Files for Training/Inferencing Mixture-of-Experts Transformers for Deorbitalized Meta-GGA Density Functionals

This dataset contains the pre-processing, training and inference files to learn deorbitalized surrogate models for orbital-dependent, semi-local meta-GGA exchange correlation functionals using Mixture-of-Experts transformers. Expert domains are defined using K-means clustering of principal components of orbital-dependent functional predictions. Expert models are trained using a transfer learning protocol of pre-training and fine-tuning. A gating network is trained to route integration grid points to each expert model. The files in the "Inference" and "System Input Files" directories are required to run the model scripts in "https://github.com/amitmc1/MetaGGA-Transformers/tree/main/Inference".

本数据集包含用于借助混合专家Transformer（Mixture-of-Experts Transformers）构建轨道依赖型半局域meta-GGA（meta-Generalized Gradient Approximation）交换关联泛函的去轨道化替代模型的预处理、训练与推理文件。专家域通过对轨道依赖型泛函预测结果的主成分进行K-means聚类定义。专家模型采用预训练与微调的迁移学习协议进行训练。门控网络经训练后可将积分网格点分配至各专家模型。运行仓库"https://github.com/amitmc1/MetaGGA-Transformers/tree/main/Inference"中的模型脚本时，需使用"Inference"与"System Input Files"目录下的文件。