Sustaining progress in Earth system science requires an ongoing commitment to research in emerging technologies, evolving workflows, and the growing challenges of data management and analysis. As we prepare to enter the era of exascale computing, that research includes understanding how our existing software will perform on new systems and how to adapt for the future.
Ongoing research commitment
Scientists anticipate being able to perform innovative science in the exascale era that would not have been possible on traditional CPU-based supercomputing systems. As the technology rapidly evolves, CISL software engineers evaluate how models, visualization, and data-analysis software work on new and promising computational hardware to anticipate how to meet future needs for new NCAR high-performance computing systems.
GPUs play an increasingly large and important role as they make it possible to tackle science problems at significantly higher scales. Our software engineers work closely with NCAR scientists and others to modify climate and weather models to take advantage of the enhanced parallelism that GPU technology makes possible. This enables scientists here, throughout the United States, and internationally to better address complex issues such as estimating the regional and local impact of weather events and climate trends.
Ever-increasing volumes of data present a significant challenge for scientists in any field, and particularly for those whose work involves running numerical simulations of complex processes like those involved in understanding our weather and climate. CISL computer scientists’ research priorities include exploring new approaches to storing, managing, and exploring these vast volumes of numerical data.
Focus areas include how to use lossy compression to reduce file sizes while preserving essential data; harnessing emerging processor architectures to speed visualization and analysis; and employing cloud computing to scale workflows.
While compressing a few images or text files is a simple matter, compressing scientific data residing on multi-dimensional grids is another story. Studying new techniques for compressing data reliably and efficiently is essential for reducing our data footprint.
Cloud computing offers the potential to scale a center’s on-premise computing resources on demand by “bursting” overflow to a public cloud computing service provider such as AWS, Google Cloud, or Microsoft Azure. Bursting provides an economical means to expand available computing resources on a short-term basis. Moreover, providers of vast troves of Earth sciences data, essential to the NCAR mission, are increasingly turning to commercial cloud resource providers to host their data collections. Effectively making use of these capabilities in tandem with traditional HPC computing resources is an area of active research and development for CISL.
Artificial intelligence research
Weather and climate models are among the most computationally intensive of all software programs, and they push the performance limits of today’s supercomputers. The pace of computer hardware advancement no longer meets the need for timely development of future model components, so we are investigating the use of artificial intelligence (AI) emulators to approximate the behavior of more complex weather and climate model components. Our objectives are to increase the models’ overall capabilities while significantly lowering the computational costs of running them. We are working with scientists throughout the community to validate the physical and computational performance of CISL-developed AI emulators for NCAR’s weather and climate models.
As AI development for Earth system science applications evolves from research toward operational use, we also need to ensure that the AI systems are worthy of trust by a wide range of potential users. In partnership with the National Science Foundation’s AI Institute for Research on Trustworthy AI for Weather, Climate, and Coastal Oceanography, we continue to investigate how to make AI systems more explainable, physically consistent, and robust to noisy and flawed data.
Data assimilation research
Data assimilation is the process of combining a model forecast with observations to produce more accurate forecasts than are possible using either one alone. DA is also used to improve numerical models and to provide comprehensive data sets that describe components of the Earth system. We focus our scientific research in this realm on the Data Assimilation Research Testbed (DART), a community software facility that supports data assimilation for many different kinds of geophysical models and many types of observations.
Our work spans a broad range of applied and fundamental research, all with an eye on improving our ability to understand, diagnose, and predict important aspects of the Earth system. For example, DART allows modelers and experimenters to use state-of-the-art data assimilation (DA) algorithms with a minimum investment in software and algorithm development and without having to become experts in DA. To ensure that DART meets the continuing challenge of scaling efficiently on the latest hardware, our scientists develop novel algorithms that can run fast on NCAR supercomputers as well as other hardware in use in the Earth system science community.
We also focus on researching applications like space weather, flood prediction, chemical constituents in the atmosphere, and other Earth system components. Each new application presents novel challenges that can be addressed by collaborative research between DA scientists, model experts, and observation specialists. Research results can include improved predictions, improved models, and a better understanding of the value of existing and proposed observing systems.