Can We Trust AI for Climate Forecast?

AI is rapidly transforming the way we understand Earth’s systems. From forecasting hurricanes to tracking air quality, machine learning is bringing speed, scale, and new levels of accuracy to problems once governed solely by physics-based models.

The most recent leap? Aurora — Microsoft’s newly published foundation model for weather and climate forecasting, featured in Nature. Aurora was trained on over 1 million hours of Earth system data and, in benchmark tests, it matched or outperformed conventional numerical weather prediction models. All while generating global forecasts in seconds — not hours.

This is a major shift. With Aurora, we now have an AI model capable of providing 10-day forecasts, tracking tropical cyclones, modeling ocean waves, and even estimating air pollution — all with stunning speed. But we also have to ask:

WHAT ARE WE TRADING OFF?

Unlike physics-based models grounded in first principles, AI models like Aurora operate as black boxes. Their predictions are powered by correlations in data, not scientific causality. Their performance depends on the quality and diversity of training datasets. And they remain largely unproven for long-term climate projections — the kind of insights that matter most for adaptation, investment, and resilience planning. This raises a key question:

Can we trust a system we don’t fully understand — especially when lives, infrastructure, and global stability are on the line?

Toward a Hybrid Future

The future isn’t AI versus science — it’s AI with science. The most promising path forward is integration: pairing machine learning with the rigor of domain expertise, transparent validation methods, and ethical safeguards. Aurora gives us a glimpse of what’s possible. Now, we need to ensure those possibilities translate into responsible, resilient climate solutions.

Let’s not confuse speed with certainty. Let’s use these tools wisely — to strengthen science, not replace it.

Aurora - a Microsoft's foundation model for the Earth System (Nature, 641, 1180-1187 (2025))

[First published in Substack "Ginci Insights" on May 28, 2025: https://gincinno.substack.com/p/can-we-trust-ai-for-climate-forecast?r=2cxt8s]