Scientific Modeling out of distribution (Scientific-Mood) ML Challenge

The HDR ML Challenge program is hosting its second FAIR challenge, this year presenting three scientific benchmarks for modeling out of distribution in three critical areas: Neural Forecasting, Climate Prediction using Ecological Data, and Coastal Flooding Prediction over time. Machine learning problems are often driven by the quality of the available training datasets. Models are very effective at interpolating across their training datasets to find patterns and trends. In this challenge, we ask models to extend beyond their training by performing out of domain extrapolation to practical critical scientific process that have not yet been well studied. As with the first challenge, we will host three distinct sub-challenges on different scientific problems, with the fourth being the combined challenge. Our focus will be on the critical problems of:

Neural Forecasting: We forecast the activations of a cluster of neurons given previous signals from the same cluster. This targets the critical problem of brain-artificial neuron interfaces, and these models can be used in brain-chip interfaces for artificial limb control, amongst many others.
Climate prediction using ecological data: We predict drought conditions at field sites over short, medium, and long timescales using images of an important group of ecological indicator organisms (ground beetles) collected there. The resulting models can be used to understand how the impacts of future climate conditions may be reflected in observable features of sentinel taxa.
Coastal flooding prediction over time: We model the sea levels at various sites over decades, with the aim of predicting future coastal floods. The resulting model will be essential for understanding the real-world impacts of climate change.