Becky Ham at SSA wrote a very kind announcement about the award that you can read here: https://www.seismosoc.org/award-recipient/kevin-milner/

And speaking of generous collaborators, one of my dear friends Scott Callaghan decided to “make a fuss” and surprised me with a sea of my heads on sticks while I accepted the award:

Photo credit: Mike Blanpied

The paper on the ERF was recently published in BSSA, give it a read here, or see the abstract and selected figures below. Stay tuned for a formal press release from the USGS and publication of the NSHM overview paper.

K. R. Milner, A. E. Hatem, R. W. Briggs, J. A. Thompson‐Jobe, A. L. Llenos, A. J. Michael, A. M. Shumway, E. H. Field, and K. L. Haynie, “The U.S. Geological Survey 2025 Puerto Rico and U.S. Virgin Islands Time‐Independent Earthquake Rupture Forecast,” Bulletin of the Seismological Society of America, Mar. 2026, doi: 10.1785/0120250040

Abstract

We present the 2025 U.S. Geological Survey Puerto Rico and U.S. Virgin Islands (PRVI) time‐independent earthquake rupture forecast (ERF), developed for the 2025 update to the National Seismic Hazard Model (NSHM) for PRVI. The updated ERF improves upon a prior model from 2003, including an expanded fault inventory with slip‐rate estimates, updated seismicity catalogs, and refined subduction zone geometries and deformation models. It applies the fault‐system inversion methodology to solve for rates of ruptures on modeled faults, adapted from the 2023 NSHM (NSHM23) for the western United States, including the first application of the inversion to model rates on a U.S. subduction interface. Off‐fault and intraslab seismicity are constrained by observed seismicity and use updated methods developed for NSHM23. Uncertainties in model components are substantial, and the ERF represents epistemic uncertainties through a comprehensive logic tree consisting of 1.7 billion logic‐tree branches combined across all sources.

This has been my primary project for the previous 3 years. Specifically, I worked on the fault-system earthquake rate model for the western U.S. As part of this process, I recently published a paper documenting the inversion methodology that we used in Bulletin of the Seismological Society of America. Give it read here, or see the abstract and selected figures below.

Kevin R. Milner, Edward H. Field; A Comprehensive Fault‐System Inversion Approach: Methods and Application to NSHM23. Bulletin of the Seismological Society of America 2023; doi: https://doi.org/10.1785/0120230122

Abstract

We present updated inversion‐based fault‐system solutions for the 2023 update to the National Seismic Hazard Model (NSHM23), standardizing earthquake rate model calculations on crustal faults across the western United States. We build upon the inversion methodology used in the Third Uniform California Earthquake Rupture Forecast (UCERF3) to solve for time‐independent rates of earthquakes in an interconnected fault system. The updated model explicitly maps out a wide range of fault recurrence and segmentation behavior (epistemic uncertainty), more completely exploring the solution space of viable models beyond those of UCERF3. We also improve the simulated annealing implementation, greatly increasing computational efficiency (and thus inversion convergence), and introduce an adaptive constraint weight calculation algorithm that helps to mediate between competing constraints. Hazard calculations show that ingredient changes (especially fault and deformation models) are the primary driver of hazard changes between NSHM23 and UCERF3. Updates to the inversion methodology are also consequential near faults in which the slip rate in UCERF3 was poorly fit or was satisfied primarily using large multifault ruptures that are now restricted by explicit b‐value and segmentation constraints.

Figure 22 Western U.S. NSHM23 hazard maps of PGA 2% in 50 yr hazard. (a) The full mean hazard map. (b) The effect of individual fault‐based logic‐tree branches. Organized into rows for each branching level, each smaller map in panel (b) shows the ratio of hazard computed with each specific submodel (but averaged across all other branches) to the full mean hazard map. In other words, they show how the mean map would change if all other choices at that branching level were eliminated. In addition to the logic tree branches introduced in Figure 2, we include branches for deformation models (Pollitz et al., 2022) and scaling relationships (Shaw, 2023).

