Batch active learning for accelerating the development of interatomic potentials
Batch active learning for accelerating the development of interatomic potentials
复制标题
DOI：
10.1016/j.commatsci.2022.111330
复制
发表时间：
2022-06
影响因子：
3.3
通讯作者：
Nathan Wilson;D. Willhelm;Xiaoning Qian;R. Arróyave;X. Qian
中科院分区：
文献类型：
--
作者：
Nathan Wilson;D. Willhelm;Xiaoning Qian;R. Arróyave;X. Qian
经典分子动力学（MD）已被广泛用于研究原子机制和涌现行为的材料在长度和时间尺度上的第一性原理方法的能力之外。经典分子动力学模拟的成功依赖于经典原子间势的能力，以准确地映射到有效的原子的少体相互作用系统的电子和原子核的复杂的多体相互作用系统。在实际中，原子间相互作用势的发展是一个不平凡的过程，需要相当大的努力。近年来，机器学习已成为加速原子间相互作用势发展的一种有前途的方法。然而，这些机器学习方法通常是计算和数据密集型的，因为它们需要来自第一原理计算的大量训练数据，例如许多原子结构的总能量、原子力和应力张量。在这里，我们提出了一种主动学习方法，结合第一性原理理论计算，以加快机器学习原子间势的发展。特别是，我们开发了一种批量主动学习方法，它结合了能量不确定性和结构相似性度量，以有效地对难以预测的高度不确定结构进行采样。这种主动采样方法最大限度地提高了每个批次中数据集的效用，并生成具有高度准确和鲁棒模型系数的原子间势，这是传统采样方法难以实现的。为了证明这批主动学习方法，我们开发了一个主动学习的潜力单层GeSe，一个二维的铁电-铁弹材料，并比较质量和鲁棒性的主动学习的潜力与从随机抽样获得的潜力。批量主动学习方法为加速开发强大而准确的机器学习潜力开辟了途径，这些机器学习潜力使用了一小部分原子结构，这对计算材料，物理和化学社区都很有价值。
Classical molecular dynamics (MD) has been widely used to study atomistic mechanisms and emergent behavior in materials at length and time scales beyond the capabilities of first-principles approaches. The success of classical MD simulations relies on the ability of classical interatomic potentials to accurately map complex many-body interacting systems of electrons and nuclei into effective few-body interacting systems of atoms. In practice, the development of interatomic potentials is a nontrivial process and requires considerable amount of effort. Recently, machine learning has become a promising approach to accelerate interatomic potential development. However, these machine learning approaches are often computation and data intense, as they require a large amount of training data from first-principles calculations, such as total energies, atomic forces, and stress tensors of many atomistic structures. Here we propose an active learning approach combined with first-principles theory calculations to expedite the development of machine learning interatomic potentials. In particular, we develop a batch active learning method which combines both energy uncertainty and structure similarity metrics to efficiently sample the highly uncertain structures that are difficult to predict. This active sampling approach maximizes the utility of the dataset in each batch and generates interatomic potential with highly accurate and robust model coefficients which are difficult to achieve with conventional sampling approaches. To demonstrate this batch active learning method, we develop an active learning potential for monolayer GeSe, a two-dimensional ferroelectric-ferroelastic material, and compare the quality and robustness of the active learning potential with the potential obtained from random sampling. Batch active learning method opens up avenues for accelerating the development of robust and accurate machine learning potential using a small set of atomistic structures which will be valuable for computational materials, physics, and chemistry community.