中一些最难的问题。我们希望这将是深度学习通用量子模拟迈出的新的一步。”
这一突破性成果不仅为量子物理学研究开辟了新的道路，也为人工智能在科学领域的应用提供了新的可能性。未来，DeepMind 的研究团队将继续探索 NES-VMC 方法的应用潜力，例如，将其用于设计新型材料和催化剂，以及理解复杂生物过程等。
英语如下：
AI Cracks Quantum Puzzle! DeepMind’s Precise Calculation of Excited States Published inScience
Keywords: AI, Quantum, Breakthrough
Content:
August 23, 2024, Beijing Time – A major breakthrough has been achieved in the field of artificial intelligence! Researchers from Google DeepMind have developed a novel method that successfully uses AI to precisely calculate quantum excited states. This achievement is considered the first time deep learning has accurately solved some of themost challenging problems in quantum physics. The related research paper has been published in the top academic journal Science.
Previously, DeepMind’s developed Fermionic Neural Network (FermiNet) excelled in modeling the quantum ground states of alarge number of electrons, but its initial applications were primarily focused on the ground states of molecules. When molecules and materials are stimulated by energy sources like light or high temperatures, electrons transition to higher energy states, known as excited states. Excited statesare crucial in fields like physics and chemistry, but achieving scalable, accurate, and robust calculations of their properties from first principles has been a significant challenge for researchers.
The new method developed by DeepMind researchers breaks through the limitations of traditional methods and can be applied to any type of mathematical model, including FermiNet and otherneural networks. The method estimates the excited states of quantum systems through variational Monte Carlo, eliminating the need for free parameters or orthogonalization. Instead, it transforms the problem into finding the ground state of an extended system. The researchers have named this method Natural Excited State VMC (NES-VMC).
The NES-VMC method accurately recovers the excitation energies and oscillator strengths of a range of molecules, achieving remarkable results on benchmark systems ranging from single atoms to molecules the size of benzene. The research results show that the method’s accuracy on first-row atoms closely matches experimental findings.
David Pfau, the first author and correspondingauthor of the paper, excitedly stated, “This is the first time deep learning has accurately solved some of the most challenging problems in quantum physics. We hope this will be a new step towards general-purpose quantum simulation with deep learning.”
This groundbreaking achievement not only opens up new avenues for quantum physics research but also providesnew possibilities for the application of AI in science. In the future, DeepMind’s research team will continue to explore the application potential of the NES-VMC method, such as using it to design new materials and catalysts, as well as understand complex biological processes.
【来源】https://www.jiqizhixin.com/articles/2024-08-23-6
Views: 1