Explainable AI for Deep Learning Based Seismo-Acoustic Applications (MSO Project)

Characterisation of seismo-acoustic sources with Deep Learning

Artificial Intelligence (AI) and Deep Learning (DL) are increasingly gaining consensus as powerful methods to support geosciences and help unveil novel scientific discoveries. However, their applications in research domains, such as Seismology, are still in the early stages.

At KNMI we have investigated DL methods, such as Convolutional Neural Networks (CNN), in order to classify continuous seismic waveforms recorded in the Groningen region into three categories: earthquake, noise and other seismo-acoustic event-type (e.g., ordnance explosion and sonic boom) (Trani et al., 2020).

DL is often perceived as a black box for the lack of transparency of its processes, the complex non-linear models and the many parameters involved. A major challenge for the adoption of an operational DL system remains the building of understanding, trust and confidence in such methods and their results. In other words they need to be made “explainable” to users e.g., seismologists, operators and analysts.

This project aims to investigate requirements that must be fulfilled in order to cover the last mile towards the operational adoption of an AI-based seismo-acoustic decision making system. Building on Explainable AI (XAI) we will research effective methods to make AI and DL trustworthy and usable, for instance, by relating expert knowledge with learned features and predictions.