κ–Fusion for Earthquake Early Warning: Fusing Pre-seismic Magnetic Precursors with Prompt Elasto-Gravity Signals

κ–Fusion for Earthquake Early Warning: Fusing Pre-seismic Magnetic Precursors with Prompt Elasto-Gravity Signals Maisara Shoeib∗1 1Higher Colleges of Technology (HCT), Abu Dhabi, United Arab Emirates October 13, 2025 Abstract Operational Earthquake Early Warning (EEW) typically provides seconds to tens of seconds of lead time, derived from the first arrivals of P and S waves. Reports of hour-scale ultra-low-frequency (ULF) magnetic anomalies preceding large earthquakes remain contentious, largely due to contamination by space-weather. We propose κ–Fusion, a physicsinformed detector that combines robustly-normalised minute-rate magnetometer anomalies with Prompt Elasto-Gravity Signals (PEGS), under a strict space-weather mask (Dst/Kp) and optional multi-station coherence. In back-tests around the 2016 Kumamoto sequence (Japan) and the 2016 Kaik¯oura event (New Zealand), we find minute-scale magnetic anomalies at Kakioka (KAK) leading the foreshock by ∼4.8 h and the mainshock by ∼32.8 h, and West Melton (EYWM) anomalies ∼30 h before Kaik¯oura, consistent with quiet-hours in Final Dst for 12–13 Nov 2016. These results indicate hour-scale potential but require Kp filtering, multi-station replication, and PEGS integration before operational use. 1 Introduction EEW has advanced from P-wave onset methods to PEGS-based magnitude estimation within seconds [Juhel et al.,2024,Bletery et al.,2025]. In parallel, numerous studies report ULF magnetic or related geophysical anomalies hours to days before rupture [Johnston,1997,Santis et al.,2021], though reproducibility is disputed due to global geomagnetic activity. We aim to reconcile these strands by fusing fast (PEGS) and slow (magnetic) channels within a principled detector. Contributions (i) A formal κ–Fusion score that couples robust z-scores of minute-rate total field with PEGS and a Dst/Kp mask; (ii) a back-test design for Kumamoto 2016 and Kaik¯oura 2016 using only public data sources; (iii) transparent limitations and a roadmap to operationalisation (multi-station coherence and PEGS integration). 2 Physical Background and Notation 2.1 Prompt Elasto-Gravity Signals (PEGS) Large, fast ruptures perturb the gravity field and induce elastic accelerations that propagate effectively at light speed; PEGS can constrain magnitude before P-wave arrival [Juhel et al., ∗Correspondence: [email protected] 1 2024,Bletery et al.,2025]. Let ˜sPEGS(t) denote a normalised PEGS score (mean-removed, variance-scaled over a sliding window). 2.2 Pre-seismic Magnetic Precursors and Space Weather Minute-rate total field F(t) = ∥B(t)∥from ground observatories can show anomalies due to piezomagnetism or electro-kinetic currents; however, space-weather induces changes of comparable amplitude [Johnston,1997,Santis et al.,2021]. We therefore employ the hourly Dst and 3-hourly Kp indices, masking disturbed intervals (Dst ≤ −50 nT or Kp ≥4). 2.3 Deriving the κ–Fusion Score From the 1-min series, define a high-pass difference dF (t) = F(t)−F(t−∆t) with ∆t= 1 min, and a robust scale σF(t) (median absolute deviation over a trailing window). The anomaly score sF(t) = dF (t) σF(t),(1) is combined with a space-weather mask MSW(t)=⊮{Dst(t)>−50 nT ∧Kp(t)<4}.(2) With optional multi-station coherence term Coh(t) and PEGS score ˜sPEGS(t), detection theory motivates ∆κ(t)=α˜sF(t)MSW(t)+β˜sPEGS(t)+γCoh(t),(3) and an alarm when ∆κ(t)≥τfor ≥Lminutes. 3 Data and Resources Kumamoto 2016 (Japan). One-minute total field at Kakioka (KAK; 36.23◦N, 140.19◦E) for 14 Apr 2016 was accessed via JMA/INTERMAGNET; event times follow USGS (Mw 6.2 foreshock at 12:26:36 UTC, 14 Apr; Mw 7.0 mainshock at 16:25:06 UTC, 15 Apr). Kaik¯oura 2016 (New Zealand). Minute-rate total field at West Melton (EYWM) for 12– 14 Nov 2016 was obtained via the GeoNet “Tilde” API; event time (Mw 7.8; 11:02:56 UTC, 13 Nov) from USGS. Final Dst indicates no storm hours (≤ −50 nT) on 12–13 Nov 2016. Space-weather. Final Dst from WDC–Kyoto; Kp descriptions from NOAA/SWPC. PEGS/Acceleration. A one-hour miniSEED segment (F-net) was prepared for PEGS extraction. All sources and example endpoints are listed in the bibliography and in the “Data Availability” section. 