Transfer Learning from ALEPH Cuts SLD Jet Angular Error by 33 Percent via Parnassus Pretraining

{"The experiment adapts a fast truth-to-reco mapping network across two 1990s-era e+e- detectors that share Z-pole physics goals but differ in silicon, calorimetry, and archived data formats. Researchers froze the ALEPH-trained encoder, replaced the decoder head, and continued training on a newly released AI-ready SLD Monte Carlo sample. The procedure required only modest SLD statistics because low-level detector-response features transferred directly. This approach bypasses the original SLD reconstruction chain, which is no longer maintained.","Performance gains appear at both particle and jet levels. Angular resolution for charged tracks improved by a factor of 5.2 relative to a randomly initialized network trained on the same SLD data. Jet angular resolution tightened from 1.73 to 1.15 times the full-simulation reference. The largest relative gains occurred in the forward region where detector technologies diverged most. These metrics indicate that shared electromagnetic and hadronic shower physics dominate over technology-specific details when the network is properly initialized.","Legacy e+e- datasets remain scientifically valuable for precision electroweak and QCD studies yet face software obsolescence. The result demonstrates that reusable pretrained simulators can extend the lifetime of archived data without recreating vanished reconstruction pipelines. Similar transfer strategies could apply to archived DELPHI or OPAL records and, prospectively, to future circular-collider detectors that reuse LEP-era concepts. The released SLD sample lowers the barrier for external groups to test such methods.","Next steps include quantifying how much target data is required before transfer gains saturate and testing whether multi-detector pretraining further improves robustness. Experiments without access to legacy code would benefit most, but validation against independent full-simulation samples remains essential before physics analyses adopt the outputs."}