SCEC Award Number 25303 View PDF
Proposal Category Individual Research Project (Single Investigator / Institution)
Proposal Title Performance evaluation of conditional generative modeling for ground motion using NGA-West2 Database
Investigator(s)
Name Organization
Rie Nakata International Computer Science Institute Michael Mahoney International Computer Science Institute Pu Ren Lawrence Berkeley National Lab Nori Nakata Massachusetts Institute of Technology Maxime Lacour University of California, Berkeley
SCEC Milestones C1,2,3-1, C1-2, A1-2 SCEC Groups GM, RC, Seismology
Report Due Date 03/15/2026 Date Report Submitted 04/29/2026
Project Abstract
 Generative Artificial Intelligence (GenAI) is revolutionizing seismic hazard analysis by offering advanced tools for predicting ground motion waveforms and intensities. We have been developing the conditional generative model (CGM) framework for ground motion modeling. Focusing on the San Francisco Bay Area, two CGM-based models were developed: one for waveform generation (CGM-GM) and another for learning path effects using FAS (CGM-FAS). Both models outperformed traditional empirical non-ergodic ground motion models (GMMs) in robustness and efficiency, demonstrating improved predictions of spatial heterogeneity in ground motion. The current SCEC proposal improves the CGM-GM performance by incorporating contrastive learning and heavy tailed distributions. This work aims to transition CGM applications from demonstration to statewide implementation in California, emphasizing performance evaluation using the established NGA-West2 database and GMMs. The models will replicate ergodic NGA-West2 GMMs using their data subsets and parameters, facilitating unbiased comparisons. Key challenges include extending models to large-magnitude earthquakes, managing epistemic uncertainty, and optimizing training strategies. This study aligns with SCEC priorities by synthesizing ground motion simulations, bridging observational gaps, and validating machine learning methods for seismic applications. Anticipated outcomes include the development of GenAI-based ground motion models (CGM-NGA-West2), improved data preparation for AI, and enhanced predictive accuracy for seismic hazard analysis. These advancements will pave the way for integrating machine learning into mainstream earthquake hazard research.
Intellectual Merit This project advances SCEC objectives by evaluating conditional generative models (CGM) using the NGA-West2 database and benchmarking against established ergodic NGA-West2 GMMs. The results show that CGM can reproduce both median ground motion and associated standard deviation, as well as key scaling behavior with magnitude, distance, and site conditions. By validating CGM behavior within a well-established database, this work supports the application of CGM to non-ergodic ground motion modeling and contributes to ongoing efforts to incorporate data-driven approaches into seismic hazard analysi
Broader Impacts
The project supports SCEC’s broader goals by advancing AI-based methods for seismic hazard analysis with potential societal benefits in earthquake risk assessment. It provided training in interdisciplinary research at the intersection of seismology and machine learning, contributing to workforce development in emerging AI-driven geophysics. The collaboration across ICSI, LBNL, UC Berkeley, and MIT strengthens research networks and infrastructure. The developed framework can be extended to non-ergodic hazard modeling and large-scale simulations, supporting improved hazard characterization relevant to engineering and public safety.
Project Participants Rie Nakata (ICSI/LBNL) served as PI and led the project. Michael Mahoney (ICSI/LBNL/UC Berkeley) contributed to machine learning methodology. Maxime Lacour (ICSI/UC Berkeley) worked on training and implementation for ground motion modeling. Pu Ren (LBNL) supported development of the conditional generative modeling framework. Nori Nakata (LBNL/UC Berkeley/MIT) contributed expertise in both seismology and machine learning. The project involved collaboration across ICSI, LBNL, UC Berkeley, and MIT.
Exemplary Figure Figure 2. Median response spectra for M = 5–8 comparing CGM-ASK and CGM-BSSA with NGA-West2 empirical models (ASK14, BSSA14) and observations at RJB = 10 km and VS30 = 800 m/s. The CGM models reproduce the overall spectral shape and magnitude scaling across periods, demonstrating consistency with established ground motion models while enabling stochastic generation of variabili
Linked Publications

Add missing publication or edit citation shown. Enter the SCEC project ID to link publication.