Scaling-law-informed neural point processes for earthquake sequence forecasting
summary
The gist
Earthquake sequence forecasting requires models that can learn nonlinear history dependence while retaining robust statistical structure The paper develops a scaling-law-informed neural marked point…
In short
The Fusion model combines neural sequence learning with structured inputs from earthquake laws (ETAS for triggering and Gutenberg–Richter for magnitude) to forecast earthquake sequences. The resulting model shows superior temporal likelihood compared to ETAS and NPP, especially when considering low-to-intermediate earthquakes. However, it offers no predictive improvement regarding target event magnitude beyond the Gutenberg–Richter distribution.
Key concepts
- Fusion Framework
- This is a neural point process that learns earthquake sequences by combining catalog history with structured inputs from seismicity laws. It separates the retained catalog history from the events being predicted to better model timing and magnitude.
- ETAS-derived Triggering
- The Epidemic-Type Aftershock Sequence (ETAS) model provides information about which earthquakes are likely to trigger others. This feature is used in the temporal branch of Fusion to help predict the timing of future events based on past activity patterns.
- Gutenberg–Richter Law
- This law describes the statistical distribution of earthquake magnitudes, showing how frequently different sizes occur. In this paper, it is used to create a feature for candidate event magnitudes in the model, though it does not provide predictive power beyond this known distribution.
- Temporal Information Gain (TIGA)
- TIGA measures how much more information a model provides about the timing of an event compared to a baseline. A positive TIGA indicates that the model successfully captures useful temporal patterns in the earthquake sequence.
Terminology used across episodes
This episode discusses
The paper
Scaling-law-informed neural point processes for earthquake sequence forecasting · Read on arXiv
Kunming University of Science and Technology · Ben-Gurion University of the Negev
Transcript
Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.
Kai: We're looking at the paper: "Scaling-law-informed neural point processes for earthquake sequence forecasting".
Mira: Here's the gist of the paper. Earthquake sequence forecasting requires models that can learn nonlinear history dependence while retaining robust statistical structure The paper develops a scaling-law-informed neural marked point process called Fusion that combines neural representations of catalog history with temporal features derived from the Epidemic-Type Aftershock Sequence model and magnitude information derived from the Gutenberg–Richter law Model Development The proposed Fusion framework uses ETAS-derived triggering and Gutenberg–Richter magnitude quantities as structured prior inputs rather than strict constraints, while retaining the flexibility of neural sequence learning The framework explicitly separates the retained catalog history from the evaluated target-event set The temporal branch combines an LSTM history representation with an ETAS-derived triggering feature, whereas the magnitude branch uses a Gutenberg–Richter-informed candidate-magnitude feature without direct recurrent-history input Likelihood Evaluation Protocol The target-event likelihood is defined as the marked point-process likelihood of the process restricted to events with m ≥ Md The forecasting target threshold is fixed at Md = 3.0 for all models and evaluations The unified objective for the log-likelihood is given by Equation 14, which includes event log-intensity and magnitude log-density terms multiplied by maski, whereas the cumulative temporal intensity is accumulated across every retained interval Temporal Information Gain Analysis The event-wise temporal information gain for a model A relative to a baseline model B at time tin is defined as TIGAB n = log λM t,n − log ΛM d,r(∆r Hr) The temporal cumulative information gain of Fusion relative to a benchmark model B is TCIGFusionB(t) = X n:tin≤t TIGAFusionB n The positive-event fraction is defined as P AB + = 1/Nd X Nd n=1 I TIGAB n > 0 × 100% Performance and Interpretation The principal result is that empirical seismicity laws and neural sequence learning provide complementary information for target-event timing Fusion gives higher temporal likelihood than both ETAS and NPP when low-to-intermediate input cutoffs retain small-to-moderate earthquakes The magnitude comparisons provide no consistent evidence of predictive information beyond the Gutenberg–Richter distribution The cutoff dependence clarifies the contribution of lower-magnitude seismicity, as lower-magnitude events provide a denser description of sequence evolution between target events Benchmark Results Across the benchmark catalogs, Fusion also gives the highest mean temporal likelihood under the model-specific configurations evaluated The magnitude comparisons show that Fusion has higher magnitude log-likelihood than NPP in all five benchmark catalogs and Campotosto, whereas the ETAS/Gutenberg–Richter reference is highest in five of the six panels The cross-catalog results therefore provide no consistent evidence of predictive information about target-event magnitude beyond the Gutenberg–Richter distribution Ablation Study The comparison between Fusion and Fusion-m shows that direct recurrent-history input does not consistently improve the magnitude branch in these experiments The recurrent representation remains available to the temporal branch, where recent retained history can inform timing, while the magnitude branch is informed by the Gutenberg–Richter-derived feature without direct LSTM input Fusion maintains elevated intensity across multiple aftershock-active intervals and is less dominated by isolated peaks in the displayed configurations Overall, the results support using empirical seismicity laws as structured inputs to neural point processes for target-event timing The timing advantage depends on an informative lower-magnitude history, whereas the magnitude results show no detectable incremental information beyond the Gutenberg–Richter distribution These conclusions are based on retrospective, nonspatial forecasts and model-specific benchmark configurations The value of lower-magnitude events may depend on catalog completeness.
