Sunstone is pleased to introduce a publicly-available prototype of the Dominican Republic Global Earthquake Model (GEM) for the Flamingo platform of OASIS.

Sunstone initiated a partnership with GEM in early 2018 in order to facilitate the use of GEM’s models for (re)insurance purposes (You can read more about the partnership between GEM and Sunstone here).

Sunstone values OASIS as a robust and open model platform in which to host the GEM models and at the same time serve the financial modelling needs of the industry.

We are pleased to invite you to try this prototype and share your thoughts and questions, not only about the prototype but about any GEM models you would like to see in the OASIS platform in the future.

About the model

This earthquake risk model is an implementation of GEM’s OpenQuake model running on the Oasis modelling platform Flamingo. The OpenQuake model for the Dominican Republic is a product of the Caribbean and Central America earthquake Risk Assessment (CCARA) project.

Specifications

• Length of simulation: 5000 years.
• Number of events: 5619.
• Geocoding support: lat-lon, municipality, province.
• Hazard grid resolution: 4-level nested population-weighted variable resolution grid, with 6112 cells of size ranging from 0.03 deg (~3 km) to 0.004 deg (~400 m).
• Hazard intensity: peak ground acceleration and spectral acceleration at three periods (0.3, 0.6 and 1.0 sec) with full uncertainty.
• Vulnerability assignment: dependent only on GEM building taxonomy (51 types).
• Coverages: structural (building).

Connecting OpenQuake and Oasis

Here we introduce the approach used to bridge the two platforms and highlight some of the fundamental differences in the model architectures:

Resolution

OpenQuake and Oasis differ fundamentally in their methodology in one important area. Whereas OpenQuake can calculate on-the-fly the ground motion at any location using the ground motion prediction equations, Oasis requires that all hazard intensities are pre-calculated. This means that the geographical resolution of the OpenQuake model is limited only by the resolution of the component with the lowest resolution, which is the Vs30 soil data that is on a native resolution of 30 arc-seconds (< 1 km). Ideally, we would like to pre-calculate the Oasis model hazard also at this native resolution, but this would lead to a very large amount of data needed to be saved with the model for larger regions. Therefore, we have optimised the grid using population density, so that the maximum resolution compatible with the Vs30 soil data is achieved in population centres and coarser resolution is used in rural are.s. This provides a reasonable balance between precision and data size.

The native resolution of the hazard model can be viewed through the model’s soil map (Fig. 1), calculated using grid-box averages of the 1 km USGS Vs30 data. The value of the variable resolution grid is seen in the greater detail visible in the capital Santo Domingo on the central south coast.

Fig. 1: Vs30 soil data used in the model, calculated as grid-box averages of the USGS global data.

Hazard Maps

The seismology of the region is illustrated well in the 500-year hazard maps for the four intensity measures (Fig. 2), calculated from the 5000-year event set.

Fig. 2: The 500-year return level of ground motion for the four intensity measures in the Oasis model.

Uncertainty

The OpenQuake model employs the concept of a logic tree to address uncertainty in the hazard. This means that during the hazard modelling, there are several possible scientific decisions to be made regarding the modelling of fault sources or ground motion prediction equations, giving rise to different branches of the tree. The OpenQuake model quantifies the likelihood of each branch and employs a probability-weighted sampling methodology during the simulation of events. This provides a robust way of statistically capturing uncertainty.

Oasis has no knowledge of logic trees, but still provides a way to capture the uncertainty by optionally providing the probability distribution of hazard for each location and event. We have run each branch of the logic tree and used the OpenQuake probabilities to define the distribution of shaking at each location for each event. During the Oasis model run, ground motion is randomly sampled from these distributions. While this gives an accurate representation of the uncertainty of the OpenQuake ground motion at each point, it does not capture the correlation between sites that the OpenQuake logic tree approach provides.

Case Study – the GEM Exposure Database

Here we present a comparison of the modelling of the GEM Exposure Database in OpenQuake and Oasis.

As a first check of the Oasis model, we present the per-event loss comparison between Oasis and OpenQuake (Fig. 3), with OpenQuake configured to use the same gridded event set that was used to create the Oasis model. In both models, the event loss presented is the probability-weighted mean of the samples. All events match within 1%.

Fig 3. Event-by-event loss comparison between OpenQuake and Oasis, with OpenQuake configured to use the same gridded event set as Oasis.

The validation of the Oasis model against the OpenQuake model run in its standard configuration (calculating hazard at each site using local site conditions rather than using a pre-calculated grid) is best validated using the OEP curve (Fig. 4). As can be seen, the move to a gridded hazard introduces a small difference in the loss exceedance probabilities.

Fig 4. Exceedance probability of ground-up loss for Dominican Republic OpenQuake and Sunstone-Oasis models run with GEM economic exposure, including Oasis model with exposure geocoded at lat-lon (dark blue), Oasis model with exposure geocoded at municipality (light blue), OpenQuake using hazard calculated at the location of the exposure (brown; 386 unique locations), and OpenQuake with hazard calculated on the same grid as the Oasis model (pink).

How to access the model

The Sunstone Flamingo installation has been decommissioned. You are free to download and access the model using the oasislmf python library. Instructions can be found on the model’s GitHub page:

How to use the model

Please refer to the README file in the GitHub repository, or otherwise refer to the oasislmf package documentation.

Sample exposure files can be downloaded here.

Exposure coding

• • Only the building taxonomy is used to lookup the vulnerability function.
  • The model supports 51 unique typologies. A table of these along with their construction characteristics can be downloaded here. Any risks with a typology not in this list will be neglected.
  • More information about the GEM taxonomy, including a tool for creating these, can be found here.
  • Only the structural (buildings) coverage is modelled, for which the TIV should be entered under the EQCV1VAL column.

---