Neural Networks and Deep Learning (CNNs, RNNs/LSTMs): With the rise of deep learning, more sophisticated AI models entered weather forecasting. Convolutional neural networks (CNNs), which are great at interpreting images, have been used to identify patterns in weather maps and satellite images. For example, a CNN-based approach can scan satellite imagery to detect the telltale swirl of a nascent cyclone or the comma-shaped cloud patterns of a severe storm system [5] – something that previously required a trained human eye. Recurrent neural networks (RNNs), especially Long Short-Term Memory (LSTM) networks, have been applied to capture time-series trends in weather data. These models can ingest sequences (like hourly temperature or pressure readings) and learn the temporal dependencies, helping forecast how conditions evolve over time. In practice, deep learning has been tested on problems like short-term rainfall prediction (nowcasting) and refining longer-range forecasts. The key advantage of neural nets is their ability to recognize complex patterns (spatial or temporal) that might be too subtle for humans to code explicitly.

Transformer-based and Graph Neural Network models: In the last few years, AI models based on transformer architectures (which revolutionized language AI) have been adapted for weather with stunning results. Notably, Google DeepMind’s GraphCast model and ECMWF’s Artificial Intelligence Forecasting System (AIFS) are breakthrough examples in this category. These models leverage enormous historical datasets and clever neural network designs to forecast weather with high speed and accuracy. GraphCast, for instance, uses a graph-neural-network (a type of transformer-based model that treats the weather grid like nodes in a network) to achieve medium-range forecasts (up to 10 days) in under a minute of computation [1]. Similarly, ECMWF’s AIFS – which blends AI with their Integrated Forecast System – runs alongside traditional models and has demonstrated accuracy gains (it improved tropical cyclone track forecasts by ~20% in tests) [6]. These transformer-based models can capture the relationships between many points in the atmosphere at once (much like how language models capture relationships between words in a sentence). The result is an AI that can essentially “learn” physics from data and produce a forecast much faster than solving equations the old way. GraphCast and AIFS show that modern AI models aren’t just academic experiments; they’re now competing with, and in some cases outperforming, the trusted physics-based approaches in operational settings.

The insurance sector is on the front lines of climate and weather risks. Losses from natural catastrophes have surged in recent years, with insurers paying out $108 billion for disaster claims in 2023 alone (well above the prior decade’s average of ~$89 billion) [12]. Approximately one in four property & casualty insurers now employs AI models to evaluate extreme weather threats like storms [13]. These AI-driven platforms can analyze vast historical datasets and real-time meteorological inputs to estimate the likelihood and severity of events – from hailstorms to wildfires – for specific locations or portfolios. Some insurers even offer parametric products that automatically trigger payouts when an AI-monitored weather metric (e.g. hurricane wind speed or rainfall amount) is exceeded, speeding relief to customers.