Giacomo Mastella , Federico Pignalberi, Giulio Poggiali, Marco Scuderi

Laboratory acoustic emissions (AEs) represent microslip events analogous to small- scale earthquakes, providing valuable insights into the mechanics of frictional instabilities. With technological advancements in acoustic monitoring, thousands of AE waveforms can now be collected in minutes of experimental time, requiring efficient methods for their detection and analysis. In this study, we introduce a deep learning model for automatically detecting AEs in laboratory shear experiments. Our dataset consists of about 30,000 manually picked AEs waveforms collected under different experimental boundary conditions using two fault gouge materials: Min-U-Sil quartz gouge and glass beads. By adapting the PhaseNet model, originally developed for natural earthquake phase detection, we train AEsNet, a robust AE picker that outperforms pre-existing picking methods for the tested materials. To investigate whether the trained models can generalize across different boundary conditions and materials, and overcome the limitations of small, manually labeled datasets, we apply transfer learning to analyze performance relative to training size and material diversity. Our results indicate that model performance is largely independent of experimental conditions but strongly dependent on material type. This finding suggests that direct transfer of models trained on one material to another is often ineffective due to distinct frequency characteristics of AEs, which are closely linked to the microphysical processes driving emissions in the different granular materials. However, quick fine- tuning significantly enhances pre-trained AEsNet performance, even surpassing that of a fine-tuned PhaseNet model pre-trained on natural earthquakes. This underscores the importance of customizing models to the specific attributes of laboratory-generated AEs—a conclusion consistent with findings from transfer learning applications in natural seismicity. In conclusion, our approach provides an efficient tool for enhancing AE detection, even with limited data from diverse laboratory conditions, enabling the creation of reliable AE catalogs that can significantly advance our understanding of fault mechanics in controlled experimental settings.