colabfit/discrepencies_and_error_metrics_NPJ_2023_interstitial_enhanced_training_set

差异和误差指标NPJ 2023间隙增强训练集包含了经过增强处理的原子结构数据，这些数据通过包含迁移间隙原子来修改原始的mlearn_Si_train数据集，目的是开发具有更好扩散性评分的模型。增强的验证集包括50个结构，其中20个来自原始训练数据集被替换的结构，11个是来自AIMD模拟的空位罕见事件快照，19个是来自AIMD模拟的间隙原子罕见事件快照。此外，还构建了间隙原子罕见事件和空位罕见事件的测试集，每个测试集包含100个原子构型快照，分别来自1230 K下AIMD模拟的单个迁移空位或间隙原子。

Structures from discrepencies_and_error_metrics_NPJ_2023 training set, enhanced by inclusion of interstitials. The full discrepencies_and_error_metrics_NPJ_2023 dataset includes the original mlearn_Si_train dataset, modified with the purpose of developing models with better diffusivity scores by replacing ~54% of the data with structures containing migrating interstitials. The enhanced validation set contains 50 total structures, consisting of 20 structures randomly selected from the 120 replaced structures of the original training dataset, 11 snapshots with vacancy rare events (RE) from AIMD simulations, and 19 snapshots with interstitial RE from AIMD simulations. We also construct interstitial-RE and vacancy-RE testing sets, each consisting of 100 snapshots of atomic configurations with a single migrating vacancy or interstitial, respectively, from AIMD simulations at 1230 K.