An exploration of contrastive self-supervised learning for reconstructed atom probe tomography data

Abstract Atom probe tomography (APT) enables spatially resolved chemical analysis at the nanometer scale, generating large 3D atomic datasets. Segmentation of subvolumes with similar composition and properties is crucial for data interpretation, but is often hindered by measurement aberrations and user bias. In this work, we explore contrastive self-supervised learning alongside a graph neural network to derive meaningful representations of 3D atomic environments in APT data. By employing clustering to group the learned representations, this approach requires no user-supplied labels, supporting an exploratory analysis. Using both experimental and artificial APT datasets, we systematically investigate the influence of input data variations on model performance. Specifically, we vary per-atom features, size of atomic environments and APT measurement aberrations, namely positional inaccuracies and limited detection efficiency. To assess capabilities, the self-supervised workflow is benchmarked against a fully supervised model. We demonstrate that, given suitably sized environments and expressive per-atom features, both approaches achieve closely matched clustering and classification performance. The supervised model is less sensitive to hyperparameters, whereas the self-supervised workflow, guided by label-free metrics, mitigates user bias.