Transfer Learning and Benchmarking for Induced Seismic Event Detection: Insights From Oklahoma

Abstract Machine learning models for microseismicity detection are often limited by the scarcity of large and high‐quality labeled data sets in many regions. To address this need, we introduce the Oklahoma Labeled AI Dataset (OKLAD), a manually curated data set compiled by the Oklahoma Geological Survey (OGS). OKLAD is designed to support studies of induced seismicity and serves as a benchmark for evaluating deep‐learning detection models in local and regional monitoring contexts. Using OKLAD, we fine‐tuned several established phase‐picking models and observed substantial improvement in local and regional detection. The best performing model achieved recalls of 91.1% for first arrival P‐ detection and 89.8% for first arrival S‐wave detection. Validating this model on continuous data in Oklahoma, we recovered 96.8% of the OGS‐cataloged events and identified 146.8% more events after associative comparison with the events reported by routine network operations. Comparable improvements were also observed when applying the best performing models to other induced seismicity settings, such as west Texas. These results establish OKLAD as a benchmark data set for induced s