Abstract Forecasting is an essential part of risk mitigation, where the mitigation efficacy depends strongly on the quality of forecasts. We explore the neural temporal point process as a deep learning framework to forecast induced earthquakes. We train our deep learning model using numerous enhanced geothermal systems and hydraulic fracturing test cases. We find that our model's performance is comparable to that of a modified Epidemic Type Aftershock Sequence; the “winning” model varies, depending on the test case in question. The addition of supplementary input data (e.g., seismic moment release, cumulative volume, injection pressure, hydraulic energy) tends to reduce model performance, as compared to simply using traditional metrics (i.e., magnitudes and injection rates). Our model's architecture allows for a flexible and data‐driven inference of the inter‐event time distribution and the injection forcing function. We find that the inter‐event time distribution is compatible with an Omori‐like decay of seismicity rates. On the other hand, our model does not recover a linear proportionality between injection and seismicity rates—instead preferring a simple on/off relationship. Finally, we discuss the statistical/physical implications of these results for and suggest future improvements. Overall, this model will likely be an important part of an ensemble of forecasting approaches that constrain seismicity risks.
Schultz et al. (Sun,) studied this question.