(Shutterstock)
The opening plenary session at TPC25 examined how science is evolving alongside advances in AI and high performance computing. This summary highlights insights on large-scale models from Rick Stevens, high performance hybrid computing from Thierry Pellegrino, and agent-based simulation from Flora Salim. Together, their talks explored how new tools and approaches are reshaping scientific work across domains.
Reinventing Discovery: Accelerating Science in the Age of Artificial Super-Intelligence
Tuesday’s opening keynote came from Rick Stevens, associate laboratory director for computing, environment, and life sciences at Argonne National Laboratory. As a founding member of TPC, Stevens outlined how the consortium, as well as the research community, are adapting to the growing capabilities and accessibility of frontier AI models.
“I think we’re entering a very interesting state, with continued progress on the commercial models, and on open models as well, towards this amazing capability that is challenging all of our assumptions about what was possible, 5 or 10 years ago, in terms of how fast AI would develop,” he said.
Stevens described today’s frontier models as partners that can co‑design research questions rather than merely speeding up workflows. He framed recent advances with two illustrations. First, there are newer reasoning models earning gold‑level scores on the International Mathematical Olympiad, a milestone Stevens says highlights how rapidly AI is matching elite human problem‑solving. Second, he described an experiment using a slide from a colleague’s talk. Using OpenAI’s Deep Research, Stevens gave the model a single page of scientific figures, and within minutes, it drafted a concise summary before expanding that summary into an 18‑page research review.
The talk then shifted to the growing importance of open models. Stevens highlighted Kimi K2, a trillion‑parameter model released by China’s Moonshot AI, as a landmark for the open community and a sign that science‑focused models of comparable scale are now within reach
“It’s the first open model that’s really a trillion-parameter model. This is a milestone in some sense of its relevance to the TPC because we imagined a few years ago that there would eventually be the prospect of creating open models for science that would be at this scale, and so we’re seeing the first emergence of this coming,” Stevens said.
He also highlighted a 32‑billion‑parameter reasoning model, AM-Thinking-v1, that was entirely built using the open-source Qwen2.5-32B base model. Stevens said it is good for coding and easily runs on his laptop, showing how advances in smaller network infrastructure could put advanced AI within reach of many laboratories.
Looking ahead, Stevens expects a new wave of embodied AI. Pairing language‑vision models with capable robotics will let researchers build systems able to carry out physical experiments and gather data with minimal human intervention. The scientific workforce, he said, will need to rethink how it integrates these new assistants while preserving reproducibility and integrity.
To prepare for this future, Argonne and partner institutions are sketching what Stevens called a “scientific AI platform.” The concept is a layered network: large reasoning models hosted locally or in the cloud, on top of knowledge graphs, simulation codes, automated labs, and domain‑specific foundation models. He stressed that the platform must remain open and collaborative so that academic users can adapt it to their own disciplines.
Stevens concluded by offering a rough estimate of the computing resources required to equip each scientist with an AI assistant capable of working about one hundred times faster than a person. He estimated each researcher would need about two to four GPUs. When multiplied across the whole scientific workforce, that demand grows fast and raises clear energy concerns, but things are happening fast, just like another time in history.
Another time such as the Industrial Revolution, to which Stevens likened today’s AI surge. Just like when steam power was able to be converted into machine labor, today’s datacenters, Stevens says, are turning electrical power into cognitive labor, raising urgent questions about how society should steer that capability and what goals it ought to serve.
“What’s happening with large-scale AI is it is the ability to take power and energy and translate that into cognitive labor,” he said, noting that AI’s ability to steer enormous amounts of cognitive labor leaves scientists to wrestle with many questions.
“It’s the question of our age, which is, what is the relationship between people and scientists and academics, and this engine that converts power into cognitive labor? How do we work that relationship? What should that relationship be, and what should the goals of an engine able to produce such output be? Those are the questions that I hope people are thinking about as we move along.”
HPC and the Science: The Need for Hybrid by Thierry Pellegrino
With rapid advances in GenAI, quantum hardware, and cloud infrastructure, the high-performance computing landscape is changing fast. Thierry Pellegrino, who is the Global Head of Advanced Computing at AWS, argued that hybrid models will be the key to scaling innovation across industries.
To explain why this shift matters, Pellegrino pointed out that high-performance computing is already quietly driving many parts of modern life. From weather forecasting to drug development and complex engineering, HPC plays a central role in both daily routines and major industries. What is changing now, he said, is the pace. “We can innovate and start a supercomputer in a matter of minutes.” That level of speed and flexibility would have been unthinkable a decade ago.
This shift is making advanced computing far more accessible. Researchers and organizations that once had to wait years for deployment or lacked the resources to build large systems now have tools available on demand.
However, accessibility alone is not enough. Pellegrino stressed that to meet the growing complexity of science and technology challenges, “we must move from isolated systems to integrated ones.” That means adopting hybrid models that bring together AI, cloud infrastructure, and quantum computing to accelerate discovery and scale results.
Thierry Pellegrino at TPC 2025
So what do these hybrid models actually look like in practice? Pellegrino described three dimensions of hybrid HPC that are already reshaping how science and industry move forward.
The first is the integration of AI and HPC. AI now supports everything from generating synthetic data to accelerating simulations and interpreting results. It is changing how scientists work day to day, making complex research faster and more efficient.
