Dataset of Carbon Nanostructures for "Prediction of Carbon Nanostructure Mechanical Properties and the Role of Defects Using Machine Learning"

This repository contains the full 3D structure database associated with the publication "<b>Prediction of Carbon Nanostructure Mechanical Properties and the Role of Defects Using Machine Learning</b>"The dataset (SI_complete_dataset) contains 1179 3D atomic structures of CNT bundles, 958 structures of CNT junctions, and 50 structures of carbon fiber cross-sections with associated mechanical properties from complete stress-strain curves up to failure for each structure (e.g., strain at break, Young's modulus, and tensile strength). The models have a size of up to 80,000 atoms and the ground truth data were derived using the reactive INTERFACE force field, IFF-R.The database is extensible and can include larger carbon nanostructures with labels, including data using multiple computational and experimental techniques as they become available. The goal is real-time prediction of stress-strain properties of carbon nanostructures of arbitrary 3D configurations.The fileshare also contains a second folder (HS-GNN_and_small_test_model.zip) that contains the hierarchical spatial graph neural network (HS-GNN) and a runscript for XGBoost to train and apply the machine learning models for property predictions as described in the publication. The files contain the complete machine learning pipeline, and results for a small test (toy) set.Third, we share the Supporting Files from the publication, which contain sample run scripts and the force field files to reproduce molecular dynamics simulations of stress-strain curves of the carbon nanostructures up to failure using IFF-R.Fourth, we share the pre-processed graphs trained on 90% of the dataset of CNT bundles, which can be used to reproduce the predictions of mechanical properties using HS-GNN.