Preprint Shows ML Pulse Classification Matches Traditional Methods in Metallic Magnetic Calorimeters

Metallic magnetic calorimeters rely on paramagnetic sensors read out by SQUID arrays to measure tiny temperature rises from X-ray absorption. The Herdrich work trained autoencoders on raw pulse traces from a 32-pixel array to discover artifact classes without manual labels, then used gradient-boosted trees for regression of rise-time and amplitude features. On a held-out test set of 120,000 pulses the ML pipeline reproduced the 2.7 eV FWHM resolution at 6 keV obtained by conventional matched filtering.

Prior MMC literature, including the 2021 PTB-NIST collaboration paper in Journal of Low Temperature Physics, already flagged pulse-shape distortions from flux trapping and electromagnetic interference as the dominant barrier to scaling beyond a few hundred pixels. The new ML approach directly targets that bottleneck by replacing per-pixel template libraries with a single model retrained on streaming data. This mirrors the shift seen in cryogenic bolometer arrays for neutrinoless double-beta decay searches, where CNN-based denoising reduced analysis latency by more than an order of magnitude.

A key limitation remains the absence of an independent test dataset acquired on a different MMC geometry or sensor material; performance may degrade when sensor inductance or heat capacity changes. Cross-facility validation on shared raw-pulse repositories would strengthen claims of generalizability.

Next steps include embedding the trained models on FPGA readout electronics to enable real-time vetoing at kilopixel scale, a milestone targeted by the ECHo and HOLMES collaborations within three years.