Скачать или смотреть Synthetik | Advanced Flood Modelling | Austin, TX

Synthetik's state-the-art flood modeling platform:

1) Improves and expands the data associated with insured properties to better understand the location, type, and vulnerability of the NFIP portfolio

2) Enhances the quality, resolution, and speed of temporal and spatial flood models leveraging Physics-informed Neural Networks, and;

3) Quantifies flood impact by leveraging engineering analytics to determine the type and severity of flood damage to each insured property, with greater accuracy and resolution.

Unique Selling Point
Funded and co-developed with the National Flood Insurance Program (NFIP), FEMA, and the Department of Homeland Security (DHS) in the US, Synthetik believes that the key innovation of our flood modeling platform is the true ‘end-to-end’ capability delivered by the technology. Every aspect from flood hazard, to exposure, to flood damage/loss is addressed with tangible enhancements delivered through the application of state-of-the-art innovation.

The recent emergence of Physics-informed Neural Networks (PINN) (2020) has not yet been applied in a commercial application, and this provides Synthetik with a significant differentiator in this area. Furthermore, PINN studies conducted by academia have typically leveraged only 2D data, whereas we intend to also ingest detailed 3D datasets.

The proposed technical approach has already been successfully applied as part of our insurance analytics consultancy work for global insurance clients, focusing on a range of perils including terrorism, hailstorm, earthquake, and wildfire. Consequently, we are extremely confident that the underpinning methodology is sound and that it can deliver real benefit to the insurance market.

Physics-informed Flood Hazard Data Model
The ubiquitous availability of accurate high-resolution flood data is a critical and yet unresolved problem. Existing methods to obtain high-resolution data from relatively disparate and finite data measurements (e.g., flood gages) can be categorized into two major groups: 1) dynamic, and 2) statistical, each of them facing its own challenges. In response, we propose a novel physics informed neural network (PINN) based approach for high-resolution flood data analysis. Supported by physics-based modelling and measurement data, this approach will leverage 3D geometric data (previous methods have only leveraged 2D data) to achieve accurate flood data estimation at an ultra-high spatial resolution. By ingesting both the physics-based forcing fields and the geospatial information (e.g., ground topology and building data), the proposed approach will preserve not only the relevant physics represented in a typical dynamic downscaling model, but also the resolution of the local geometry/topology, which makes the whole network both accurate at very fine resolutions and generalizable at large spatial scales.

The framework is location and data agnostic and can be extended to any geographic area, variable, measurement, and application use-case. Furthermore, as the proposed framework incorporates hard-coded physical constraints, there is an opportunity to leverage state-of-the-art uncertainty estimation approaches such as Monte Carlo (MC) Dropout without sacrificing the physical consistency of our results.