Kevin R. Milner, Bruce E. Shaw, Edward H. Field; Enumerating Plausible Multifault Ruptures in Complex Fault Systems with Physical Constraints. Bulletin of the Seismological Society of America 2022;; 112 (4): 1806–1824. doi: https://doi.org/10.1785/0120210322

Abstract

We propose a new model for determining the set of plausible multifault ruptures in an interconnected fault system. We improve upon the rules used in the Third Uniform California Earthquake Rupture Forecast (UCERF3) to increase connectivity and the physical consistency of ruptures. We replace UCERF3’s simple azimuth change rules with new Coulomb favorability metrics and increase the maximum jump distance to 15 km. Although the UCERF3 rules were appropriate for faults with similar rakes, the Coulomb calculations used here inherently encode preferred orientations between faults with different rakes. Our new rules are designed to be insensitive to discretization details and are generally more permissive than their UCERF3 counterparts; they allow more than twice the connectivity compared with UCERF3, yet heavily penalize long ruptures that take multiple improbable jumps. The set of all possible multifault ruptures in the California fault system is nearly infinite, but our model produces a tractable set of 326,707 ruptures (a modest 29% increase over UCERF3, despite the greatly increased connectivity). Inclusion in the rupture set does not dictate that a rupture receives a significant rate in the final model; rupture rates are subsequently determined by data constraints used in an inversion. We describe the rupture building algorithm and its components in detail and provide comparisons with ruptures generated by a physics‐based multicycle earthquake simulator. We find that greater than twice as many ruptures generated by the simulator violate the UCERF3 rules than violate our proposed model.

Figure 1 Three‐dimensional view looking north of fault model 3.1 from the Third Uniform California Earthquake Rupture Forecast (UCERF3) model broken up into 2 km × 2 km patches for Coulomb calculations. Patches overlap slightly to fill the fault surfaces completely. In this example, eight subsections of the Garlock fault are used as sources (green) with 1 m displacement and Coulomb stress changes (ΔCFF⁠⁠) are computed to all other patches (with contributions summed across all source patches). Receiver patches are colored by their sign with darker colors indicating greater amplitude, and subsection outlines are colored by the sum across all receiver patches (red is positive, blue negative). This shows the Coulomb‐preferred corupture direction of the left‐lateral Garlock fault connecting to the Mojave section of the right‐lateral San Andreas fault (SAF). Coastlines are overlaid in black.

Figure 16 Allowed fault‐to‐fault connection points in the UCERF3 and proposed models, drawn between the middles of the connected subsections. (a) Entire model region, (b) zoom in on the bay area, and (c) zoom in on southern California and the eastern California shear zone. Connections that are unique to the proposed model are drawn in red, unique to UCERF3 drawn in blue, and common to both models drawn in green. Changes in the connection strategy can result in different connection points between the same pair of faults between the two models, and the modified plausibility filters and increased jump distance result in many more connections (776) in the proposed model than were allowed in UCERF3 (369).

You can listen to those tracks directly via YouTube below. Enjoy, and please consider supporting Céleigh’s music by purchasing it. She has vinyl available on bandcamp!

Supercomputer creates over 700,000 years of simulated earthquakes

Simulating 800,000 Years of California Earthquake History to Pinpoint Risks

Kevin R. Milner, Bruce E. Shaw, Christine A. Goulet, Keith B. Richards‐Dinger, Scott Callaghan, Thomas H. Jordan, James H. Dieterich, Edward H. Field; Toward Physics‐Based Nonergodic PSHA: A Prototype Fully Deterministic Seismic Hazard Model for Southern California. Bulletin of the Seismological Society of America doi: https://doi.org/10.1785/0120200216

Abstract

We present a nonergodic framework for probabilistic seismic‐hazard analysis (PSHA) that is constructed entirely of deterministic, physical models. The use of deterministic ground‐motion simulations in PSHA calculations is not new (e.g., CyberShake), but prior studies relied on kinematic rupture generators to extend empirical earthquake rupture forecasts. Fully dynamic models, which simulate rupture nucleation and propagation of static and dynamic stresses, are still computationally intractable for the large simulation domains and many seismic cycles required to perform PSHA. Instead, we employ the Rate‐State earthquake simulator (RSQSim) to efficiently simulate hundreds of thousands of years of M≥6.5 earthquake sequences on the California fault system. RSQSim produces full slip‐time histories for each rupture, which, unlike kinematic models, emerge from frictional properties, fault geometry, and stress transfer; all intrinsic variability is deterministic. We use these slip‐time histories directly as input to a 3D wave‐propagation code within the CyberShake platform to obtain simulated Fmax=0.5 Hz ground motions. The resulting 3 s spectral acceleration ground motions closely match empirical ground‐motion model (GMM) estimates of median and variability of shaking. When computed over a range of sources and sites, the variability is similar to that of ergodic GMMs. Variability is reduced for individual pairs of sources and sites that repeatedly sample a single path, which is expected for a nonergodic model. This results in increased exceedance probabilities for certain characteristic ground motions for a source–site pair, while decreasing probabilities at the extreme tails of the ergodic GMM predictions. We present these comparisons and preliminary fully deterministic physics‐based RSQSim–CyberShake hazard curves, as well as a new technique for estimating within‐ and between‐event variability through simulation.

