Sum – Product Networks for Eruption Forecasting

Volcanic eruption forecasting is a crucial aspect of hazard mitigation and public safety. This study explores the application of Sum Product Networks (SPNs) as a novel approach for forecasting volcanic eruptions. SPNs are a type of probabilistic graphical model known for several successful applications in image classification, speech recognition or natural language processing. They offer a flexible and expressive framework for probabilistic modelling and inference and can capture complex dependencies and make accurate predictions.

To investigate the potential of SPNs for volcanic eruption forecasting in New Zealand, we trained and tested a model using continuous seismic amplitude measurements (RSAM), local tectonic earthquakes rate, the number of long-period and very-long-period earthquakes and SO2, CO2 and H2S gas fluxes at Whakaari – White Island volcano between 2010 and 2020, during which time six eruptions occurred. We tested different forecasting periods ranging from 2 to 90 days before an eruption to investigate the SPN’s performance and ability to model complex relationships within the data. Our preliminary results indicate that an SPN-based approach is capable of accurately predicting an eruption more than 70% of the time within a time window of 40 days or more prior to eruption. Finally, we conducted a thorough comparison with other machine learning models to demonstrate the competitive performance of SPNs.