Abstract Atomistic modeling is important for studying physical and chemical properties of materials. Recently, machine learning interaction potentials have gained much more attentions as they can provide density functional theory level predictions within negligible time. The symmetry function descriptor based atomistic neural network is the most widely used model for modeling alloys. To precisely describe complex potential energy surfaces, integrating advanced metrics, such as force or virial stress, into training can be of great help. In this work, we propose a virtual-atom approach to model the total energy of symmetry function descriptors based atomistic neural network. Our approach creates the computation graph directly from atomic positions. Thus, the derivations of forces and virial can be handled by TensorFlow automatically and efficiently. The virtual atom approach with AutoGrad within TensorFlow allows for efficient training to not just energies and forces, but also virial stress. This new approach is implemented in our open-source program TensorAlloy, which supports constructing machine learning interaction potentials for both molecules and solids. The QM7 and SNAP/Ni–Mo datasets are used to demonstrate the performances of our program. Program summary Program Title: TensorAlloy Program Files doi: http://dx.doi.org/10.17632/w8htd7vmwh.1 Licensing provisions: LGPL Programming language: Python 3.7 Nature of problem: Modeling interatomic interactions with the symmetry function descriptor based atomistic neural networks. Solution method: The TensorAlloy program is built upon TensorFlow and the virtual-atom approach. TensorAlloy can build direct computation graph from atomic positions to total energy. Atomic forces and virial stress tensors are handled by TensorFlow automatically and efficiently. Additional comments including restrictions and unusual features: This program needs TensorFlow 1.13.*. Neither newer or older TensorFlow is supported.

中文翻译：

---

TensorAlloy：合金自动原子神经网络程序

摘要 原子建模对于研究材料的物理和化学性质具有重要意义。最近，机器学习交互潜力得到了更多的关注，因为它们可以在可以忽略不计的时间内提供密度泛函理论水平的预测。基于对称函数描述符的原子神经网络是最广泛使用的合金建模模型。为了精确描述复杂的势能面，将力或力压等高级指标整合到训练中会大有帮助。在这项工作中，我们提出了一种虚拟原子方法来模拟基于原子神经网络的对称函数描述符的总能量。我们的方法直接从原子位置创建计算图。因此，TensorFlow 可以自动有效地处理力和维里尔的推导。TensorFlow 中 AutoGrad 的虚拟原子方法不仅可以对能量和力进行有效训练，还可以对力应力进行有效训练。这种新方法在我们的开源程序 TensorAlloy 中实现，它支持构建分子和固体的机器学习相互作用势。QM7 和 SNAP/Ni-Mo 数据集用于演示我们程序的性能。程序摘要 程序名称：TensorAlloy 程序文件 doi：http://dx.doi.org/10.17632/w8htd7vmwh.1 许可条款：LGPL 编程语言：Python 3.7 问题性质：使用基于原子神经网络的对称函数描述符建模原子间相互作用. 解决方法：TensorAlloy 程序建立在 TensorFlow 和虚拟原子方法之上。TensorAlloy 可以构建从原子位置到总能量的直接计算图。原子力和维里应力张量由 TensorFlow 自动有效地处理。包括限制和异常功能在内的其他评论：此程序需要 TensorFlow 1.13.*。不支持较新或较旧的 TensorFlow。