From Short Histories to Long Futures: Horizon-Aware Graph Neural Networks for Long Horizon Forecasting

2026-05-28 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Environmental Science & Earth Systems · Depth: Expert, quick

Summary

A multi-horizon Graph Neural Network (GNN) emulator is proposed for accurate long-range prediction of geophysical systems, addressing challenges like non-linear dynamics and error accumulation in traditional methods. Unlike most deep neural networks that drift over long forecast horizons, this GNN learns state-to-state transitions from a single current time to multiple future lead times within one unified model. It represents the physical domain as a graph, with nodes for spatial locations and edges for local interactions, predicting ice thickness and ice velocities using a shared GNN backbone and separate output branches. Stability is enhanced by predicting state increments relative to the current state. The model is trained by jointly optimizing all lead times with a unified regression objective, and inference employs a coarse-to-fine rollout strategy. Experiments on multi-decadal Pine Island Glacier simulations demonstrate superior long-range accuracy and stability compared to initial-state baselines and standard single-step autoregressive rollouts, making it a more reliable emulator for climate and sea-level studies.

Key takeaway

For research scientists developing long-horizon geophysical emulators, this GNN approach offers a robust solution to overcome instability and error accumulation. You should consider implementing multi-horizon training with state increment prediction and a coarse-to-fine rollout strategy. This can significantly improve the accuracy and stability of your models for multi-decadal climate and sea-level studies, providing more reliable long-term forecasts.

Key insights

A multi-horizon GNN emulator improves long-range geophysical forecasting by learning state transitions and predicting increments.

Principles

• Long-range forecasting benefits from multi-horizon training.
• Predicting state increments enhances model stability.
• Graph representations capture spatial interactions effectively.

Method

The method uses a GNN with a shared backbone and separate output branches to predict future states. It trains jointly across lead times, predicting state increments, and employs a coarse-to-fine rollout for inference.

In practice

• Apply GNNs for multi-decadal climate predictions.
• Use increment prediction for stable long-term models.
• Implement coarse-to-fine inference for efficiency.

Topics

• Graph Neural Networks
• Long-Horizon Forecasting
• Geophysical Systems
• Climate Modeling
• Pine Island Glacier
• Multi-horizon Prediction

Best for: AI Scientist, Research Scientist

Related on AIssential

• Utilizing AMD Instinct GPU Accelerators for Weather and Precipitation Forecasting with NeuralGCM — NeuralGCM combines GCMs with ML for stable, accurate, and physically consistent weather and precipitation forecasts.
• Physics-Conditioned Synthesis of Internal Ice-Layer Thickness for Incomplete Layer Traces — A novel network synthesizes complete ice-layer thickness from incomplete radar data by integrating geometric and transformer learning with physical climate model conditioning.
• BALLAST: Bayesian Active Learning with Look-ahead Amendment for Sea-drifter Trajectories under Spatio-Temporal Vector Fields — BALLAST optimizes Lagrangian observer placement by simulating future trajectories using sampled vector fields.
• Time-TK: A Multi-Offset Temporal Interaction Framework Combining Transformer and Kolmogorov-Arnold Networks for Time Series Forecasting — Multi-offset temporal embedding combined with KANs and Transformers significantly improves time series forecasting accuracy.
• How Far Can You Actually Forecast? The Lyapunov Ceiling — Forecast quality in time series models decays predictably with increasing prediction horizon.
• Graph In-Context Operator Networks for Generalizable Spatiotemporal Prediction — In-context operator learning with diverse operators significantly outperforms classical methods for complex spatiotemporal prediction.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.