Asymmetric Buckling Columns (ABC)

Link to the manuscript "Learning Mechanically Driven Emergent Behavior with Message Passing Neural Networks" is forthcoming. All code necessary to generate the ABC dataset is available on GitHub (https://github.com/pprachas/ABC_dataset). For questions, please contact Emma Lejeune (elejeune@bu.edu).The Asymmetric Buckling Columns (ABC) dataset contains spatially heterogeneous columns with fixed-fixed boundary conditions that are classified to be buckling left (label of 0) or right (label of 1). The dataset is split into 3 subdatasets: sub-dataset 1, sub-dataset 2, and sub-dataset 3, each with increasing levels of geometric complexity. For each sub-dataset, we provide information to reconstruct the domain geometry as txt files (subdataset*_geometry.zip), graphs from Simple Linear Iterative Clustering (SLIC) segmentation with varying degrees of node density as json files (subdatset*_sparse_graphs.zip, subdatset*_medium_graphs.zip, subdatset*_dense_graphs.zip) , and output labels as txt files (subdataset*_labels.zip). In brief, sub-dataset 1 is generated by stacking blocks with varying widths, sub-dataset 2 consists of overlapping rings of identical size, and sub-dataset 3 consists of overlapping and trimmed rings of varying sizes. For sub-dataset 1 geometry files, "x.txt'' indicates the centers of each block and "l.txt'' gives the length of each block. Each block is stacked top to bottom. For sub-dataset2, the files in folder "x'' and folder ``y'' give the "x'' an "y'' coordinates for each ring with inner radius of 0.15w and outer radius of 0.25w. Note that there is a different number of rings in each structure. For sub-dataset 3, "x.txt'' and "y.txt'' contain the "x'' and "y'' coordinate of each ring, and "outer.txt'' and "inner.txt'' give the ring outer thickness and the ratio of inner thickness to outer thickness respectively. Note that the coordinates for all domains are defined such that the origin is in the top left. Details on how to generate ground truth domain geometry with the provided geometric information and the corresponding graphs and finite element meshes are provided on GitHub (https://github.com/pprachas/ABC_dataset). The json graph files are the graphs used in the corresponding manuscript, and the code to load the json files with Pytorch and Pytorch Geometric is also provided. All subdatasets contain 25,000 simulation results. For the manuscript, 20,000 data points are used to train the ML model with 2,500 datapoints used for validation, and 2,500 datapoints held out as test data. The graphs provided are first shuffled and then split into train, validation and test data. Details of our protocol for shuffling and splitting the data are also provided on GitHub