The second dimension involves combining on-premise systems with cloud infrastructure. Pellegrino stressed that this is “not the story of cloud or on-prem,” but rather a model that benefits from both. Cloud access allows teams to scale quickly, test new hardware, and collaborate more easily.
The third dimension is the convergence of classical and quantum computing. Though still in early stages, quantum is starting to speed up select workloads, while HPC and AI are playing a key role in designing the quantum systems of tomorrow.
Pellegrino closed by highlighting the importance of making powerful systems easier to use. As computing continues to evolve, he said it is just as important to simplify the experience as it is to expand capabilities. Scientists and engineers need tools that help them focus on their work, not the complexity underneath.
He emphasized the value of removing barriers so that more people can engage with advanced computing. “We can make it a big way to take that weight off the shoulders of everyone,” he said, pointing to the need for intuitive frameworks that meet researchers where they are.
This summary only scratches the surface. Pellegrino’s full talk includes stories from the field and a closer look at how AI, cloud, and quantum are starting to work together. If you want to hear the full talk, including audience questions and his thoughts on what is coming next, the complete video and transcript will be available soon at tpc25.org.
Tackling Complex Behavior with Agentic AI by Flora Salim
Sometimes it’s best to start with the conclusion. Flora Salim, professor in the School of Computer Science and Engineering at the University of New South Wales (UNSW) Sydney, Australia, has an expansive vision.
“I see the world where we can actually build digital twins, but digital twins not data-driven, but model-driven. A lot of people think about digital twin as actually data points. Actually, they are models, models of the world. It could be agents that simulating mobility, behavior, energy and grit, and we can do rapid, very rapid water scenarios, scenario planning. We can actually then build decision making systems that are robust,” is how Salim closed her talk, the third, in yesterday’s TPC conference plenary session.
Getting from here to that vision will require many advances, not least in the ability to effectively collect and deal with the vast amounts of complex spatial and dynamic data. The heart of her talk —Understanding Complex Behaviour in Dynamic Cyber-Physical-Social Systems — discussed approaches for dealing with that data.
She began by emphasizing that data capturing real-world behaviors — such as mobility routines, energy-use patterns, and decision-making in urban and digital environments — are inherently noisy, context-dependent, and often only partially observed. She touched on recent progress in understanding behavior at scale through data-driven modeling and simulation, highlighting the convergence of data-efficient learning, generative models, and agentic AI for complex systems analysis
“To be able to understand the dynamic, complex behavior in our cyber, physical, social systems, we have lots of time series and spatial temporal data currently being collected in many domain because of the advances of industry. 4.0, IoT and I believe some of you are working on time series as well,” noted Salim.
“Better understanding this time series data can help us to understand different domains. [Take time-series data] It is so challenging. The time series themselves, it has dynamics, but also imperfect. A lot of sensor data, they’re noisy. By default, they’re noisy,” she said, citing accelerometers as an example; these devices ae widespread in smartphones, smart watches, cars, and their differences must be accounted for.
Salim’s talk (82 slides) left virtually no stone unturned is discussing the problems of data cleaning and ranking. The full scope of these remarks are experienced in full and links to her recorded talk and slides will be made available by TPC shortly.
The fruits of all that work, of course, is being able to use the data in effective models. Salim walked through several example but dwelled most on her team’s work on “SOCIA: an Agentic Simulation Generation Engine for Cyber, Physical, and Social Behaviours” (see associated paper)
“This is version one. So what it is, we actually using social media to generate codes, but these are simulation codes. It could be simulating the physical space, it could be also simulating group behavior, or social behaviors, a combination of any of these. What makes it different from other work is basically a lot of the other agent-based work typically only focuses on one system, either online social media behavior or simulating behavior movement. But we actually integrate cyber-physical-social system design thinking in SOCIA.
“So this is a multi-agent system that actually monitors agents that work under the workflow to understand the task and the data that are given to it, and also feedback given to it. And this is basically the agents working on the workflow manager that generate codes, verify the codes, execute the stimulation, give you, give results, but then keep iterating until you get the best result, because we also provide the evaluation metrics.”
Again, Salim’s extensive remarks and slides are best seen directly and will be made available by TPC and the Tabor family of publications. She emphasized that her group open sources its data and code. Below is the abstract from their paper on SOCIA.
Abstract from SOCIA paper
This paper introduces SOCIA (Simulation Orchestration for Cyber-physical-social Intelligence and Agents), a novel end-to-end framework leveraging Large Language Model (LLM)- based multi-agent systems to automate the generation of high-fidelity Cyber-Physical-Social (CPS) simulators. Addressing the challenges of labor-intensive manual simulator development and complex data calibration, SOCIA integrates a centralized orchestration manager that coordinates specialized agents for tasks including data comprehension, code generation, simulation execution, and iterative evaluation-feedback loops. Through empirical evaluations across diverse CPS tasks, such as mask adoption behavior simulation (social), personal mobility generation (physical), and user modeling (cyber), SOCIA demonstrates its ability to produce high-fidelity, scalable simulations with reduced human intervention. These results highlight SOCIA’s potential to offer a scalable solution for studying complex CPS phenomena. SOCIA is still under development, and we will release the code once it becomes available.
Contributing to this article were Ali Azhar, Douglas Eadline, Jaime Hampton, Drew Jolly, and John Russell.