Perilyx's core is a set of GPU-accelerated neural PDE solvers trained to approximate the physical propagation of climate hazards through terrain, vegetation, hydrology, and built infrastructure — then to model how each hazard changes the conditions for the next one.
Continuous ingestion of satellite soil moisture and canopy density, utility grid telemetry, USGS stream gauges, and NOAA ensemble forecasts, resolved to parcel-level grids.
Neural PDE solvers — trained against physical hazard-propagation equations, not just historical loss curves — model heat transfer, fuel combustion, saturation, and debris-flow dynamics.
A persistent per-parcel state vector tracks how each hazard event shifts underlying conditions, so cascade probability compounds correctly instead of resetting to a historical base rate.
Actuarial loss tables are excellent at pricing events that look like the past. They structurally cannot anticipate a compound sequence — say, three consecutive wet-then-dry seasons priming a watershed for post-fire debris flow — because that exact sequence may never have appeared in the training window for that geography.
| Capability | Historical loss models | Perilyx cascade engine |
|---|---|---|
| Prices known historical patterns | Yes | Yes |
| Models novel hazard sequences | Limited | Yes |
| Tracks parcel-level physical state over time | No | Yes |
| Forward lead time on compound events | Days–weeks | Months–18mo |
| Portfolio correlation from shared cascade exposure | No | Yes |
Cascade scores, lead-time windows, and confidence intervals are delivered by REST API or scheduled batch export, with pre-built connectors for common underwriting and GIS platforms.
Query cascade probability, expected severity, and forward lead time for any parcel or portfolio, refreshed nightly.
Book-level exposure clustering by shared cascade pathway, delivered ahead of renewal cycles.
Our science team can walk your team through a live simulation on a region of your choosing.