One neuron is more informative than a deep neural network for aftershock pattern forecasting

29 August 2018: “Artificial intelligence nails predictions of earthquake aftershocks”. This Nature News headline is based on the results of DeVries et al. (2018) who forecasted the spatial distribution of aftershocks using Deep Learning (DL) and static stress feature engineering. Using receiver operating characteristic (ROC) curves and the area under the curve (AUC) metric, the authors found that a deep neural network (DNN) yields AUC = 0.85 compared to AUC = 0.58 for classical Coulomb stress. They further showed that this result was physically interpretable, with various stress metrics (e.g. sum of absolute stress components, maximum shear stress, von Mises yield criterion) explaining most of the DNN result. We here clarify that AUC c. 0.85 had already been obtained using ROC curves for the same scalar metrics and by the same authors in 2017. This suggests that DL - in fact - does not improve prediction compared to simpler baseline models. We reformulate the 2017 results in probabilistic terms using logistic regression (i.e., one neural network node) and obtain AUC = 0.85 using 2 free parameters versus the 13,451 parameters used by DeVries et al. (2018). We further show that measured distance and mainshock average slip can be used instead of stress, yielding an improved AUC = 0.86, again with a simple logistic regression. This demonstrates that the proposed DNN so far does not provide any new insight (predictive or inferential) in this domain.

💡 Research Summary

The paper revisits the claim made by DeVries et al. (2018) that a deep neural network (DNN) dramatically improves the forecasting of aftershock spatial patterns compared with traditional Coulomb stress models. DeVries et al. trained a six‑layer DNN with 13 451 parameters on twelve static stress features derived from the full stress tensor and reported an area‑under‑the‑receiver‑operating‑characteristic curve (AUC) of 0.85, a substantial increase over the AUC of 0.58 obtained with a simple Coulomb stress metric. However, the same authors had already achieved an AUC of roughly 0.85 in 2017 by applying ROC analysis directly to the same scalar stress metrics, suggesting that the deep‑learning approach did not actually add predictive power.

To clarify this issue, the present study reformulates the 2017 results in a probabilistic framework using logistic regression, which can be interpreted as a neural network with a single output node (one neuron). The logistic model incorporates three scalar stress descriptors—sum of absolute stress components, maximum shear stress, and the von Mises yield criterion—and learns only two free parameters (a weight vector and an intercept). Despite this extreme reduction in model complexity, the logistic regression reproduces the DNN’s AUC of 0.85. This demonstrates that the DNN’s performance can be fully explained by a simple linear combination of the same stress features, and that the massive parameter count of the DNN offers no tangible advantage in terms of predictive accuracy.

The authors then explore whether alternative, physically motivated predictors could further improve performance. They replace the stress‑based inputs with two readily observable quantities: the Euclidean distance from a candidate location to the main‑shock rupture plane, and the average slip (or average displacement) of the main shock. Both variables have clear geophysical interpretations—distance captures the well‑known decay of aftershock density with distance, while slip reflects the amount of energy released and thus the potential for triggering secondary failures. When these two variables are fed into the same logistic regression framework, the model attains an AUC of 0.86, marginally surpassing the DNN’s result.

Beyond raw performance, the study emphasizes interpretability and parsimony. In the logistic model, each coefficient directly quantifies the contribution of its associated predictor, allowing seismologists to assess the relative importance of distance versus slip, or of the various stress components. In contrast, the DNN’s thousands of weights are entangled in a non‑transparent manner, making it difficult to extract physically meaningful insights. Moreover, the low‑parameter logistic model is less prone to overfitting, requires far fewer training examples, and can be trained quickly on modest hardware—practical advantages that are especially valuable in seismology, where high‑quality labeled datasets are limited.

The broader implication of the work is a cautionary note against the uncritical adoption of deep learning in domains where simpler statistical models already capture the essential physics. While deep learning excels in tasks with massive, high‑dimensional data (e.g., image or speech recognition), its benefits do not automatically transfer to problems like aftershock forecasting, where the underlying processes are governed by well‑understood physical laws and the available predictor set is relatively small. The authors advocate for a balanced approach: start with physically motivated, low‑dimensional features, evaluate baseline models such as logistic regression, and only resort to more complex architectures if they demonstrably improve predictive skill or uncover new physical relationships.

In summary, the paper shows that (1) the previously reported DNN performance can be matched by a one‑neuron logistic regression using the same stress features; (2) replacing stress with distance and average slip yields a modest but real improvement; (3) model simplicity, interpretability, and robustness outweigh the marginal gains offered by a heavily parameterized DNN; and (4) future research in aftershock forecasting should prioritize physically grounded, parsimonious models before turning to deep‑learning solutions.