Covering Scientific & Technical AI

“Decoding the Genesis Mission: A New Era for AI-Driven Science with Rick Stevens” is a presentation recently given by Rick Stevens, Associate Laboratory Director for Computing at Argonne National Laboratory, as part of a series hosted by the Google Cloud Advanced Computing Community. The talk is a detailed technical and strategic overview of the Department of Energy’s Genesis Mission, a national initiative for building a platform for AI-driven scientific discovery, energy innovation, and national security.

Rather than transcribing the talk verbatim, AIwire has organized a selection of Stevens’ remarks into a question-and-answer format to highlight the key themes. The questions below were written by AIwire to reflect the structure of the talk and to focus on the Genesis Mission’s impact on the high performance computing and AI for science communities.

AIwire: What is the Genesis Mission, and how is the Department of Energy positioned to lead it?

Rick Stevens

Rick Stevens: The Genesis Mission is the new national AI for science, national security, and energy initiative that’s been launched by the Department of Energy. Over the last four or five years, those of us in the national labs have been trying to launch a large-scale AI for science initiative. We’ve had workshops, meetings, and planning, and with the new administration, there was an opportunity to reframe it and pitch it.

We’re super fortunate to have Dario Gil come in as the undersecretary for science. He was a leader at IBM Research for many years in semiconductor work and later in quantum and is very aware of the state-of-the-art targets for advanced computing. He’s now at DOE, and one of his first actions essentially was to get this AI initiative going.

The name is Genesis Mission. The framing of this is to be really groundbreaking in revving up the U.S. national level of awareness and investment in AI and associated infrastructure and applications to create a new way of doing science and a new way of doing engineering and technology development. And it’s being framed in the same spirit as the Manhattan Project or the Apollo Program. There’s a massive effort to change what’s possible and what’s new. Think of it as phase three of the great eras of American science.

Why launch Genesis now, and what problems is it intended to address for U.S. scientific productivity?

One of the arguments for doing a massive AI initiative is that the scientific output, as a function of resources going into it, has been stagnating, or actually decreasing, over time. And this is a well-studied phenomenon, not a recent phenomenon. It’s been going on for decades.

And there are some hypotheses about why this might be happening. One is that low-hanging fruit has disappeared, or science problems are getting harder, or there’s friction of larger and larger teams needing to attack open questions. But one of the notions is that AI and high performance computing, coupled with quantum, may be able to tackle this challenge, which is trying to increase scientific output as a function of input. That framing is one of the important goals of Genesis Mission.

The long-term mission is to double U.S. research productivity over the next 10 years, and not just in national labs or government-related things, but across the entire ecosystem, the entire country. The U.S. spends about a trillion dollars a year in R&D, about three and a half percent of the GDP, and the goal of Genesis is to have the impact as if we were spending twice as much.

(Credit: DOE)

What is “closed-loop discovery,” and how does the convergence of AI, exascale computing, and DOE’s experimental facilities change how scientific discovery can be carried out?

We’re starting to see across many fields — chemistry, biology, materials science, other areas of science and engineering — this emergence of a of a concept of a closed loop discovery paradigm where we have AI, maybe agentic networks, that is able to come up with hypotheses, do simulations, control simulations. Maybe it’s searching for a new material, or, say, designing a protein, then carrying out experiments to validate the conjectures that the AI and simulations have come up with, then taking measurements and feeding that information back into the models and loop.

There are a number of companies and startups, about a half a dozen in the U.S. and more internationally, that are building business models essentially on this concept. And this idea, we think, has legs. It was not possible until recently to imagine doing this because AI wasn’t powerful enough, and we didn’t have the ability to glue everything together.

DOE has been a leader in high performance computing for decades, and we have exascale computers. We have the world’s largest collection of large-scale experimental facilities under one logical roof. There are about 30 user facilities that DOE manages, and we have a huge volume of scientific data from historical operations of facilities, about 80 years of data in many areas.

So this idea is an initiative that integrates all of that and identifies these things called Lighthouse challenges. Lighthouse challenges are just another term for grand challenges, but aimed at the future and possibly attackable with AI. Can we accelerate in energy, basic science, and national security by integrating all of these things with AI and building a platform to do that? That’s the motivation for Genesis, a national initiative to build a platform.

(S. Singha/Shutterstock)

When you describe Genesis as a platform, what are its core components, and how does that platform change the kinds of scientific problems researchers can realistically tackle?

What we mean by platform here is the combination of the hardware, software stack, data sets, and AI models needed to accelerate discovery science and national security and drive energy. That’s the whole goal of it in one sentence, essentially.

Another thing that we’re trying to do is enable us to work on harder problems. Increasing research productivity doesn’t just mean that the same people produce twice as many papers. That’s not what we’re after. What we’re trying to do is bring in timelines. So something that a community might believe is possible in 10 years, we’ll try to bring that in, say, by a factor of two, to five years.

