[VT | July 22, 2026 | Washington D.C.]
From protein structure prediction to mathematical proofs, X-ray imaging, power-grid simulation, and the design and manufacturing of flight-test vehicles, artificial intelligence is moving into different stages of scientific research.
At the Technical Keynote of the inaugural Genesis Mission Annual Summit, Darío Gil, Under Secretary for Science at the U.S. Department of Energy (DOE) and leader of the Genesis Mission, outlined the mission’s vision for AI for Science — applying artificial intelligence to scientific research and discovery — alongside four Science Lightning Talks focused on mathematics, advanced imaging, the electric grid, and national security.
The scientific problems across these fields are very different. Together, however, they address a more specific question: When AI for Science moves from national research strategy into laboratories and scientific workflows, what does it actually change?
From 200,000 Protein Structures to 200 Million
Gil began with protein structure prediction.
In 1971, Brookhaven National Laboratory helped establish the Protein Data Bank. For decades, scientists determined three-dimensional protein structures through experimental methods such as X-ray crystallography, using diffraction patterns to reconstruct molecular structures. After roughly 50 years of work across the global scientific community, experimentally determined structures had grown to about 200,000.
AI changed the scale.
DeepMind’s AlphaFold uses neural networks to predict three-dimensional protein structures from amino acid sequences, expanding the number of available predicted protein structures to roughly 200 million.
The change is not simply one of speed. Protein folding involves enormous combinatorial complexity: an amino acid sequence can potentially take on a vast number of three-dimensional configurations. Science has traditionally addressed this problem through painstaking experimental measurement and accumulation. AI can learn structural patterns from existing data and generate predictions across search spaces that were previously difficult to navigate.
Gil extended the same idea to genomics. Researchers are now training foundation models on large-scale DNA sequences to learn biological patterns across species. He cited training data on the scale of approximately 9 trillion base pairs across 128,000 species.
From proteins to genomes, AI is beginning to address scientific problems previously constrained by search space, data scale, and computational capacity. But AI does not enter every discipline in the same way. The four Science Lightning Talks that followed showed how, as AI enters different research workflows, the bottleneck begins to move.
Mathematics: From “Proof Scarcity” to “Proof Abundance”
Terence Tao, Professor at the University of California, Los Angeles (UCLA) and a Fields Medalist, brought the discussion into mathematics.
Tao described AI entering multiple parts of mathematical research, including problem solving, code generation, automated formalization of proofs, and literature search. In some tasks, he said, researchers are already seeing improvements of tenfold or even a hundredfold. But the change is uneven: many forms of mathematical collaboration still resemble the traditional model of a few researchers working together.
A more fundamental shift is taking place around the proof itself.
Tao pointed to a collection of roughly 1,200 significant mathematical problems from the 20th century. About a year ago, roughly 300 had been solved; that number is now approaching 500. He observed that new model releases have been accompanied by noticeable jumps in progress on the list.
But an AI system producing an answer does not mean the answer is correct.
That makes formal verification increasingly important. AI can generate candidate proofs, while formal systems can independently check whether their logical steps hold.
Tao described the transition in two terms: “proof scarcity” and “proof abundance.”
For centuries, mathematics has largely operated under proof scarcity: the difficult part was finding a proof. If AI systems can continuously generate growing numbers of candidate proofs, mathematics may increasingly face a different environment. The scarce resource is no longer only the answer. It becomes the ability to verify results, determine which ones matter, and decide which questions are worth pursuing.
In other words, AI does not eliminate scarcity in research. It moves the bottleneck:
As generation becomes abundant, verification, selection, and judgment become more important.
The same pattern appears differently at large scientific facilities.
X-Ray Imaging: When Data Arrives Faster Than It Can Be Interpreted
Laurent Chapon, Associate Laboratory Director for Photon Sciences at Argonne National Laboratory (ANL), showed how AI is being applied to advanced X-ray imaging.
Large light-source and neutron facilities operated by U.S. national laboratories can produce extremely bright beams for examining the internal structure of materials, including reconstructing complex three-dimensional circuitry inside semiconductor chips without destroying the sample.
But the detector does not simply produce a photograph.
It records complex diffraction data that requires substantial computation before researchers can reconstruct a scientifically meaningful three-dimensional image.
Chapon said facilities of this kind can generate data volumes equivalent to roughly 150 million movies per year. That creates a striking imbalance in the research workflow:
An experiment may take minutes or hours, while data processing and image reconstruction can take days or weeks.
AI is compressing that gap.
Chapon showed AI models used for image reconstruction. As X-rays scan a sample, the models can accelerate the conversion of detector data into three-dimensional images, bringing processes that once took hours or days closer to real time. In related tasks, he said, the improvement can reach at least 100-fold.
The result is not simply a faster image.
If researchers can see reconstructed results while an experiment is still underway, they can identify anomalies earlier, adjust experimental conditions, and decide where to look next. A workflow that once looked more like:
experiment → data processing → result
can begin to move toward:
experiment ↔ real-time analysis ↔ next experiment
Once image reconstruction no longer requires days of waiting, another bottleneck emerges: Which anomalies in a continuous stream of information are actually worth pursuing?
