Slice through the predictions of our semantic segmentation network applied to a validation simulation. Left panel: Ground truth representation showing in red the voxels/particles belonging to a DM halo at z = 0 and in blue those particles that do not belong to a DM halo. Central panel: Probabilistic predictions of the semantic network with colour-coded probabilities for halo membership. Right panel: Pixel-level error map indicating true positive (green), true negative (blue), false negative (black), and false positive (red) regions resulting after applying a semantic threshold of 0.589 to our predicted map. The network effectively captures complex halo boundaries and exhibits high validation accuracy (acc = 0.86) and F₁-score (F₁ = 0.83).