Fathom and Verisk Use Diffusion AI to Reshape Cat Modeling

The standard catastrophe model runs on physics: combine decades of atmospheric science, engineering research, and claims data to simulate a probability distribution of natural disaster losses. The Financial Times reports that risk modelers including Fathom and Verisk are now using AI diffusion models to bypass the limits of those physics-based 'cat' models, synthesizing disaster scenarios that extend event sets beyond what conventional approaches can reach.

Fathom, acquired by Swiss Re in December 2023, combines physics-based modeling with machine learning in targeted applications. In early 2026, according to InsureTechTrends, Swiss Re announced it was integrating Fathom's flood hazard and terrain data into its own internal catastrophe model, building 50,000-year probabilistic flood event sets that leverage AI-enhanced climate models to capture extreme scenarios and remove historical biases. Fathom's FathomDEM+ product uses machine learning to generate high-resolution terrain data that feeds into those physics-based simulations. Professor Paul Bates, Fathom's Chairman and Co-Founder, put the underlying logic plainly: 'the foundation of better flood prediction starts with better terrain data—specifically digital elevation models (DEMs)—not just more sophisticated algorithms.' Verisk is pursuing a parallel approach at the property level, using AI analysis of images from satellites and low-flying aircraft to add a data layer conventional models could not access at scale.

The urgency behind this shift is legible in the loss numbers. Insured natural catastrophe losses exceeded $137 billion in 2024, the fifth consecutive year above $100 billion. Thin historical data for rare, extreme events introduces pricing gaps that translate directly into underwriting shortfalls for carriers and reinsurers.

The caveat worth taking seriously was flagged in follow-on coverage: diffusion models carry their own failure mode. The Decoder published a response under the headline 'Insurers turn to generative AI for catastrophe modeling, but hallucinations and sales logic could get in the way.' A separate industry perspective on the modeling question makes the validation point plainly: 'open science and peer-reviewed methods should be non-negotiable criteria' when evaluating catastrophe models. What the reporting does not resolve is how regulators and rating agencies plan to treat AI-generated scenario sets relative to the physics-grounded standards they already accept.

If that question gets answered, the upside is material: smaller insurers who previously could not access high-resolution physics-based models could gain granular event sets, and the firms building them, Fathom and Verisk chief among them, can offer forward-looking scenarios that embed climate trajectories rather than extrapolating from sparse historical records.