Molecular dynamics (MD) simulations are widely used to study tribological deformation at the atomic scale, but translating the resulting high-dimensional data into interpretable deformation pattern maps remains a largely manual and resource-intensive task. In this work, we present a data-driven workflow that aims to automate this process using unsupervised and supervised machine learning techniques. Grain-orientation-colored tomographic images from CuNi alloy simulations were first compressed into a 32-dimensional representation using an autoencoder, demonstrating that microstructural features can be retained under strong dimensionality reduction. The predictive model was trained independently using a dual-branch CNN–MLP architecture that combines image data with simulation metadata to classify dominant deformation patterns (transient and final states). The model achieves > 95% accuracy on spatially independent validation data. To assess generalization, entire simulations were excluded from training, yielding 100% final-state accuracy, as these conditions lie distant from regime boundaries. Evaluation on an artificial pressure–composition grid results in 87% accuracy, with misclassifications concentrated near regime boundaries. For transient states, the model achieves 68% accuracy on unseen simulations, increasing to > 92% when accounting for physically plausible misclassifications between adjacent regimes. A comparison with an image-only ResNet-50 model (~70% accuracy on spatially independent data, ~50% on unseen simulations) highlights the importance of integrating metadata. These results demonstrate that tribological deformation mechanisms can be automatically identified from microstructural data. This work provides a proof of concept for data-driven construction of deformation maps and represents a step toward predictive frameworks that may reduce reliance on large-scale MD simulations.