DOE Awards $159M to Bring Scientific AI Into Fusion, Chip Design, and Quantum Research

A fusion experiment, a geothermal reservoir, and a chip built to survive intense radiation have something in common: each can require researchers to explore an enormous space of possible designs before finding one that works. The U.S. Department of Energy’s latest Genesis Mission awards aim to put AI inside that search process, helping scientists decide what to simulate, build, and test next.

DOE announced $159 million for 12 new Phase II projects on October 8, 2026, alongside six additional Phase I awards. The department says the mission’s first-year cohort is now complete at 297 projects. Its broader Phase II portfolio contains 14 projects, including GridFM 2.0 and Prometheus, supported through the offices of Electricity and Nuclear Energy.

The scope is unusually broad. The selected work stretches from fusion and geothermal energy to RNA structures, rare-earth separation, scientific software, and quantum computing. What connects the projects is a practical ambition: use AI to shorten the distance between an idea and a scientifically useful result.

What DOE means by “Super Intelligence”

DOE describes the initiative using the term “Super Intelligence,” abbreviated SI. That terminology warrants care. The announcement establishes a research portfolio and its intended applications; it does not present evidence that a general-purpose AI system has surpassed human capability across science.

The initial Genesis project selections announced in July describe a mix of advanced models, agent frameworks, scientific software, and high-performance computing. That is a useful way to understand the new awards: as investments in AI-enabled research systems, with capabilities tailored to particular disciplines and infrastructure.

There is an important distinction between generating a plausible scientific answer and completing a reliable scientific workflow. A model may propose a material, but its properties still need measurement. An agent may generate simulation code, but the code must preserve the equations and assumptions researchers intended. An optimizer may recommend a machine setting, but it must respect operating limits.

The interesting question is therefore how these projects connect prediction and reasoning to evidence. That connection will determine whether the investment produces better science rather than simply more scientific-looking output.

Fusion: a digital twin for planning SPARC operations

Commonwealth Fusion Systems is developing an AI-enabled digital twin to simulate and optimize SPARC operations. A digital twin is a computational representation of a physical system, useful for exploring how that system might behave under different conditions.

The company already has a relevant technology foundation. In January, CFS announced work with Siemens and NVIDIA combining engineering tools with simulation and visualization technologies. That earlier collaboration provides context for the new award, although DOE’s announcement does not establish that every component of the January architecture is part of the funded project.

CFS explains in its technical account of AI and simulation that surrogate models can learn from computationally expensive physics simulations and approximate their outputs more quickly. Such models can help researchers explore many candidate operating conditions without rerunning the most demanding calculation for each one.

For a tokamak, which uses magnetic fields to confine hot plasma, this creates a potentially valuable planning tool. Researchers can examine candidate scenarios, identify promising experiments, and compare model predictions with measurements. The difficult part is knowing where a fast approximation remains trustworthy and where the full physics calculation is needed.

SPARC is a fusion demonstration machine, with CFS targeting net fusion energy in 2027. That target concerns fusion performance; it should not be read as a claim that SPARC is already supplying electricity to the grid. The near-term significance of the AI project is its potential to improve experimental planning and operation.

Geothermal and rare earths: using AI to navigate complex physical systems

UC Irvine’s MAESTRO project takes a multi-agent approach to mapping, stimulating, and managing deep geothermal reservoirs. The name expands to Multi-agent AI Expert for Subsurface Reasoning and Optimization.

The geological challenge is substantial. DOE’s enhanced geothermal systems overview explains that a productive geothermal resource needs heat, fluid, and permeability. Where natural conditions are insufficient, engineering can create or improve pathways that allow fluid to circulate through hot rock and carry heat toward the surface.

That background helps explain why an AI system could be useful. Reservoir decisions involve imperfect subsurface information and interacting physical processes. A multi-agent architecture could divide analytical work among specialized components, but DOE’s brief announcement does not specify MAESTRO’s full implementation or autonomy limits.

Our assessment is that its value will depend on how well the combined system handles uncertainty. Producing a confident reservoir description is less useful than identifying what is known, what remains uncertain, and which measurement would most improve the next operational decision.

The University of Illinois’ AXIS project applies AI to electrochemical processes for extracting rare-earth elements from domestic sources. Its focus on redox-active ligand design points toward the chemistry governing selective separation.

The underlying design problem is different from geothermal engineering, but the logic is similar: search a complicated space of candidates while respecting physical constraints. For separation technology, meaningful evaluation should extend beyond whether a candidate looks promising in a model. Selectivity, recovery, energy consumption, and performance in realistic mixtures would be useful measures of progress. Those are proposed evaluation criteria, not results reported in the award announcement.

