Tel Aviv University Turns Silent Seismic Sensors Into Nuclear Test Detectors
A new framework turns physics-based expert models into interpretable fit scores, beating random forests on seismic monitoring when training data is scarce.
- New framework turns expert physics models into interpretable goodness-of-fit features for classification with informative missing data.
- Applied to CTBT seismic monitoring, where 58% of auto-detected events are false and need expert review.
- Decomposes log-likelihood into detection, non-detection, and observed-value components, treating missingness as signal.
- Logistic regression on these features hit AUROC 0.925 versus 0.865 baseline, doubling fake-event rejection at 95% recall.
- Outperforms random forests on raw data in small-sample regimes and tolerates ±25% expert-model misspecification.
- Simulation code on GitHub; CTBTO bulletin data remains restricted to contractors.
A new preprint from Tel Aviv University and the Soreq Nuclear Research Center proposes expert-guided class-conditional goodness-of-fit scoring, which converts a mechanistic model’s fit and sensor non-detections into features for a conventional classifier. Applied to seismic monitoring for the Comprehensive Nuclear-Test-Ban Treaty (CTBT), the method nearly doubled the share of invalid candidate events rejected while retaining 95% of valid events.
Silent stations carry signal
The CTBT bans nuclear explosions and relies on a global monitoring system that includes seismic stations. Automated software associates signals from those stations into candidate events, which analysts review before adding them to operational bulletins. In the study’s data set, experts rejected more than 58% of automatically generated candidates, creating a substantial review workload.
Sparse station coverage makes automated association difficult. Signals from unrelated events can be combined into a malformed candidate, while distance, station sensitivity, background noise, event depth, and magnitude all affect whether a station reports a detection. Analysts therefore consider both the recorded signals and the stations that remained silent.
A missing detection can contradict the candidate event’s estimated properties. If a nearby, low-noise station would probably detect an event of the proposed magnitude and location, its silence lowers the candidate’s plausibility. Generic imputation can erase that evidence by replacing the missing observation without modeling why it is absent.
Compress physics into classifier features
The framework turns an expert model’s likelihood calculations into a small set of interpretable features:
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
