This paper proposes AtmoLearn, a novel framework that learns and incorporates atmospheric attractor structure into ML weather forecasting models, improving physical consistency and robustness over extended forecast horizons.
Key findings
AtmoLearn combines delay embedding reconstruction, neural operator learning, and physics-informed regularization.
Significantly reduces error accumulation during autoregressive rollout and maintains physical consistency at extended lead times.
Theoretical analysis establishes convergence guarantees and robustness bounds for the approach.
Comprehensive experimental protocol comparing AtmoLearn against state-of-the-art baselines on standardized benchmarks.
Limitations & open questions
The framework's scalability to real-time operational deployment needs further investigation.
Potential failure modes and mitigation strategies for operational deployment require detailed risk analysis.