CURC User Success story- Accelerating Discovery: How Arjun Rao Uses High-Performance Computing to Transform Satellite AI
Written by:
Addison Schneider
Arjun Rao is a PhD student in Computer Science whose research brings together machine learning, large-scale computation, and satellite imagery. He has long been drawn to computer science, with a particular focus on making machine learning faster and more efficient. He is especially interested in applications within the space industry. While publishing his findings in respected journals is one of his goals, Arjun is equally motivated by the desire for his work to create a broader social impact. He hopes his research can reach the places where it is needed most.
Along the way, he has faced challenges, such as accessing computational resources needed to advance his work. Thanks to the University of Colorado Research Computing (CURC), Arjun has been able to address these computational resource challenges using Grace Hopper 200 superchips. These superchips are designed to merge the strengths of CPUs and GPUs, delivering the kind of power that cutting-edge AI research, such as Arjun’s, requires. A team of experts is also available to assist researchers with any high-performance computing needs. A quick Google search and a three-minute sign-up form was all it took to get started. Today, he credits CURC as being unlike any other university resource, providing essential support for his projects.
Arjun has already published two papers with his advisor, Dr. Esther Rolf. His first project, titled, Using Multiple Input Modalities Can Improve Data-Efficiency and O.O.D. Generalization for ML with Satellite Imagery demonstrates that combining additional geographic data, such as elevation models and land-use maps, with satellite images, leads to smarter and more reliable AI. This approach helps models learn effectively, even with scarce labeled data, and allows them to perform well in new regions. The study also shows that straightforward methods for fusing data can often outperform more complex techniques, making the models both efficient and accurate for real-world use. Arjun's most recent work, Measuring the Intrinsic Dimension of Earth Representation, studies geographic implicit neural representations, which are AI models that represent maps, terrain, or spatial data as continuous mathematical functions, rather than fixed grids or pixels. His research revealed that these models, once thought to be highly complex, can actually be described using only two to ten simple variables. He also developed a metric to measure their true information content and how easily AI systems can learn from them.
Venturing into uncharted territory, Arjun is one of the few researchers in HPC tackling these unique challenges. What surprises him most is how small improvements in computation can create a large impact for practitioners in the real world. He emphasizes that the support of his advisor, Dr. Rolf, along with CURC’s resources and team, especially Brandon Reyes (CURC User Support), has been essential, noting that 90% of his research has relied on CURC resources.
Looking ahead, Arjun hopes to continue being empowered by CU resources and has started his next, highly compute-intensive project. This work focuses on location encoders, which are models that take longitude and latitude as input and predict what can be observed at a given location. While promising, these models have a fixed resolution, meaning they cannot zoom in as much as researchers might want. Arjun’s aim is to study the mathematical properties of these models and address this limitation. By pushing the limits of machine learning and computation, Arjun hopes not only to advance the space industry but also to develop tools that can make a real difference for people here on Earth.