Maximum observed magnitude of earthquakes induced by fluid injection: associated factors and predictive performance across projects

Abstract

Reliable assessment of the maximum observed magnitude (M_obs) of injection induced earthquakes is important for seismic risk analysis. We analyzed 50 induced seismicity records from 39 projects using three complementary analytical frameworks, with all validation grouped by Project ID. Regression adjusted for operation type assessed lithology’s additional contribution to M_obs. Nested validation grouped by project compared a low capacity regularized multilayer perceptron (MLP) with ridge regression to assess predictive performance. Log linear models assessed associations between M_obs and individual continuous predictors measured before the event, adjusting for fracturing depth. Operation type provided comparatively stable information, whereas broad lithological categories added no stable independent information after adjustment. Among the continuous variables, cumulative injection volume before the event defining M_obs (V_pre) showed the clearest positive association. In the model containing fracturing depth and V_pre, the coefficient of log₁₀(V_pre) was 0.640 Mw per decade (95% confidence interval: 0.439 to 0.841; p < 0.001). Maximum injection pressure, maximum injection rate, and fault length showed no stable independent associations. Nested leave one project out validation using 25 complete records from 21 projects yielded an R² of 0.501, a root mean square error of 0.776 Mw, and a mean absolute error of 0.590 Mw for the MLP. Performance differences from ridge regression were uncertain, and performance was similar to the linear model containing depth and volume (R² = 0.486). Overall, operation type and cumulative injection volume provided the most consistent empirical information, with no stable additional predictive benefit from the MLP.