Introduction to Surya: A Foundation Model in Heliophysics

Artificial Intelligence (AI) has emerged as a transformative tool for establishing connections between data and acting as a tool that empowers humans to solve a wide variety of problems for which there were no easy solutions. In heliophysics, AI has been applied to a wide variety of problems including space weather forecast, data calibration and homogenization, computer vision (i.e. segmentation), coronal field extrapolation, etc.   To this date, the overwhelming majority of AI applications in Heliophysics are targeted and designed to address specific problems.   However, the rise of very powerful multi-purpose models has revolutionized natural language applications (e.g. GPT) and have the potential to revolutionize how we derive scientific value from vast quantities of data.

Large multi-purpose AI models, i.e. “Foundation Models”, are large AI models that are self-trained on vast quantities of data to enable a wide range of applications, while paying the lionshare of the training cost upfront. Successful foundation models, such as GPT-4, can then be fine-tuned at a fraction of the cost by anyone with access to the pre-trained model. In this talk we will introduce the concept of Foundation models, how they are built, and how they can be used. We will focus on Surya, a NASA-funded foundation model, constructed using data from the Solar Dynamics Observatory (SDO). We’ll show preliminary results, our validation strategies, and how we aim to push the Surya project beyond the training of large models with vast quantities of data. Our vision involves the creation of frameworks alongside the trained model that enable an entire community to use the foundation model for scientific inference and we’ll discuss our efforts to achieve that. Finally, we will discuss some interesting potential applications of Surya that can elevate our use of foundation models.

Andrés Muñoz-Jaramillo, SwRI, Madhulika Guhathakurta, NASA HQ, and the Surya team