Data-Efficient Unsupervised Recalibration of Calorimeter Sensor Arrays Using Wasserstein Adversarial Learning.

Large distributed sensor arrays require repeated recalibration as radiation damage, material aging, gain variation, and readout drift alter channel responses. We studied a high-granularity calorimeter as a large sensor array and addressed unsupervised recalibration from two unpaired datasets: a nominal reference response and an aged response with attenuated cell-wise signals. Aging was modeled by a deterministic sensor-wise base field with reading-level stochastic variation; the base coefficients were used only for post-training evaluation. We evaluated a Wasserstein adversarial calibration field against an evaluation-only global-mean coefficient predictor and two non-adversarial estimators, a per-cell mean-energy ratio and independent per-cell Wasserstein matching. The adversarial objective supplied an adaptive event-level discrepancy over the full sensor array. In a hierarchical data-efficiency study with four reference/aged sampling-seed combinations and three configured-seed adversarial fits per seed combination, the adversarial method achieved an RMSE from 0.0238±0.0014 to 0.0173±0.0012 and a positive R2 from 0.828±0.022 to 0.910±0.013 across the tested event counts. Its RMSE was also below the approximately 0.0579 global-mean reference at every event count, demonstrating the recovery of cell-wise coefficient variation beyond the global mean. The mean-energy-ratio and Wasserstein-only estimators remained below the R2=0 reference in this sparse benchmark.