Paper summary: Kai: So to wrap up what we just discussed, this paper, "Scaling-law-informed neural point processes for earthquake sequence forecasting," is about developing a new framework called Fusion. The central thesis is that combining neural learning of history with structured inputs from the ETAS model and Gutenberg–Richter law gives better earthquake sequence forecasts than using either component alone.
Mira: They claim that this works because the model explicitly separates the retained catalog history from the target-event set, which allows them to test how keeping lower magnitude earthquakes affects timing predictions for a fixed target set. The goal is to see if these empirical scaling laws can complement neural sequence learning for better forecasting.
Lev: From an implementation standpoint, what’s interesting is that they use ETAS derived triggering and Gutenberg–Richter magnitude quantities as structured priors instead of strict constraints on the neural network. This means the model gets flexibility from the neural part while still being guided by known physical processes in the input data.
Kai: They evaluated this using a marked point process likelihood where they set their target-event threshold at three point zero for all models, and they look at event log intensity and magnitude log density terms multiplied by a mask i, all accumulated over every retained interval. This is how they define the unified objective for the log likelihood.
Mira: And then they analyze temporal information gain, defining it as the difference between the target event likelihood and the density of events within a certain range over history, which tells you how much better one model is at predicting timing than another baseline.
Lev: And they quantify this with cumulative temporal information gain, which sums up that event-wise gain across time t sub n less than or equal to t, giving a measure of sustained timing gains through substantial portions of the sequence.
Conclusion: Kai: So, looking at this paper, "Scaling-law-informed neural point processes for earthquake sequence forecasting," the implication is that we don't have to rely on one type of model for seismic forecasts. Instead, blending empirical knowledge about triggering and magnitude scaling with deep learning on catalog history gives a better picture of when earthquakes are likely to happen.
Mira: It means the real power here isn't just in predicting which events will be big or small, but more in understanding the precise timing structure of an earthquake sequence. The paper shows that incorporating lower-magnitude historical events actually provides a denser description of how the sequence evolves between major events.
Lev: For us, this suggests we should focus on building systems where we can effectively feed structured physics—like ETAS or Gutenberg–Richter scaling—into neural architectures for timing prediction rather than just letting the network learn everything from scratch.
Kai: The authors found that this combination gives higher temporal likelihood than both the ETAS and NPP models when you keep the lower magnitude events in your history, which is a concrete result showing that those smaller events matter for timing accuracy.
Mira: And crucially, they found that while the timing improves with lower-magnitude history, the results for predicting magnitude show no consistent predictive gain beyond what we already expect from the Gutenberg–Richter distribution. That tells us exactly where the added value lies in these forecasts.
More episodes
- 2611.14082-Two atoms with strong light-induced dipolar interactions
- 2611.14040-Characterization of BNLs 20 and 30 micrometer thick LGADs for the Electron Ion Collider
- 2611.13399-Lift-element-based dynamic network of vortex gust-airfoil interactions
- 2611.13476-Hopfion Superlattices in 3D Twisted Moir'e Light Fields
- 2611.13642-Investigation of the timing performance of double-sided p-type 3D silicon sensors produced by CNM
- 2611.13830-SiPM optical readout of a PICOSEC Micromegas precise timing detector
- 2611.14152-Where are the shortest-path nodes in spatial networks?
- 2611.14219-Sub-Ice: semi-automated mapping of basal channels from ice shelf surface elevation data
- 2611.16341-SUMNO: Spatially Unrolled Multigrid Neural Operator for Fluid Flow Surrogates
- 2611.16090-Multi-UTE Myelin Fraction Mapping: Repeatability and Cross-Field-Strength Reproducibility at 1.5T and 3T