Impact of machine-learning phase picking on seismic tomography at Popocatépetl Volcano, Mexico

Machine-learning (ML) algorithms are increasingly used to build seismic catalogs. However, their transferability and impact on downstream applications remain uncertain. Here, we evaluate how ML-based phase picking and association affect local earthquake tomography at Popocatépetl Volcano, Mexico, one of the most active and hazardous volcanoes in North America. We tested multiple pipelines that combine the PhaseNet neural network—trained on publicly available datasets—with two phase associators, GaMMA and PyOcto. These pipelines were used to generate fully automatic catalogs of volcanic earthquakes. The pipelines show substantial variability in detection rates and in the spatial distribution of seismicity. This reflects how transferability depends on both the training context and the association algorithm. To overcome these limitations, we built a hybrid catalog using a systematic workflow that merged high-quality events from different pipelines, yielding the most complete dataset suitable for high-resolution seismic tomography. The resulting three-dimensional Vp, Vs, and Vp/Vs models reproduce several structures reported in previous tomographic studies and extend the resolvable volume to greater depths beneath the volcano revealing the presence of a vertically organized magmatic system composed by at least two high-Vp/Vs regions separated by a low-Vp/Vs in the firsts 18 km of the crust. We interpret these structures as two crystal-mush reservoirs connected by a volatile-bearing layer. Our results demonstrate that carefully constructed ML-based catalogs can reliably support high-resolution seismic imaging in complex volcanic settings • Machine-learning picker and associator pipelines were tested at Popocatépetl Volcano. • A hybrid catalog combining multiple pipelines improved seismicity completeness. • Synthetic tests reveal that catalog construction influences tomographic resolution. • Vp and Vp/Vs models agree with previous studies and extend resolution to greater depth. • Results show vertically organized magma reservoirs separated by volatile-saturated crust.