Abstract
We demonstrate transfer learning capabilities in a machine-learned algorithmtrained for particle-flow reconstruction in high energy particle colliders.This paper presents a cross-detector fine-tuning study, where we initiallypre-train the model on a large full simulation dataset from one detectordesign, and subsequently fine-tune the model on a sample with a differentcollider and detector design. Specifically, we use the Compact Linear Colliderdetector (CLICdet) model for the initial training set, and demonstratesuccessful knowledge transfer to the CLIC-like detector (CLD) proposed for theFuture Circular Collider in electron-positron mode (FCC-ee). We show that withan order of magnitude less samples from the second dataset, we can achieve thesame performance as a costly training from scratch, across particle-level andevent-level performance metrics, including jet and missing transverse momentumresolution. Furthermore, we find that the fine-tuned model achieves comparableperformance to the traditional rule-based particle-flow approach on event-levelmetrics after training on 100,000 CLD events, whereas a model trained fromscratch requires at least 1 million CLD events to achieve similarreconstruction performance. To our knowledge, this represents the firstfull-simulation cross-detector transfer learning study for particle-flowreconstruction. These findings offer valuable insights towards building largefoundation models that can be fine-tuned across different detector designs andgeometries, helping to accelerate the development cycle for new detectors andopening the door to rapid detector design and optimization using machinelearning.