ICLR 2024 recap

The semi-annual Climate Change AI workshop took place last week at ICLR in Vienna, featuring a packed schedule of engaging talks and poster sessions. It was my first meeting in person and I was glad to finally meet all the actors of this vibrant community face to face.

My discussions with the crowd of scientists focused on bridging the hap between research and action. Climate Change is is after all an activist topic that has immediate impact on the society. Two topics emerged:

Collecting data from relevant locations. For example, earth observation models are often calibrated in wealthy countries but they are applied in poorer countries, where the conditions differ. For example, the farms in the US Midwest have very different scale from the small plots cultivated in some regions of Africa.

Making research more accessible. There was a strong consensus that university should shape policies by providing the latest facts on various topics. Some papers deliver crucial insights beyond academic circles. Engaging with journalists or policy makers is necessary for disseminating these findings. Policy makers should access these facts through simpler ways like policy briefs or blog posts. Academics should be encouraged to creating policy-shaping content as part of their professional duties.

Here are some topics that I would love to see being addressed by the community:

Predicting forest health. Can we predict the health (or even the development) of a forest on a 5-year or 20-year horizon? This prediction is crucial for relying on forestry as a carbon sink.

Modeling and understanding resilience We are entering uncharted territory as far as climate change is concerned. The last 4 years broke consecutive temperature records for the ocean. Can we use AI to identify early signs that ecosystems or earth systems are at the brink of non-linear rapid changes?

This year marked a clear paradigm shift in using AI for long-term predictions. AI-based methods are more accurate and much faster than pure physics-based models, especially for extended periods. [[Emily Shuckburgh]] from Cambridge University, in her talk about AI and Climate change. This is a non-obvious and intriguing result: while we have excellent models for many physical phenomenons over short periods (for example, the Navier-Stokes equation for fluids), AI's superior performance over longer times suggests the existence of undiscovered laws governing these systems over larger time scales.