September 17, 2020

Abhineet Gupta, Director of Resilience Labs

RE: Review of One Concern Earthquake Model


Dear Abhineet,

We are pleased to summarize our thoughts and impressions based on our review of One Concern’s model to estimate earthquake damage to buildings. Our comments are based primarily on our meeting with you and your colleagues on April 1, 2020, where we heard presentations and had discussions about One Concern’s earthquake model. As the detailed outcomes of this meeting are nicely summarized in the report you prepared (Report TWG-2020/01, Technical Working Group, Seismic USA), in this letter we will just focus on high points from our review.


As we all know, the development of models to estimate earthquake damage to buildings and its consequences is extremely challenging, due to the limited data from past earthquakes upon which to calibrate and validate these models. Therefore, earthquake damage and loss models are traditionally developed based on a combination of theory and judgment, informed by available data from past earthquakes and detailed analyses of building performance, which is supported by laboratory tests of building materials and components. In contrast to other existing software for regional damage and loss analyses, such as HAZUS, One Concern’s cloud-based platform offers a promising new approach that leverages statistical machine learning and artificial intelligence (ML/AI) techniques to integrate observed and/or simulated data to characterize earthquake ground shaking, detailed building inventories, and building performance. While the ML/AI techniques have great potential, it is important to carefully train and validate the models so as to understand their capabilities and limitations.


Following up on suggestions that we previously raised, the presentations and discussion at the April 2020 review meeting focused largely on the topic of training and validation of One Concern’s earthquake models. This included comparative studies of damage estimates between


One Concern’s model with (1) observed data from the 1994 Northridge Earthquake, (2) simulated data from the 2019 Haywired earthquake scenario study sponsored by the USGS, and (3) observed data from the 2016 Kuramoto City earthquake in Japan. In each case, we reviewed the overall results of the analyses (e.g., comparisons between estimated building damage) along with the constituent exposure data (i.e., building inventories) and earthquake ground motions. The review also considered the spatial distribution and correlations of damage, sensitivity to ground motion intensity, and the characterization and influence of geotechnical site effects. In the case of the Japan study, even though the types of buildings in Kumamoto City are different from those in the United States, the study is of special interest because the large, well-documented, building-damage data set permits further validation of One Concern’s modeling approach. The study also augments One Concern’s emerging market in Japan.


The following is a summary of the key observations from our review:

  • Overall, the generally good agreement between the damage and loss results from the One Concern model and the observed or independently calculated results of the Northridge, Haywired and Kumamoto City studies is encouraging. Where there were systematic differences in the results, they could be traced to differences in the exposure databases or other factors (e.g., localized ground motions, construction practices, etc.). The detailed report of our review identifies some of the observed discrepancies along with suggestions on how to better identify the underlying reasons for the differences and/or to improve the One Concern model. From our discussions and the follow-up report, we are reassured that the One Concern team appreciates this feedback and is receptive to following up on our suggestions.


  • To the extent that One Concern’s ML/AI approach relies on robust (1) input data on building inventories (exposure) and ground motions (hazard), and (2) training data to relate exposure/hazard to observed damage, we continue to encourage One Concern’s efforts to expand and improve these data. Building exposure data can be improved by leveraging emerging ML/AI technologies to interrogate visual/LIDAR imagery (street view, drone, and satellite imagery) along with traditional (e.g., municipal and public tax databases) and non-traditional (e.g., real estate, Microsoft Building Footprint, etc.) databases. Similarly, resolution in geotechnical information and resulting geographic variations in building performance can be improved by the application of these technologies to large public databases both in the US and abroad. We encourage the application of One Concern’s model to these data sets and comparison of resulting predictions with appropriate examples as previously published (e.g. by the Global Earthquake Model Foundation) to further validate the model. With regards to training data, we encourage One Concern to continue to collect, evaluate, and incorporate legacy data on damage and losses from past earthquakes and to engage with other groups involved with collecting data from future earthquakes (e.g., the Earthquake Engineering Research Institute, the Structural Extreme Event Reconnaissance network, and others). We encourage the continued development of comparative methods to document the contributions of ML/AI technologies. In addition, to augment limited observational data on earthquake damage, One Concern should continue to make use of synthetic data from detailed performance-based earthquake engineering analyses (e.g., FEMA P-58 analyses), while being cognizant of possible modeling biases in the simulated data sets.


  • Comparisons of the type we reviewed in the April meeting are essential to validate and promote trust in One Concern’s earthquake model, and we strongly encourage continuing efforts in this regard. Further, we encourage One Concern to engage more with the risk analysis research community in natural hazards engineering, including publishing some aspects of its work in scholarly journals, which we believe will lead to improvements in One Concern’s models and credibility among experts in the field.


Finally, beyond efforts that we reviewed to validate One Concern’s model, we think it is important to demonstrate the unique capabilities of One Concern’s technology that is not available in other damage and loss assessment methods. Certainly, cloud-based computing and information technologies offer computing speed, immediate access, and high-resolution output that are exemplary. Aside from convenience, these capabilities offer an interactive decision support system for investment in disaster risk mitigation alternatives. Ideally, the models could extend beyond the analysis of physical damage and loss to existing building inventories to a broader consideration of socio-economic aspects that can inform the design of effective risk mitigation strategies to safeguard communities.


We appreciate this opportunity to review and provide input on One Concern’s earthquake model. We also acknowledge the cooperation and responsiveness of your team to presenting information and responding to questions during our review.


Yours truly,


Earthquake model review sig

About One Concern

One Concern is a Resilience-as-a-Service solution that brings disaster science together with machine learning for better decision making. With operations in the US and Japan, the company quantifies resilience from catastrophic perils, empowering leaders to measure, mitigate, and monetize risk so disasters aren’t so disastrous.