Physics-informed deep learning to forecast [Formula: see text] during hydraulic fracturing.

Area of Science:

• Geophysics
• Artificial Intelligence
• Earthquake Science

Background:

• Short-term forecasting of maximum seismic event magnitudes is vital for managing induced seismicity risks during fluid injection.
• Existing methods often depend on real-time injection data, which may not always be accessible.
• Developing data-driven approaches independent of real-time injection data is therefore a critical research need.

Purpose of the Study:

• To propose and evaluate two deep learning (DL) models for forecasting seismic event magnitudes using only historical seismicity patterns.
• To compare a direct DL forecasting approach with a physics-informed DL approach that forecasts seismicity rates.
• To assess the practical utility and limitations of these DL methods for induced seismicity risk mitigation.

Main Methods:

• Developed two DL models: one for direct magnitude forecasting, and another for seismicity rate forecasting integrated with physical constraints.
• Utilized two distinct data-partitioning strategies to train and test the DL models.
• Applied the models to a hydraulic fracturing monitoring dataset from western Canada.

Main Results:

• The direct DL approach accurately forecasts magnitudes based on past seismicity but exhibits a time lag, limiting its real-time application.
• The physics-informed DL approach effectively predicts seismicity rate changes but shows variability in estimating maximum magnitudes.
• Significant exceedances in forecasted magnitudes may indicate an impending runaway fault rupture event.

Conclusions:

• Deep learning models can forecast induced seismicity magnitudes using historical seismic patterns, offering an alternative to real-time data-dependent methods.
• A physics-informed approach enhances the forecasting of seismicity rates, providing valuable insights into dynamic changes.
• The proposed methods contribute to improved risk assessment and mitigation strategies for fluid-induced seismicity, with potential early warnings for fault rupture onset.