{"id":3035,"date":"2026-02-11T08:00:14","date_gmt":"2026-02-11T08:00:14","guid":{"rendered":"https:\/\/research.reading.ac.uk\/met-darc\/?p=3035"},"modified":"2026-02-12T16:43:23","modified_gmt":"2026-02-12T16:43:23","slug":"xai-ai-data-assimilation-in-earth-sciences","status":"publish","type":"post","link":"https:\/\/research.reading.ac.uk\/met-darc\/2026\/02\/11\/xai-ai-data-assimilation-in-earth-sciences\/","title":{"rendered":"Can you explain that again? Connecting explainable AI and data\u00a0assimilation\u00a0in Earth sciences\u00a0"},"content":{"rendered":"

By Ieuan Higgs, February 2026<\/p>\n

Perception\u00a0of AI<\/span><\/h4>\n

Artificial intelligence (AI) has become closely associated, in the public imagination, with large language models (LLMs). From a user perspective, AI chatbots such as ChatGPT or Claude operate on a simple and intuitive principle: text goes in and text comes out (or some combination of text, images, sound, and\/or video). This framing has made AI feel familiar to millions. However, it has also narrowed the way in which we might sometimes think about AI, the forms which AI can take, and the potential for beneficial scientific AI applications.<\/span><\/p>\n

Weather prediction and Earth system scientists are developing AI in ways that look very different from conversational chatbots (even if they use many of the same underlying components and optimisation methods). Rather than generating text or media, these models operate on large amounts of numerical data. Their purpose is to attempt to predict the future evolution of Earth system processes. These AI applications can directly benefit people by informing hazards and emergency responses, including storm warnings, flood risk, and heatwave risk. This contrasts quite heavily with some well-known AI applications, which have confusing-at-best and deliberate-misinformation-at-worst impacts on society (is a cat video on social media even funny if it isn\u2019t real?).<\/span><\/p>\n

\"A
Figure 1: Screenshot from a “YouTube Short” of AI generated \u201cTop 5 Caterpillar Videos<\/a>\u201d (left). An AI generated forecast<\/a> of 2m temperature and 10m wind from ECMWF (right). Here, I question whether the application of AI in one of the domains has more potential benefit to society than the other.<\/figcaption><\/figure>\n

Earth system AI models can already offer gains in some operational settings. This is because AI generated forecasts are faster to compute than traditional physics-based model forecasts (while maintaining comparable accuracy).<\/span>\u00a0However,\u00a0if we are using these models in safety critical settings, how do we\u00a0<\/span><\/b>know\u00a0<\/span><\/i><\/b>we\u00a0can\u00a0trust the output?<\/span><\/b><\/p>\n

Trusting\u00a0AI in Earth Science<\/span><\/h4>\n

The main factors affecting trust in AI models usually reflect the differing development paradigms of AI and physics-based models.<\/span><\/p>\n

At a basic level, the two types of model can look very similar. AI weather models (e.g., AIFS<\/a> or FourCastNet<\/a>) often take inputs and produce outputs in the same format as physics-based models. These inputs and outputs are gridded fields of atmospheric variables, including temperature, wind, and pressure. The model evolves the gridded input fields forward in time to produce output fields. These grids contain spatial, temporal, and multivariate structure which corresponds to physical processes such as weather fronts, atmospheric waves, and cyclones.<\/span><\/p>\n

However, in traditional physics-based models, this structure arises from model equations which explicitly encode physical laws. Conservation laws, thermodynamics, and fluid dynamics constrain how information propagates through the system. The complex behaviour of the atmosphere can emerge from these first principles. Decades of development and validation have built confidence in how these models behave. Importantly, this includes an understanding of their limitations.<\/span><\/p>\n

Beyond the AI models being new and lacking this time-in-operation, they also approach the problem from a fundamentally different direction. They do not encode physical laws directly. Instead, the AI models train on examples of atmospheric evolution (e.g., a historical record of combined model runs and real observations<\/a>). They must then infer internal representations that allow them to make predictions on unseen examples (i.e., data which the model is not trained on). If successful, the model has captured something meaningful about atmospheric dynamics. However, this immediately raises some critical questions:<\/span><\/p>\n