A multiparametric and unsupervised-learning approach to characterize seismic and deformation variability during Villarrica’s eruptive cycle (2018–2019)

The Villarrica volcano in southern Chile is one of the most persistently active basaltic systems in South America, characterized by continuous open-vent degassing, sustained tremor, and episodic lava bursts. These conditions generate a complex seismic environment where traditional event-based analyses may overlook subtle changes in system behaviour. This study focuses on the period between December 2018 and September 2019, during which multiple eruptive pulses were documented by Villarica Volcano Observation Project (POVI) during the austral summer, autumn, and mid-winter, followed by a quieter interval in August and renewed activity in September. The identification and delimitation of the study period is based on long- and very long-period classifications and visual observations, but these data were not considered as analytical variables. This natural alternation between eruptive and calm phases provides an ideal framework for evaluating temporal patterns in seismic and deformation signals.

Continuous broadband seismic data at 100 Hz are segmented into 3-minute windows (18,000 samples), producing thousands of high-resolution segments per day across several stations and components. From each window, several statistical and spectral features are extracted using the tsfresh package (python), creating a high-dimensional representation of signal variability. In parallel, an eight-station GNSS network (2012–2024) provides deformation context to interpret the analysed interval within Villarrica’s broader inflation–deflation behaviour.

Unsupervised learning methods are applied to the feature space to identify latent patterns without imposing predefined classes. Preliminary results indicate that feature-based representation captures clear differences between eruptive and quiescent intervals, suggesting that changes in the seismic signal statistical structure may reflect shifts in fluid dynamics and conduit conditions. The method also reveals intermediate states that do not coincide directly with eruptive pulses, pointing to possible transitions in the underlying system.

This work presents an integrative framework linking high-frequency seismic variability, eruptive observations, and GNSS-derived deformation. The results highlight the potential of unsupervised learning to identify transitions in volcanic behaviour and to support future multiparametric monitoring strategies at Villarrica and similar open-vent systems.