Anvil

Multi-Model ML System · Historical & Nowcast

Named for the cumulonimbus incus — the flat, ice-crystal cap that forms when a storm's updraft punches through the troposphere — Anvil is nature's signal translated into machine intelligence.

Anvil is a multi-model ML system producing two calibrated outputs: hail probability and hail size. Both run across two modes — historical, for post-event analysis and ground-truth validation, and nowcast, for real-time early detection before hail reaches the ground.

What Anvil Predicts

Three Models, One Unified System

A multi-model prediction system producing three distinct, calibrated outputs every two minutes.

Hail Probability

Calibrated chance of hail at your exact location

Hail Size

Predicted diameter from pea to softball

Storm Trajectory

Speed, direction, and ETA

P(hail|x) = Φ(Σfᵢ(x;Θ))· d̂ = Q₀.₉₅(D | MESH, ZDR, ρ)· x̂(t+τ) = x(t) + v⃗τ + ½a⃗τ²· ETA = ‖xₜ − x̂‖ / ‖v⃗‖· MESH = ∫Ŵ(Z)·dh· ΔZ/Δh = [Z(eₖ₊₁) − Z(eₖ)] / Δh· ℒ = −Σ[y·log(p̂) + (1−y)·log(1−p̂)]· v⃗ = Δx⃗/Δt ± σᵥ· P(hail|x) = Φ(Σfᵢ(x;Θ))· d̂ = Q₀.₉₅(D | MESH, ZDR, ρ)· x̂(t+τ) = x(t) + v⃗τ + ½a⃗τ²· ETA = ‖xₜ − x̂‖ / ‖v⃗‖· MESH = ∫Ŵ(Z)·dh· ΔZ/Δh = [Z(eₖ₊₁) − Z(eₖ)] / Δh· ℒ = −Σ[y·log(p̂) + (1−y)·log(1−p̂)]· v⃗ = Δx⃗/Δt ± σᵥ·

Anvil Engine

Hail Probability

73%

calibrated confidence

Predicted Size

38mm

golf ball

ETA

12min

moving SSE at 35 mph

Agentic AI

From Prediction to Action

Anvil's outputs feed directly into an autonomous AI agent that evaluates context, makes decisions, and takes action — no human bottleneck.

Detection

Anvil outputs arrive in real time

73% probability 38mm hail 12min ETA

Decision

AI evaluates user context

2.3 mi away asset: vehicle severity: high

Action

Automated response dispatched

push alert sent SMS dispatched claim documented

From raw radar to automated response in under two minutes.

Observing Network

Continental-Scale Multi-Sensor Fusion

Anvil ingests from a massive, continental-scale observing network updated every two minutes. Radar — 159 stations fused with rain gauges, satellite, and model data into a seamless mosaic with dozens of derived products at 1km resolution. Numerical Weather Models — 80+ atmospheric variables at 3km grid resolution providing instability, shear, and thermodynamic context.

Satellite & Lightning — geostationary imagery and lightning detection networks capturing storm-top signatures and electrical activity. Surface Observations — mesonets, weather stations, and automated sensors providing ground-level context. Verified Ground Truth — years of confirmed hail reports, damage assessments, and crowd-sourced observations forming the continuous feedback loop that trains Anvil.

160+

Proprietary Features

159

Radar Stations

<2min

End-to-End Latency

Billions

Training Observations

Feature Architecture

160+ Proprietary Features Across Six Domains

Each domain encodes domain expertise as engineered inputs — purpose-built to capture predictive signals no existing operational product provides.

Radar-Derived Foundations — Raw observables transformed through proprietary engineering into high-signal inputs far beyond standard products

Microphysical Discrimination — Distinguishing hail from rain, graupel, and mixed-phase precipitation through learned signal combinations

III

3D Convective Architecture — Three-dimensional storm structure capturing updraft intensity, vertical extent, and internal organization

Environmental Context Modeling — Atmospheric state features capturing instability, wind shear, and moisture flux

Spatiotemporal Storm Tracking — Storm motion vectors, lifecycle stages, and temporal derivatives capturing rapid intensification

Continuous Ground-Truth Learning — Closed feedback loop ingesting verified reports, damage assessments, and observational networks

Validation

Rigorous, Multi-Year Validation

Every prediction Anvil makes is held to the same standard — validated against real storm events across multiple years and storm types.

Multi-year hold-out validation across thousands of labeled events

Rare event sensitivity testing on the long tail of severe hail

Reliability and sharpness analysis ensuring calibrated probabilities

Cross-storm-type generalization from supercells to derechos

Multi-scale spatial reasoning validated at street-level precision

Hail Sentinel Research Team. "Anvil: A Purpose-Built Ensemble Learning Architecture for Probabilistic Hail Detection." Hail Sentinel Technical Report, 2025.

Research paper in preparation — 2026