Guiding the Forecast: Interpretability and AI Steering in Climate Science

While AI-based weather foundation models have revolutionized predictive capabilities, their opaque nature and susceptibility to training data biases pose significant challenges for operational trust. A prominent example is Aurora, a state-of-the-art foundation model that demonstrates exceptional hurricane tracking accuracy but consistently underestimates cyclone wind speeds. Because this bias is inherited from the underlying reanalysis data, standard retraining often fails to alleviate the systematic error.

In this work, we propose a novel paradigm for bias correction by adapting AI Steering, a technique recently established for monitoring and adjusting Large Language Model (LLM) behavior, to the domain of climate science. Rather than relying on traditional post-processing or computationally expensive retraining, steering allows us to interrogate and shift the internal neural representations of Aurora without modifying the underlying weights. By identifying the latent features associated with wind speed intensity, we can shift the model’s internal state to align more closely with high-resolution observations.

To evaluate this approach, we run forecasts initialized with IFS-HRES conditions and validate our results against IBTrACS observations. Our results demonstrate that this interpretability-driven approach helps to improve systematic biases by significantly reducing wind speed errors while preserving model integrity and maintaining Aurora’s high-fidelity track accuracy. Furthermore, we show that steering enables a form of "Human-in-the-Loop" oversight, providing a transparent mechanism for meteorologists to adjust model outputs based on physical constraints and domain expertise. By bridging the gap between LLM interpretability and AI-based weather forecasting, we highlight the potential of steering to improve operational forecasts and offer a scalable, transparency-first framework for diagnosing and mitigating failure modes in complex AI-based climate and weather models.