Periodic Reporting for period 1 - DERISK (Deep lEarning foRecasting of Induced Seismicity for risK management operations)
Reporting period: 2023-07-03 to 2025-07-02
Europe has set an ambitious goal towards achieving climate neutrality, aiming to reduce to zero the net greenhouse gas emissions by 2050 under the European Green Deal. This transition to a zero-carbon economy requires not only a transformation of energy systems but also the large-scale deployment of reliable, sustainable, and low-carbon technologies. While solar and wind power dominate the current renewable landscape, they remain dependent on weather variability and extensive storage solutions. In contrast, geothermal energy offers a continuous, base-load renewable resource, harnessing the Earth’s internal heat to provide both electricity and heating. Its vast potential across many worldwide regions makes it an attractive solution in the transition to a resilient, low-carbon energy mix.
However, geothermal energy is not without risks. The process of deep drilling and fluid injection needed to exploit geothermal reservoirs can alter subsurface stress conditions and, in some cases, trigger induced seismicity. Unlike natural earthquakes, which are caused by tectonic forces, induced seismic events are directly linked to human activity and can raise concerns in densely populated or seismically sensitive areas, highlighting the seismic risk associated to this extraction. While most induced seismic events are small and not hazardous, there remains the possibility of stronger tremors that may pose safety risks and undermine public acceptance. Balancing the immense benefits of geothermal as a stable, carbon-free energy source with the careful management of seismic hazards will be central to its role in Europe’s path to a climate-neutral future.
The DERISK project aims at leveraging novel Deep-Learning (DL) techniques to tackle the challenges of monitoring seismic activity at geothermal sites (both enhanced and natural) and reduce the risks associated with operational tasks (i.e. injection and extraction of fluids). In particular, it seeks to develop next-generation software and methods for creating enhanced microseismicity catalogs (EMC) and forecasting the maximum magnitude expected in a short-time window.
The impact of this project could lead to new strategies to be embedded to the well-known and widely used decision-making scheme (i.e. Traffic Light System, TLS) for safer energy extraction, thereby promoting and helping the spread of EGS deployment in Europe.
However, geothermal energy is not without risks. The process of deep drilling and fluid injection needed to exploit geothermal reservoirs can alter subsurface stress conditions and, in some cases, trigger induced seismicity. Unlike natural earthquakes, which are caused by tectonic forces, induced seismic events are directly linked to human activity and can raise concerns in densely populated or seismically sensitive areas, highlighting the seismic risk associated to this extraction. While most induced seismic events are small and not hazardous, there remains the possibility of stronger tremors that may pose safety risks and undermine public acceptance. Balancing the immense benefits of geothermal as a stable, carbon-free energy source with the careful management of seismic hazards will be central to its role in Europe’s path to a climate-neutral future.
The DERISK project aims at leveraging novel Deep-Learning (DL) techniques to tackle the challenges of monitoring seismic activity at geothermal sites (both enhanced and natural) and reduce the risks associated with operational tasks (i.e. injection and extraction of fluids). In particular, it seeks to develop next-generation software and methods for creating enhanced microseismicity catalogs (EMC) and forecasting the maximum magnitude expected in a short-time window.
The impact of this project could lead to new strategies to be embedded to the well-known and widely used decision-making scheme (i.e. Traffic Light System, TLS) for safer energy extraction, thereby promoting and helping the spread of EGS deployment in Europe.
The DERISK project's primary goal was to develop a waveform-based deep-learning analysis framework capable of detecting, locating, and characterizing induced and natural microseismicity. To this end, we designed a system (HEIMDALL) that integrates advanced graph neural networks with hybrid CNN-Transformer architectures, resulting in a new state-of-the-art pipeline for earthquake detection, picking, and localization. The model has demonstrated clear improvements in detection accuracy and robustness to noisy signals compared to existing benchmark methods. These achievements (including code and database) have been presented at international conferences, and a peer-reviewed open-access publication is currently in preparation.
On the forecasting side, we evaluated the potential of foundation models for time series, in particular Google-TimesFM, applied to the open-access seismic catalog of injection testing at the FORGE site (Utah, USA). Initial results are promising: the model performs best when magnitude prediction is coupled with related operational parameters such as injection volume, fluid density, or pressure. This suggests that while forecasting based on magnitude time-series data alone remains limited (almost random) and physics-agnostic, multi-source time series can support more reliable and “educated” predictions of maximum expected seismicity. The outcomes of this work have been presented at international conferences and discussed with peers both inside and outside academia.
On the forecasting side, we evaluated the potential of foundation models for time series, in particular Google-TimesFM, applied to the open-access seismic catalog of injection testing at the FORGE site (Utah, USA). Initial results are promising: the model performs best when magnitude prediction is coupled with related operational parameters such as injection volume, fluid density, or pressure. This suggests that while forecasting based on magnitude time-series data alone remains limited (almost random) and physics-agnostic, multi-source time series can support more reliable and “educated” predictions of maximum expected seismicity. The outcomes of this work have been presented at international conferences and discussed with peers both inside and outside academia.
The main research breakthrough and outcome beyond the state of the art, is provided by the HEIMDALL framework. Among a landscape of task-specific deep-learning models, our unified architecture simplifies training and deployment throughout the entire seismic pipeline (phase picking, phase association, event location, and magnitude estimation), as shown in Figure 1. Trained once for each seismic network, it eliminates the need to retrain multiple standalone models and comes with an integrated software stack for data preparation, training, and deployment. Due to its low inference cost, the system supports near-real-time operation. Another impact of the DERISK project is highlighting the potential of using foundation models for time series forecasting, as shown in Figure 2, especially for predicting the maximum magnitude, which shows promising potential. Operational robustness and generalization will be assessed in future phases.
To secure further uptake and sustained performance, we identify the following priorities: (i) a continual‑learning framework for both seismic detection and magnitude/time‑series forecasting, enabling periodic or streaming retraining as new data arrive and conditions evolve; (ii) robust domain‑adaptation mechanisms to manage site/network drift and preserve inference accuracy in production; and (iii) systematic evaluation across diverse networks to quantify generalisation and guide targeted updates. Together, these steps will close the loop between operations and model improvement, supporting the adaptive and reliable deployment of deliverables.
To secure further uptake and sustained performance, we identify the following priorities: (i) a continual‑learning framework for both seismic detection and magnitude/time‑series forecasting, enabling periodic or streaming retraining as new data arrive and conditions evolve; (ii) robust domain‑adaptation mechanisms to manage site/network drift and preserve inference accuracy in production; and (iii) systematic evaluation across diverse networks to quantify generalisation and guide targeted updates. Together, these steps will close the loop between operations and model improvement, supporting the adaptive and reliable deployment of deliverables.