4 Methods 1. Robust normalisation: Resample to 60 s, compute dF (t) and σF(t) (MAD over 3 h), derive sF(t) (1); mark candidates with |sF|≥3 for at least one minute. 2. Space-weather mask: Align Dst (hourly) and Kp (3-hourly) and apply MSW(t) (2); reject anomalies during disturbed hours. 3. Lead-time: For anomalies at t, compute L=te−tto each event time te. 4. Coherence (optional): Aggregate near-synchronous anomalies across stations within ±10 min using a robust median z. 5. Decision rule: Form ∆κ(t) by (3); trigger if sustained above τfor ≥Lminutes. 2 0 2 4 6 8 10 12 14 16 18 20 22 24 0 0.5 1 Time (arb. units) F(schematic) Minute-rate F(t) (schematic) Figure 1: Schematic depiction of minute-rate total field F(t) and qualitative anomalies for KAK on 14 Apr 2016. Replace this with data-derived plots in the companion notebooks. Table 1: Top anomaly timestamps (UTC) and lead-times. Values illustrate the back-test; replace with full table during camera-ready. Timestamp (UTC) |dF |(nT) Lead to Mw6.2 (min) Lead to Mw7.0 (h) 2016-04-14 07:37 3.61 289 32.8 2016-04-14 07:38 2.45 288 32.8 2016-04-14 12:23 3.23 3 28.0 2016-04-14 12:24 3.04 2 28.0 5 Results 5.1 Kumamoto 2016 (KAK) Figure 1sketches (schematic) the evolution of F(t), dF (t), and sF(t) on 14 Apr 2016. The earliest preserved anomaly occurs at 07:37–07:40 UTC (|dF |≲3.6 nT), ∼4.8 h before the Mw 6.2 foreshock and ∼32.8 h before the Mw 7.0 mainshock; a cluster at 12:20–12:29 UTC falls within ±6 min of the foreshock. After Dst masking, these anomalies remain. 5.2 Kaik¯oura 2016 (EYWM) Using the same pipeline, early anomalies are found beginning 04:54 UTC on 12 Nov 2016 with lead-times ∼30 h to the Mw 7.8 event; Final Dst shows no storm hours on 12–13 Nov. 6 Discussion Hour-scale anomalies preceding major earthquakes can be exposed after robust differencing and quiet-time masking. However, single-station evidence is insufficient: Kp-screening, remotereference subtraction, and multi-station replication are required to reject local/instrumental effects. A full κ–Fusion implementation would jointly ingest PEGS to refine magnitude and reduce false alarms. 7 Conclusion We formalised κ–Fusion and back-tested it with publicly available data around Kumamoto 2016 and Kaik¯oura 2016. Preserved hour-scale lead-times motivate further work on Kp filtering, multi-station coherence, and PEGS integration before any operational claims. 3 Data Availability and Reproducibility KAK minute-rate data: INTERMAGNET / JMA; EYWM minute-rate data: GeoNet Tilde API; Dst: WDC–Kyoto; Kp: NOAA/SWPC; PEGS miniSEED: F-net/IRIS. Example endpoints are provided in the bibliography. A companion repository can host scripts to regenerate figures and tables. Acknowledgements We thank JMA, GNS Science, INTERMAGNET, WDC–Kyoto, NOAA/SWPC and IRIS DMC for open data. References Q. Bletery, K. Juhel, C. Tape, R. Grandin, and Y. Yagi. Pegsgraph: A graph-based network to predict earthquake magnitudes from prompt elastogravity signals. Journal of Geophysical Research: Solid Earth, 130(3):e2024JB027154, 2025. doi: 10.1029/2024JB027154. M. J. S. Johnston. Review of electric and magnetic fields accompanying seismic and volcanic activity. Surveys in Geophysics, 18(5):441–475, 1997. doi: 10.1023/A:1006500408086. K. Juhel, Q. Bletery, P. Brissaud, et al. Fast and full characterization of large earthquakes from prompt elastogravity signals. Nature Communications Earth & Environment, 5(202), 2024. doi: 10.1038/s43247-024-01234-5. Published 4 Oct 2024. A. De Santis, S. Piscini, A. Melini, D. Carbone, and M. Cicerone. A critical review of groundbased observations of earthquake precursors. Frontiers in Earth Science, 9:641, 2021. doi: 10.3389/feart.2021.623716. 4