PES-Learn: An Open-Source Software Package for the Automated Generation of Machine Learning Models of Molecular Potential Energy Surfaces

We introduce a free and open-source software package (PES-Learn) which largely automates the process of producing high-quality machine learning models of molecular potential energy surfaces (PESs). PES-Learn incorporates a generalized framework for producing grid points across a PES that is compatible with most electronic structure theory software. The newly generated or externally supplied PES data can then be used to train and optimize neural network or Gaussian process models in a completely automated fashion. Robust hyperparameter optimization schemes designed specifically for molecular PES applications are implemented to ensure that the best possible model for the data set is fit with high quality. The performance of PES-Learn toward fitting a few semiglobal PESs from the literature is evaluated. We also demonstrate the use of PES-Learn machine learning models in carrying out high-level vibrational configuration interaction computations on water and formaldehyde.

我们介绍一种免费且开源的软件包（PES - Learn），它在很大程度上使分子势能面（PESs）的高质量机器学习模型的生成过程自动化。PES - Learn包含一个通用框架，用于在PES上生成网格点，该框架与大多数电子结构理论软件兼容。新生成的或外部提供的PES数据随后可用于以完全自动化的方式训练和优化神经网络或高斯过程模型。专门为分子PES应用设计的稳健超参数优化方案得以实施，以确保为数据集找到尽可能最佳的模型且具有高质量拟合。对PES - Learn在拟合文献中的一些半全局PES方面的性能进行了评估。我们还展示了PES - Learn机器学习模型在对水和甲醛进行高级振动组态相互作用计算中的应用。