A hybrid CNN-Transformer model predicts solar flares up to three days early.

Solar flares are intense bursts of energy from the Sun. Strong flares, specifically classes C, M, and X, can damage satellites, disrupt radio signals, and harm power grids on Earth. Predicting them early helps protect this critical technology, but forecasting is challenging. To solve this, researchers designed a hybrid system called ResNet-TransformerTransformerA type of AI designed to understand sequences of data, such as time series or sentences.. It pairs a convolutional neural network named ResNet50, which analyzes patterns across space in solar images, with a standard Transformer, which tracks changes across time. Working together, the system examines image data to estimate whether a flare will occur, determine its exact class among C, M, and X, and compute its probability across 24-hour, 36-hour, and 72-hour windows. The team tested the model against traditional machine learning tools, like support vector machines, and standalone convolutional networks. ResNet-Transformer beat every alternative across every tested metric, showing that combining spatial and temporal tools yields more dependable solar flaresolar flareAn intense burst of radiation from the Sun's surface. forecasts.

Why it matters

Strong solar flares can knock out satellites, disrupt GPS, and trigger blackouts. Current predictions are often too late or inaccurate. This model gives utilities and space agencies up to three days of warning, enough time to protect critical systems.

Questions answered