An AI model to unlock the Sun's secrets

Welcome back to the Circuit Breaker, where you can find the best recaps on the latest innovations in AI, quantum computing, semiconductors, and more, from across IBM Research and beyond.

Week of August 19 - August 22

• IBM and NASA team up for a new solar physics foundation model
• The US Open is serving up new experiences with IBM
• Understanding AI through the algorithms they compute

Introducing Surya, an AI model for the Sun

This week, IBM, NASA, and eight other partners open-sourced Surya, a heliophysics foundation model that can be used for all sorts of scientific questions related to our host star, including predicting solar flares, solar winds, and the Sun’s magnetism.

Why do we need an AI model for the Sun? Two reasons: AI can make mountains of satellite data accessible in new ways. It can also potentially give us more advanced warning of impending solar storms.

When NASA’s Solar Dynamic Observatory launched into space, scientists didn’t really have the tools to analyze the satellite’s daily haul of 1.5 terabytes of data. Now, 15 years later, they do. IBM, NASA, and collaborators used nine years of raw SDO imagery to create an AI model named Surya that could not only predict space weather but provide new insights into the Sun itself.

Space weather hits Earth hardest during the Sun’s active phase, which peaks every decade or so. This phase is marked by solar flares, solar wind, and jets of plasma shooting charged particles towards Earth. If the particles interact with Earth’s magnetic lines, the energy can endanger astronauts in space and throw off satellites, power grids, and communications. Researchers hope that Surya and a related set of benchmarks can help us see solar storms coming sooner.

"We want to give Earth the longest lead time possible,” said Andrés Muñoz-Jaramillo, a solar physicist at the SouthWest Research Institute and a lead scientist on the project.

Why Surya? It’s Sanskrit for Sun, and it follows IBM’s Prithvi (Sanskrit for Earth) family of AI models for weather and climate prediction on Earth, among other applications.

How was Surya trained? Instead of learning to reconstruct partially blacked out images during training, as Prithvi was, Surya was given sequential images and asked to predict what SDO would see an hour in the future. In the process, the model inferred knowledge from the raw data essential for skillful forecasting — things like the Sun’s geometry, magnetic structure, and its unusual rotation.

Were there any surprises? Yes! One of the Sun’s quirks is that it spins faster at its equator than its poles. To do any kind of forecasting, you have to correct for the Sun’s differential rotation. Researchers at first tried to encode this knowledge directly into the model, but they got better results when they let the model be. “It’s remarkable that the model performed better when we let it figure out the rotation on its own,” said Johannes Schmude, an AI researcher who led the project on the IBM side.

☀️ Read more about what Surya can do

Game. Set. Match Chat.

IBM and the (USTA) United States Tennis Association are serving up a new level of fan engagement for the 2025 US Open. From real-time match insights to interactive assistants, this year’s AI-powered digital experience is designed to keep tennis fans close to the action. These new ways to enjoy the matches were built with IBM watsonx Orchestrate, and the Granite family of LLMs, born out of IBM Research. Here’s what’s new:

💬 Match Chat is an interactive AI assistant that's ready to answer fan questions during all 254 singles matches. Whether you’re looking for stats, head-to-head records, or even help pronouncing a player’s name, all you have to do is ask.

📊 SlamTracker is back again this year, and it’s enhanced with live “Likelihood to Win” probabilities, updating in near real-time based on player stats, expert analysis, and match momentum.

📝 Key Points is built with watsonx and generates instant TL;DR summaries of articles and match data, so fans never miss the story behind the score.

🎾 Explore all the new updates here

How a decades-old classical computing technique can illuminate the inner workings of AI

Despite the massive leaps we've taken into this new era of AI, our ability to interrogate the inner workings of large models remains surprisingly ad hoc. A new paper in Nature Machine Intelligence builds on decades-old complexity theory, proposing a new way to bring rigor and transparency to assessing the computational complexity of large AI systems

Mid-20th century innovations like Turing machines and the birth of computational complexity theory gave us formal tools to define computability and categorize problems by their inherent difficulty.

• 📈 Modern large language models deliver impressive results, but they evade traditional complexity hierarchies. This leaves us without a systematic way to analyze their time and memory requirements, or to verify the steps they take when "reasoning."
• 1️⃣ solution includes casting AI inference as circuit computations: mapping operations onto directed acyclic graphs and measuring metrics like circuit depth. This technique, argue the researchers behind the new paper, yields a more precise framework to quantify what these models can and cannot compute.
• 🧪 Testing AI on algorithmic tasks with known complexity produces verifiable outputs and intermediate steps, exposing flawed reasoning, guiding benchmark design, and laying a principled foundation for building more reliable, efficient AI.

👀 Read more about the research here

IBM Quantum platform updates from Jay Gambetta

What is shadow AI?

Sports fans are turning to AI

Making it easier for anyone to deploy agents

Take the IBM Quantum annual feedback survey

Highlighting new publications from IBM researchers that we liked the sound of:

• Quantifying artificial intelligence through algorithmic generalization
• SkyStore: Cost-Optimized Object Storage Across Regions and Clouds
• Co-optimization of codon usage and mRNA secondary structure using quantum computing

If you liked this, please consider following IBM Research on LinkedIn. And if you want to go even deeper, subscribe to our monthly newsletter, Future Forward for more on the latest news on breakthroughs in AI, quantum computing and hybrid cloud.