Scalable Uncertainty Quantification for Extreme Weather Forecasting via Empirical Neural Tangent Kernels

2026-06-01 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Environmental Science & Earth Systems, Mathematics & Computational Sciences · Depth: Expert, quick

Summary

Neural Tangent Kernel-based uncertainty quantification (NTK-UQ) addresses the critical gap in deep learning weather models, which currently provide deterministic forecasts without crucial uncertainty estimates for extreme weather events. This scalable method utilizes last-layer empirical features. Theoretical analysis reveals UQ quality is architecture-dependent, influenced by a variance collapse mechanism where aggressive eigenvalue truncation (e.g., k \u2264 10 for spectral operators) can destroy discrimination, while attention-based models tolerate full-rank computation. Decomposition performance benefits from Independent Component Analysis (ICA) for heavy-tailed extreme-event features, leveraging higher-order statistics, outperforming Singular Value Decomposition (SVD). A data-driven rule selects between ICA and SVD based on the feature eigenspectrum concentration ratio. NTK-UQ achieves 31-37% sharper prediction intervals at 90% coverage compared to split conformal prediction and uniquely generates adaptive intervals scaling with event severity. It requires no retraining, with inference-time uncertainty needing only a single matrix-vector product per sample.

Key takeaway

For AI Scientists and Research Scientists developing deep learning weather models, integrating Neural Tangent Kernel-based uncertainty quantification (NTK-UQ) is crucial. This method provides adaptive, sharper prediction intervals for extreme weather, a capability standard conformal prediction lacks. You should consider NTK-UQ to enhance forecast reliability and support high-stakes decision-making, especially when dealing with heavy-tailed data distributions. Implement the data-driven ICA/SVD selection rule to optimize decomposition performance.

Key insights

NTK-UQ provides scalable, architecture-aware uncertainty quantification for deep learning weather models, outperforming conformal prediction.

Principles

• UQ quality is architecture-dependent.
• Variance collapse impacts UQ discrimination.
• ICA excels for heavy-tailed extreme event features.

Method

NTK-UQ uses last-layer empirical features, applying a data-driven rule to select ICA or SVD based on feature eigenspectrum concentration for decomposition, then computes uncertainty via a single matrix-vector product.

In practice

• Apply NTK-UQ for adaptive weather forecast intervals.
• Use ICA for heavy-tailed extreme event features.
• Select decomposition based on eigenspectrum ratio.

Topics

• Uncertainty Quantification
• Extreme Weather Forecasting
• Neural Tangent Kernels
• Deep Learning Weather Models
• Independent Component Analysis
• Conformal Prediction

Best for: AI Scientist, Research Scientist

Related on AIssential

• Structure-Aware Epistemic Uncertainty Quantification for Neural Operator PDE Surrogates — Restricting uncertainty sampling to the lifting module improves neural operator uncertainty quantification efficiency and accuracy.
• The Role of Ambiguity in Error Prediction via Uncertainty Quantification — Disentangling input ambiguity from UQ signals significantly improves LLM error prediction, especially for Question Answering tasks.
• v271: Proceedings of ProbNum 2025 — Probabilistic Numerics integrates uncertainty quantification into numerical algorithms for robust computation.
• Beyond Surface Statistics: Robust Conformal Prediction for LLMs via Internal Representations — Internal LLM representations offer more robust uncertainty signals for conformal prediction, especially under distribution shifts.
• Deep Learning Surrogates for Emulating Stochastic Climate Tipping Dynamics — A dynamics-informed Temporal Fusion Transformer provides a 465x faster, differentiable surrogate for forecasting stochastic climate tipping events with high fidelity.
• Adversarial Robustness of NTK Neural Networks — Early stopping is crucial for adversarial robustness in NTK neural networks, as overfitting leads to vulnerability.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.