Machine Integration and Learning for Earth Systems (MILES): Bridging Key Gaps in Machine Learning for Earth System Science

Interest in machine learning (ML) has skyrocketed because these types of approaches promise to address key challenges, such as accelerating the forecast generation process, summarizing the exponentially growing amount of available weather and climate data, improving predictive accuracy, and linking observation and model data to impacts. Realizing the potential of machine learning in Earth System Sciences is currently limited by gaps between these new machine learning systems and existing weather and climate infrastructure as well as gaps in understanding between machine learning practitioners and the intended users and operational maintainers. The NCAR Machine Integration and Learning for Earth Systems (MILES) group is working to address some of these key gaps by building a convergent science culture spanning a diverse array of expertise in ML, weather and climate science, and social science; the development of new software packages; educating practitioners and users; and forming key partnerships to share expertise in filling these gaps. In this presentation, we will discuss some of our latest projects aimed at tackling issues like uncertainty quantification, tuning complex ML pipelines on HPC systems, connecting ML and Fortran models, and merging weather and societal data with ML to understand the sensitivities of potential impacts.

Our new uncertainty quantification package, miles-guess, supports training, analyzing, and interpreting neural networks that can explicitly predict both aleatoric and epistemic uncertainties. We have applied it to a range of weather applications and found that the uncertainties produced help discriminate challenging cases and are connected with physically salient features, helping guide forecasters to more challenging situations in their domains. ECHO supports hyperparameter optimization of machine learning pipelines on HPC systems and now has added capabilities to support training across multiple GPUs on multiple nodes, to optimize along the Pareto front of multiple losses, and to analyze how hyperparameter selection changes between random searching and targeted Bayesian optimization. We are benchmarking multiple libraries for incorporating ML models into Fortran-based NWP and Climate models, including both fully Fortran approaches and those that link with external ML libraries. Finally, we have continued development of CRISIS, a machine learning risk model that combines convection-allowing NWP output and dynamic population data to estimate predicted exposure to tornadoes. We have analyzed how the mix of different personal mobility patterns in an area can dramatically affect the potential tornado exposure within a short period of time and which time periods and areas are most sensitive to these changes.