Fortnet, a software package for training Behler-Parrinello neural networks

A new, open source, parallel, stand-alone software package (Fortnet) has been developed, which implements Behler-Parrinello neural networks. It covers the entire workflow from feature generation to the evaluation of generated potentials, coupled with higher-level analysis such as the analytic calculation of atomic forces. The functionality of the software package is demonstrated by driving the training for the fitted correction functions of the density functional tight binding (DFTB) method, which are commonly used to compensate the inaccuracies resulting from the DFTB approximations to the Kohn-Sham Hamiltonian. The usual two-body form of those correction functions limits the transferability of the parametrizations between very different structural environments. The recently introduced DFTB+ANN approach strives to lift these limitations by combining DFTB with a near-sighted artificial neural network (ANN). After investigating various approaches, we have found the combination of DFTB with an ANN acting on-top of some baseline correction functions (delta learning) the most accurate one. It allowed to introduce many-body corrections on top of two-body parametrizations, while excellent transferability to chemical environments with deviating energetics could be demonstrated.本研究开发了一款全新的开源并行独立软件包（Fortnet），其实现了Behler-Parrinello神经网络（Behler-Parrinello neural networks）。该软件覆盖从特征生成到生成势评估的完整工作流，并集成了原子力解析计算等高级分析功能。本软件的功能通过针对密度泛函紧束缚（DFTB）方法的拟合校正函数开展训练得到验证——这类校正函数通常用于弥补因DFTB对Kohn-Sham哈密顿量（Kohn-Sham Hamiltonian）的近似所带来的计算偏差。这类校正函数通常采用双体形式，这限制了其参数化方案在差异极大的结构环境间的可迁移性。近期提出的DFTB+ANN方法旨在通过将DFTB与近邻人工神经网络（ANN）结合，突破上述局限。经对多种方案开展研究后，我们发现将DFTB与基于基准校正函数的人工神经网络（delta learning）结合的方案精度最高。该方案可在双体参数化的基础上引入多体校正，同时其对能量偏差较大的化学环境展现出优异的可迁移性。