Geometry-Independent Machine Learning Framework for Straw-Tube Timing and Position Calibration in the Mu2e Experiment

Accurate spatial and temporal calibration of straw-tube readouts is critical for achieving the Mu2e experiment's design sensitivity to charged lepton flavor violation. We present a geometry-independent machine learning framework for deriving calibration constants directly from detector response data. The model ingests digitized tracker waveforms and noised target observables, constructed by adding fixed systematic offsets and Gaussian fluctuations to the nominal ground-truth timing or positional values, to jointly learn the mapping between detector signals and the underlying calibration corrections. This formulation allows the network to isolate systematic timing and positional shifts on a per-straw basis without relying on explicit geometrical information or analytical drift models. The framework integrates with a dedicated preprocessing pipeline that structures event-level data into grouped sequences, enabling the model to exploit correlations across hits and events. The resulting calibration parameters provide a data-driven alternative to geometry-dependent calibration methods traditionally employed in high-rate tracking detectors.