正文：
近日，Google DeepMind与柏林工业大学的研究人员合作，提出了一种名为SO3krates的Transformer架构，用于加速分子动力学（MD）模拟。该架构结合了稀疏等变表示和自注意力机制，能够显著提高计算速度和稳定性，有望推动分子模拟领域的研究进展。
分子动力学模拟是研究物质微观结构与性质的关键技术，通过模拟原子和分子的运动，揭示物质在时间上的演变。然而，传统模拟方法往往需要大量计算资源，限制了模拟时间的长度和系统的尺度。近年来，机器学习力场（MLFF）的发展为解决这一问题提供了新思路。
MLFF通过学习从大量分子数据中提取的统计信息，可以预测原子间的相互作用力，从而在减少计算资源的同时保持较高的精度。然而，MLFF在长时间尺度上的稳定性问题一直备受关注。DeepMind与柏林工业大学的研究团队意识到，等变表示虽然能够提高数据的可转移性，但相关的计算成本可能会限制其实际应用。
为了解决这一问题，研究团队提出了SO3krates架构。该架构通过使用原子邻域相对方向的过滤器，避免了昂贵的张量积计算，从而提高了计算效率。同时，自注意力机制的运用使模型能够分离不变和等变信息，进一步提高了模型的稳定性和准确性。
实验结果表明，SO3krates模型在保持与当前最先进ML模型相当的稳定性和准确性的同时，速度提高了约30倍。研究人员仅在几个小时内就完成了超分子结构的纳秒级MD模拟，这一成就为研究复杂分子系统的量子特性提供了新的可能。
这项研究的发表在《Nature Communications》上，标志着分子动力学模拟领域的一个重要突破。未来，随着计算方法的不断进步，科学家们有望更深入地理解物质的本质，推动相关领域的科技发展。
英语如下：
Title: “DeepMind Breakthrough: Simulating One Year of Molecular Dynamics in 2.5 Days”
Keywords: DeepMind, Euclidean Transformer, MD acceleration
News Content:
Recently, researchers from Google DeepMind and the Technical University of Berlin have collaborated to develop a new computational method to accelerate molecular dynamics (MD) simulations. They introduced a Transformer architecture named SO3krates, which combines sparse equivariant representations and self-attention mechanisms to significantly enhance computational speed and stability, offering promising advancements for research in the field of molecular simulation.
Molecular dynamics simulation is a critical technique for studying the microscopic structure and properties of matter, revealing the temporal evolution of substances by simulating the motion of atoms and molecules. However, traditional simulation methods often require substantial computational resources, limiting the duration of the simulations and the scale of the systems. In recent years, the development of machine learning force fields (MLFF) has provided a new approach to addressing this issue.
MLFF can predict atomic interactions by learning from statistical information extracted from vast amounts of molecular data, maintaining high precision while reducing computational resources. However, the stability issue of MLFF over long timescales has been a matter of concern. The research team, recognizing that equivariant representations can enhance data transferability, also noted that the related computational cost might limit practical applications.
To address this challenge, the research team proposed the SO3krates architecture. This architecture avoids expensive tensor product calculations by using filters for relative directions of atomic neighborhoods, improving computational efficiency. The application of self-attention mechanisms allows the model to separate invariant and equivariant information, further enhancing stability and accuracy.
Experimental results show that the SO3krates model maintains a similar level of stability and accuracy as the current state-of-the-art ML models, while achieving a speedup of about 30 times. Researchers were able to complete nanosecond-scale MD simulations of supramolecular structures within a few hours, opening new possibilities for studying the quantum characteristics of complex molecular systems.
This research was published in Nature Communications, marking an important breakthrough in the field of molecular dynamics simulation. With ongoing advancements in computational methods, scientists are poised to gain deeper insights into the nature of matter and drive technological developments in related fields.
【来源】https://www.jiqizhixin.com/articles/2024-08-09-8
Views: 3