QUEST pretrained U-Net weights for quasiparticle-ridge extraction from synthetic ARPES (seeds 42/7/123)

Pretrained network weights of the QUEST quasiparticle-ridge extractor: a U-Net (ResNet-18 encoder, depth 3, single segmentation head, 112×112 input) trained on the QUEST-wannier_v40 synthetic ARPES corpus to predict Gaussian ridge heatmaps of quasiparticle dispersions from noisy E–k momentum–energy images. The noise/background/resolution pipeline is applied online by the training and evaluation code; the network itself is a pure image-to-image model. This record ships the Dice-optimal (reference) checkpoints of the three-seed ensemble (seeds 42/7/123; mutually independent, material-disjoint 80/10/10 splits) used in the accompanying QUEST manuscript. Per-seed test Dice0.5 = 0.7026 / 0.7114 / 0.7082 (mean 0.707±0.004) on the fixed-seed noisy replay of the held-out cuts, ridge coverage 0.86–0.87, covered-peak distance ≈40 meV. Each experiment provides the EMA state_dict (best_dice.pt, plain tensor dict, 158 keys), the full training recipe (exp_info.json), the material-grouped split (split_manifest.json), training curves (history.json), frozen-protocol test metrics (test_results.json) and the per-cut noise-replay parameters of the test protocol (noise_params.tar.gz). Provenance: trained on the QUEST-wannier_v40 corpus (Zenodo DOIs 10.5281/zenodo.23857001 and 10.5281/zenodo.23857004); QUEST code commits 66a2538 (seed 42) and dbee64f (seeds 7, 123). License & attribution: CC-BY-4.0. Please cite this record and the training corpus records. Code: see the attached README.md for the loading convention (QUEST repository, ml/models/heatmap_model.py and ml/training/checkpoint.py). Limitations: weights are EMA state_dicts (no optimizer/RNG state); only the Dice-optimal checkpoint of each run is included; inputs must follow the preprocessing recorded in exp_info.json (input_config, 112×112 resize, online noise pipeline). Access by request; fully open upon journal publication.