Chips for environments ordinary electronics cannot tolerate

Fermilab’s AXESS project targets microelectronics for high-radiation, space, and extremely cold environments. Its name, Accelerating Extreme Environment Specs-to-Silicon, captures the ambition to move from engineering requirements to chip designs more quickly.

Fermilab’s earlier description of the initiative explains why these devices are difficult to develop. Scientific instruments can require custom electronics that operate under conditions outside the design assumptions of mainstream commercial chips. The laboratory describes using AI to accelerate a traditionally demanding design process.

The October announcement places AXESS within the new Genesis funding. The opportunity is to automate more of the exploration and checking involved in specialized chip design, drawing on expertise across laboratories, universities, and industry.

Design speed and device readiness are separate milestones. Even a dramatically faster design workflow still needs verification, fabrication, and physical testing. A chip intended for a particle detector or cryogenic system must demonstrate that its behavior survives the conditions specified for it. AI-generated design files alone cannot establish that.

Quantum research: making hardware useful and designing materials backward

Harvard’s ASQC project combines AI with quantum hardware to address error correction through application-aware co-design. In practical terms, co-design means considering the application, error-correcting approach, and hardware together rather than treating each as an isolated problem.

Quantum devices are sensitive to errors. Error-correcting schemes encode information across physical qubits so that a logical qubit can be more reliable. Harvard’s earlier logical quantum processor research provides background on that distinction and the effort to run computations using encoded qubits.

The new award should not be interpreted as an announcement that fault-tolerant quantum computing has been completed. Instead, it targets a difficult optimization problem: matching scientific tasks to the resources and error behavior of real hardware.

Oak Ridge National Laboratory’s selected project addresses another quantum frontier: physics-informed inverse design of quantum magnets for sensing and low-power electronics. Inverse design starts with a desired behavior and searches for a system that could produce it, reversing the familiar process of choosing a material and then calculating its properties.

For this kind of research, physics-informed modeling matters because an attractive numerical solution may be physically unattainable. A useful design must ultimately connect to materials that can be made and measured. The announcement sets that research direction; it does not report a finished sensor or commercial electronic device.

The less visible infrastructure behind scientific AI

Several other awards concern the systems that make large-scale research possible. MIT’s project connects AI and supercomputing to lattice quantum chromodynamics, a computational approach to studying the strong interaction. Lawrence Berkeley National Laboratory’s MOAT-Core develops a shared AI platform for accelerator design and operation. Northeastern’s project applies agentic assistance to the Facility for Rare Isotope Beams.

Argonne’s AI4HPC effort includes developing, optimizing, and verifying scientific computing software. That verification component is consequential: a program that runs faster but silently changes the scientific calculation is not an improvement.

Biology is also represented. UC San Diego aims to expand an RNA structure database fivefold, while the University of Washington will work on enzyme design through improved sequence–structure ensemble modeling.

These projects highlight two dependencies often overshadowed by model announcements: trustworthy data and dependable tools. A scientific agent’s usefulness is limited by what it can access and execute. Connecting agents to simulation engines, curated datasets, and instruments can make them more capable, while increasing the need to track their actions and validate their outputs.

As our coverage of Anthropic’s support for the Genesis Mission illustrates, the initiative also involves bringing commercial AI capabilities into the scientific ecosystem. The funded projects will need to establish how those capabilities translate into measurable research gains.

Two separate investments in the researchers who will use these tools

DOE also announced a Genesis Mission Graduate Fellowship Pilot. It plans up to $100 million, including $20 million in fiscal 2027, with later funding contingent on congressional appropriations. The initial cohort is expected to support approximately 100 doctoral fellows through four-year pathways combining AI and a scientific discipline.

The planned placements include both a DOE laboratory or user facility and an industry setting. That structure could help researchers understand the gap between developing an AI method and making it useful in an experimental or operational environment.

A separate Early Career Research Program announcement covers $96 million for 67 scientists, with university awards of about $875,000 and laboratory awards of $2.75 million over five years. Its scope includes AI, quantum science, and fusion, among other research areas. DOE notes that award selections remain subject to negotiations.

These amounts should not be collapsed into a single fully awarded Genesis package. The $159 million concerns the new Phase II projects; the fellowship has a conditional future funding plan; and early-career support is a separate program.

What would count as success?

The strongest test of this portfolio will be whether AI improves scientific work under realistic conditions. That means measuring the time and resources needed to reach a validated result, the quality of the result, and the system’s ability to recognize when its own methods are unreliable.

A useful fusion model should help plan better experiments. A chip-design agent should produce designs that pass verification and physical testing. A geothermal system should support decisions that remain sound when new field measurements arrive. A scientific coding tool should preserve correctness as it improves performance.

The Genesis awards create room to test those propositions across very different disciplines. Their significance lies in putting AI closer to the machinery, data, and constraints of science—and giving researchers the resources to find out where it actually helps.