ABSTRACT
AI/ML techniques are finding their way into many scientific disciplines, including in the sciences relating to Earth and its systems. They can be used to simplify a complex dataset or process, breaking it down into subregimes that are easier to study individually. A second significant activity that they facilitate is the compiling of new datasets, by detecting and extracting phenomena of interest from large data, thus making existing datastores much more valuable. AI can also be applied in order to reduce computational cost. One of the more active uses in this vein is surrogate modeling, wherein part of a simulation is delegated to AI because it can often provide predictions much faster than a physics-based simulator can. AI/ML can also be used to examine or correct biases in a first principles model, by comparing its results to observational data, thereby learning to predict systematic errors. Finally, we discuss some cases where AI has provided a leap forward in computational sciences. All the above come with benefits and challenges, the tradeoffs of which need to be considered for any particular use case.