Figure 1 Probabilistic seismic‐hazard analysis (PSHA) pathways. This study presents a new pathway, shown with red arrows, which combines a multicycle earthquake rupture simulator directly with a ground‐motion simulator to compute synthetic seismograms. Shaw et al. (2018), shown with green arrows, combined a multicycle earthquake rupture simulator with an empirical ground‐motion model (GMM). Prior CyberShake studies, shown with blue arrows, used a kinematic rupture generator to extend an empirical earthquake rupture forecast (ERF) for ground‐motion simulation. Traditional PSHA studies, shown with gray arrows, combine an empirical ERF with empirical GMMs.

Figure 3 3D perspective view looking north of faults considered in southern California, highlighting an M 7.5 simulated Rate‐State earthquake simulator (RSQSim) rupture on the Mojave section of the San Andreas fault. Darker colors represent higher patches of total cumulative slip, and major faults and cities are annotated. All other fault patches that did not participate in the rupture are shown in gray.

Figure 15 RSQSim simulation hazard curves at USC. CyberShake (3D) is plotted with thick, black lines. (a) ASK2014 GMM comparisons curves in blue, with the complete hazard curve plotted as a thick solid line. GMM curves computed from truncated lognormal distributions at 3‐, 2‐, and 1‐σ are plotted with dashed, dotted, and dotted and dashed lines, respectively. The 1D BBP hazard curve is included in yellow, and 95% confidence bounds assuming a binomial distribution (representing sampling uncertainty from a finite catalog duration) on the 3D simulated curve are depicted as a gray shaded region. (b) An enlarged view of CyberShake hazard curves, including curves computed with different RSQSim catalog lengths. The complete catalog is shown (after discarding spin‐up time) with a thick black line, and subsets of the catalog, starting with the first 50,000 simulated yr in light gray, are shown in thin and increasingly dark lines with increasing duration.

Kevin R. Milner, Edward H. Field, William H. Savran, Morgan T. Page, Thomas H. Jordan; Operational Earthquake Forecasting during the 2019 Ridgecrest, California, Earthquake Sequence with the UCERF3‐ETAS Model. Seismological Research Letters doi: https://doi.org/10.1785/0220190294

Figure 1. Percentile scenarios: Map view of Uniform California Earthquake Rupture Forecast, Version 3–epidemic‐type aftershock sequence (UCERF3‐ETAS) synthetic catalogs of aftershocks to the Ridgecrest sequence (preferred model). Shown are 30 days of simulated seismicity immediately following the M 7.1 event. UCERF3 faults in the map region are plotted with gray outlines, except for faults that participate in events in the synthetic catalogs, which are plotted in the same color as the event hypocenter. (a) A typical catalog, defined as the catalog that lies at the 50th percentile with respect to the total number of simulated aftershocks across all 100,000 simulations. That catalog has 2584 M≥2.5 events and multiple M≥5 events, but no supraseismogenic events on UCERF3 faults. (b) A less probable catalog (97.5th percentile, 7574 M≥2.5 events) in which the Garlock fault was triggered in an M≥7 supraseismogenic aftershock. (c) The most extreme catalog from the suite of 100,000 simulations with 39,550 M≥2.5 events in which an M≥7 Garlock aftershock triggered an M≥8 earthquake on the San Andreas fault. (d) The observed earthquakes with M≥2.5 within 30 days of the M 7.1, accessed from Comprehensive Catalog (ComCat) on 6 December 2019.