Or say there are problems that we believe are too complicated, like a generalized method for designing proteins for room-temperature catalysts, for example. People might not have been willing to tackle that complex a problem, but something like Genesis and the platform could make our scientists bolder and more fearless to go after harder problems.

Why is making scientific data “AI-ready” such a central challenge for Genesis, and how might the planned platform treat data and models as shared, reusable assets?

One of the main challenges is that we have a large collection of data, since the Manhattan Project, actually, and much of that data is not in an AI-ready form. Getting it into an AI-ready form is critical. Working out how to connect the basic science component of national security with the production facilities DOE operates is a target, as well as reducing red tape and improving the efficiency of all the processes on the national security side.

The way to think about the platform is it’s more like an operating system that we’re trying to build, but it’s distributed, built on ESnet, built on the integrated research infrastructure that DOE is trying to build to connect all the facilities. It also includes models and data, and we have teams working on an inventory of models — not just foundation or general-purpose models, multimodal models that we’re building — but domain-specific models for biology, chemistry, material science, cosmology, and so forth. These all become assets on the platform, and then the underlying data sets that DOE has, whether it’s material science or cosmology, high energy physics, etc., all of those become assets in the platform that users of the platform can integrate with low overhead into building new models or testing hypotheses.

(Illus_man/Shutterstock)

How are lighthouse problems shaping the Genesis platform, and what does the near-term roadmap look like?

One of the things in the executive order that launched the initiative back in November was that the Department was to deliver a list of high-priority problems to the White House. There was a list of 27 of these lighthouse problems that was sent over to OSTP and hopefully will be made public very soon. These problems were chosen to be not just important Grand Challenges that, if we solve them, will advance U.S. leadership in science and technology, but they also provide a driver function for development of the platform. So think of the lighthouse problems with the applications as having a requirement setting or a pull on the platform development, and there’ll be a feedback between advances, making the platform able to advance the application, and back and forth.

These things are supposed to demonstrate transformative AI, not just say, workflows on existing AI, but advances in AI, like increasing the modality of the AI, or as I mentioned before, the ability to reason over physics multi-scale problems. It’s also to build on top of the current best state-of-the-art AI, so that we’re building agentic networks that use the best of frontier models augmented with domain-specific models. And these things, as much of it as possible, will be public and open. Of course, some data sets can’t be open, and we have the provision for partnerships with companies where data might be proprietary. But the intent here is to have a significant amount of openness that would allow additional partners and academics to build on this.

There’s an ongoing effort right now to create an inventory and mapping of federal compute resources that can become part of the platform. We’re currently working on a data set inventory that will become part of the platform. And these AI-directed lab capabilities include things like automated labs, model cards, and so forth. We will have a significant initial set of demonstrations of the platform in March, further demonstration in July, and a release of the initial 1.0 version of the platform next fall.

What makes scaling compute for AI-driven science different from traditional HPC, and how is Genesis being designed to address that?

We need a lot of compute, and we don’t just need compute for modeling and simulation. We need a large-scale compute for AI inference. One of the challenges in trying to talk about AI for science is trying to come up with a way of talking about scaling compute. It’s similar to how we used to talk about it in HPC. We used to talk about moving from, say, petascale to exascale. With a factor of 1,000, we could draw a diagram or a chart up into the right and say, “Well, if we have 1,000 fold improvement, we can increase the resolution of a simulation, or reduce the time steps.”

(Credit: Rick Stevens, DOE)

We don’t have a way to talk about that exactly yet in AI for science. We, of course, do talk about token production and LLMs, and we’re trying to relate how many trillions of tokens we might need to do an inference to solve a given problem, and I’d say this is very much a work in progress. This table (see image above) gives you a ballpark of some of the estimates we’ve come up with based on early experiments. For example, to do an autonomous nuclear reactor design, how many trillion tokens might be needed to do that? And the estimate right now is like 10 to 50 trillion tokens. And you can think about what that might cost if you’re buying tokens, say, in Gemini or something like that.

You can see these different estimates, like co-scientists here, which is something that we’re collaborating with Google DeepMind on. We have an understanding of how token consumptive they are, but how many users in the future will want to do co-scientist type things? What’s the diffusion rate going to be and so on? That’s what we’re interested in. Our current estimate, based across the complex for people who are using state-of-the-art models to do things like vibe coding, is that we need about 120 trillion tokens a year to drive the vibe coding that we want to do. That’s in addition to much of this list. So we need a huge amount of tokens. That means we’re going to need to deploy infrastructure. We’re going to need to buy tokens from partners and really innovate in high performance AI. The platform — it’s the data, it’s the models, and it’s the infrastructure.

To see Stevens’ entire presentation, including a live Q&A session where he discusses topics ranging from scientific reasoning models to data ownership, partnerships, and the challenges of validating AI-driven discovery, visit this link. To see more Google Cloud Advanced Computing Community events, go here.

This transcript has been edited for length and clarity.

This article first appeared on HPCwire.

Related