The Electric Grid: From a Few Scenarios to Millions of Possibilities
Hendrik Hamann, Chief AI Scientist for Innovation, Science and Security at Brookhaven National Laboratory (BNL), presented another kind of computational bottleneck: the U.S. electric grid.
Hamann described a system with roughly 200 million access points, 50 million transformers, 10 million energy resources, and 7 million miles of wires. Electricity demand, generation, weather, storage, and a growing number of distributed energy resources are changing simultaneously and interacting across the system.
Operating and planning a system at this scale requires working across multiple time horizons: balancing electricity in real time, anticipating demand hours or days ahead, and determining how new generation, data centers, and other major loads can connect years into the future.
Traditional physics-based models can simulate the grid. But as the number of variables increases, the possible scenarios can reach into the billions, while actual planning can examine only a small fraction of them.
The constraint is not the absence of a model. It is the inability to calculate enough possible futures within a useful decision-making timeframe.
Hamann described a grid foundation model being developed for Genesis that learns relationships among grid topology, demand, and power flows, enabling much faster evaluation of possible scenarios. In some tasks, he said, the model has achieved improvements of up to 1,000 times and can evaluate millions of scenarios. A data-center interconnection study that might previously have taken months or even years can be compressed dramatically.
Here again, the role of AI changes.
Mathematics confronts an enormous proof-search space. Advanced imaging confronts a processing gap between the rate at which experiments produce data and the rate at which researchers can interpret it. The electric grid confronts an enormous number of possible system states.
As AI expands the number of futures that can be calculated and simulated, the bottleneck moves again:
Being able to calculate more possible futures does not automatically determine which future should be chosen.
Once simulation capacity expands, turning predictions into planning and decisions becomes the next problem.
National Security: From Digital Design to the Physical World
Sivasankaran Rajamanickam, Distinguished Member of the Technical Staff at Sandia National Laboratories (SNL), took the discussion further into design, manufacturing, and physical testing.
His team has been working on a flight-test vehicle used for high-speed research. Such systems must withstand complex thermal, vibration, and structural loads, and traditional development can involve years of design, simulation, manufacturing, sensor integration, and testing.
The team sought to connect AI-driven topology optimization with advanced manufacturing, scientific software, data, and high-performance computing capabilities accumulated across the national laboratory system.
Given engineering objectives and constraints, the system explored a large design space and generated different structural options. Rajamanickam showed 16 candidate designs, from which researchers selected one for manufacturing.
The selected structure was not the kind of design engineers would necessarily have arrived at intuitively.
That also illustrates the boundary of AI capability. AI can expand the design space and generate options that people might not otherwise consider. But generating a structure does not mean it can be manufactured — or that it will work in the physical world.
The workflow therefore involves several distinct stages:
AI explores the design space → researchers select a solution → advanced manufacturing turns the digital design into a physical object → sensors and physical testing validate the result.
The team also used AI in payload and sensor-layout design and combined it with advanced manufacturing methods to shorten production cycles. Rajamanickam said the project moved from requirements to its first production unit in approximately three months, and to its first flight test in approximately eight months.
At this point, AI is no longer simply analyzing scientific data.
It is entering the cycle connecting design, manufacturing, experiment, and validation.
What AI for Science Is Actually Changing
Mathematics, advanced imaging, the electric grid, and national security are not examples of the same AI application.
Mathematics faces proof generation and verification. Advanced imaging faces the processing gap between experimental data and image reconstruction. Grid research faces a simulation space too large to exhaust. Engineering design faces the cycle from digital search to manufacturing and physical validation.
Yet the four cases reveal a common structure:
AI expands a part of the scientific workflow that was previously constrained — and, in doing so, pushes the bottleneck to the next stage.
As generation expands, verification becomes more important. As experimental data can be processed in real time, researchers must decide what is worth investigating. As simulation spaces expand, planners must determine which scenarios matter. As design spaces expand, proposed solutions still have to be manufactured, tested, and validated in the physical world.
AI for Science, then, is not simply about giving scientists a more powerful general-purpose AI tool.
It is beginning to change parts of the scientific process itself: how researchers search, analyze, simulate, and design — and how those digital capabilities reconnect with experiments and the physical world.
That also helps explain why the Genesis Mission extends beyond AI models themselves.
From AI Models to an “Internet of Science”
Gil described the system envisioned by the Genesis Mission as an “Internet of Science.”
What must be connected is not only AI models, but scientific data, supercomputing capacity, AI agents, and physical research infrastructure — including light sources, particle accelerators, microscopes, advanced manufacturing capabilities, and other scientific facilities.
The four cases presented during the Technical Keynote show different pieces of that connection already taking shape. Mathematics connects AI generation with formal verification. Advanced imaging connects AI with data produced by major scientific facilities. Grid research connects models with high-performance computing and complex physical systems. Engineering design extends the connection into manufacturing and real-world testing.
This helps explain why the Genesis Mission involves not only AI companies, but federal agencies, national laboratories, universities, researchers, and industry partners.
For the first 278 selected teams, the next question is therefore no longer simply:
Can AI accelerate scientific discovery?
A more practical question follows:
How will researchers across different institutions actually find, access, use, and share the data, models, computing resources, AI agents, and scientific facilities distributed across this network?
The answer will help determine whether the “Internet of Science” becomes a collection of individual AI research projects — or a shared research infrastructure that can operate as a connected system.
