Machine learning dataset, model weights, and training logs for the publication "Broken neural scaling laws in learning the optical properties of solids"

This repository contains a ready-to-use machine learning dataset, model weights, and training logs associated with the publication "Broken Neural Scaling Laws in Learning the Optical Properties of Solids", published in PRX Intell. 1, 013011 (2026), https://doi.org/10.1103/wv8r-7zr1.The code implementing the graph neural network models and training workflows used with this dataset is available at: https://github.com/MaxGrossmann/optimetalThe data are provided in the form of compressed archives that are organized as follows.Dataset archive:The archive "dataset.zip" contains the following three HDF5 files: "train.h5", "val.h5", and "test.h5". These files store the training, validation, and test splits used for the reported models. Details on the data format and parsing are provided in the associated code repository.The underlying raw dataset was generated using a high-throughput ab initio workflow and is distributed across two repositories due to its size: https://doi.org/10.6084/m9.figshare.31111798 and https://doi.org/10.6084/m9.figshare.31112491.Graph archives:The "graph.zip" and "graph_e3.zip" archives each contain three corresponding PyTorch files: "train.pt", "val.pt", and "test.pt". These files store the corresponding graph representations used as input for the invariant and equivariant graph neural networks, respectively. Parsing and usage of these files are documented in the associated code repository.Model weights archive:The "adapted_uma_model_state_dicts.zip" archive contains the trained model state dictionaries for the equivariant graph neural networks used in the publication. The corresponding model weights for the invariant OptiMetal architectures are available in the aforementioned code repository.Training logs archive:The "training_logs.zip" and "training_logs_e3.zip" archives contain directories corresponding to specific groups of training runs. Within each directory, the subdirectories are named according to the model configuration, random seed, and (where applicable) additional hyperparameters. These subdirectories contain TensorBoard log files, job submission scripts (LSF files), and other outputs produced during training.Together, these data, model weights, graph representations, and training logs enable reproduction of the training, evaluation, and neural scaling law analyses reported in the publication.