The weather and climate science AI revolution isn’t revolutionary

While AI seems ubiquitous, its application in weather and climate modeling isn't a sudden leap but a refinement of established machine learning methods. Researchers have studied these techniques for years, understanding their strengths and limitations. This approach differs between weather and climate simulations.

AI in this context refers to machine learning, which involves using computers to identify patterns in data. This can range from simple linear regression to complex algorithms that uncover relationships difficult for humans to discern manually. Machine learning models learn from vast datasets, iteratively adjusting parameters to connect inputs to correct outputs.

However, limitations exist. Models can't identify what they haven't been trained on, and the quality of training data is crucial. Sometimes, the internal workings remain a "black box." Despite these challenges, machine learning algorithms often outperform human-crafted ones in computational efficiency and accuracy when used correctly.

For weather forecasting, AI models are trained on sequential weather data, running much faster than traditional physics-based models. Companies like Google, Nvidia, and Microsoft are developing these, sometimes in collaboration with academics. Major weather centers are also adopting them; for instance, the European Centre for Medium-Range Weather Forecasts (ECMWF) implemented an AI model that significantly reduces computation time and energy consumption while maintaining forecast quality.

These AI models can present challenges; for example, they might produce physically nonsensical results like negative rainfall. Addressing this involves implementing "physical guardrails" to constrain outputs and ensure consistency. The immense computational efficiency, particularly for ensemble forecasts, makes these AI models invaluable tools despite their limitations.