Subset of stochastically generated interacting molecules for CH GAP interatomic potential

This is a subset of the dataset used to train general-purpose CH GAP interatomic potential [1].

This subset contains the interacting molecules generated stochastically in a following manner. The subset is generated using active learning and uncertainty-based configuration selection. We started with randomly chosen pairs of CH-containing molecules from the QM9 database up to 7 carbon atoms. The probability of selecting the molecules is set based on the energy and size of the molecule.  The probability is lower as the energy above the convex hull is higher.  A bigger size of the molecule also lowers the probability to favor the inclusion of small structures. Then, we estimate the uncertainties for the new structures based on how far away from the existing interacting molecules in the training set they are (in configuration space), and identify those with the largest expected errors. This way, we generated about 3k structures.

本数据集为训练通用型CH GAP原子间相互作用势（interatomic potential）所用原始数据集的子集[1]。
该子集包含按下述方式随机生成的相互作用分子体系。本子集通过主动学习与基于不确定性的构型筛选方法生成。我们首先从最多含7个碳原子的QM9数据库中，随机选取含碳氢（CH）的分子对作为初始样本。分子的选取概率由其能量与尺寸共同决定：相较于凸包（convex hull）能量差更高的分子，其选取概率更低；分子尺寸越大，选取概率也越低，以此优先纳入小分子结构。随后，我们基于新结构与训练集中已有相互作用分子体系在构型空间（configuration space）中的距离，估算其不确定性，并筛选出预期误差最大的结构。通过该流程，我们共生成约3000个结构。