LLMs and weather prediction
As I mentioned in several of my previous posts, the field of numerical weather prediction is experiencing a significant transformation from traditional numerical weather prediction methods to data-driven machine learning approaches. It is hard to be up to date with the new developments, but I want to report on my latest futile attempt to do so. My thoughts below are highly speculative, but given the rapid pace of development in the field they might become a reality in the next 5-10 years.
With the recent release of an operational version of the AIFS, these days I was thinking that it is possible that in the not so far future we might end having some kind of large language models (or, more precisely, a foundational model that would be trained on an absurd amount of weather data to predict the weather, shifting from purely numerical predictions to event forecasting systems that can interpret and predict weather events in natural language.
A quick search took me to this paper by Li et al that introduces CLLMate, a model that aims to combine meteorological raster data with natural language processing capabilities, forecasting of weather events rather than just numerical variables. The author's idea is to
- First develop a knowledge graph by using an LLM to extract weather and climate events from a corpus of environmental focused news and articles.
- Map these events to meteorological raster data (I guess they mean grib or netcdf format gridded data)
- Create a supervised data set from the two previous data sets.
An example output, taken from the paper, is below.
I find this approach interesting because it creates a bidirectional mapping between numerical weather data and real-world weather events as reported in news and articles. This connection could be used in several ways:
Improved Event Detection: By learning from historical weather events and their corresponding meteorological patterns, the system could we used to identify potential severe weather events before they fully develop.
Context-Rich Forecasting: Unlike traditional numerical weather prediction models that output raw parameters, this approach can provide contextual information about how weather patterns might impact different sectors or regions based on historical precedents.
Historical Pattern Analysis: The knowledge graph built from news articles could help identify similar weather patterns from the past, providing valuable context for current forecasts and potential impacts.
On the other hand, the accuracy of a language model in interpreting meteorological data needs to be rigorously validated, and there's always the risk of the model generating plausible-sounding but incorrect forecasts (ie, hallucinations). Additionally, the training data needs to be carefully curated to avoid biases in reporting and ensure global coverage. Maybe the increasing number of publicly available data-driven models could fill the gap.
Maybe in the future this fusion of LLMs and numerical weather data could lead to more sophisticated weather communication systems. One could potentially ask a weather model directly "What will the temperature be tomorrow in location X?" without the need to do any data processing. In a way, weather apps do this already, but they rely on a particular weather model in the backend. My thoughts are more on the lines of asking an integrated LLM-weather model "run this weather model for the next 7 days and make me a dashboard of all relevant parameters for application X, Y and Z".