The White House just made American dominance of scientific foundation models a top federal R&D priority. This is the most important scientific policymaking choice of the Trump Administration. Most of America’s hardest technical and scientific problems, from the production of safe, next-generation nuclear reactors to accessible hypersonic flight, will be solved by physics models that harness the power of artificial intelligence to exponentially accelerate discovery.
For most of the modern era, the pace of scientific and engineering progress has been set by how fast we can verify an idea against physical reality. Prototypes take years. Laboratory experiments take months. High-fidelity physics simulation takes hours to days. A full aerodynamic database for one flight vehicle can consume months of supercomputer time and an entire team of experts, without any ability to use these calculations to accelerate the pace of problem-solving for the next vehicle.
On July 21, OSTP Director Michael Kratsios and OMB Director Russell Vought made a multi-billion-dollar bet. To ensure the next generation of scientific breakthroughs happens at home, the U.S. government will prioritize building the models necessary to exponentially accelerate discovery and prototyping. I commend them for it.
This reinvigoration of the Endless Frontier mantra is the clearest statement yet from any government that the same class of technology now transforming language and biology can transform our command of the physical world. If every U.S. government agency actually allocates a portion of its resources to collecting and curating the data, and building the physics models necessary to solve their hardest problems, America will win.
I spent my career building simulation and optimization methods for aerospace systems, at Stanford’s Aerospace Design Laboratory, as director of NASA’s Fundamental Aeronautics Program, and now at Luminary, America’s only frontier lab focused on building physics AI models. The models and datasets that we are building at Luminary are based on high-fidelity physical data (rather than text, code, images, and/or video) to create the foundational models necessary for accelerated scientific breakthroughs. In collaboration with America’s leading government agencies, we are positioning the United States to lead the next generation of physical discovery and engineering prowess.
But make no mistake. Labs in China and Europe are also competing to build the next generation of foundational physics models. The Chinese Academy of Sciences released its ScienceOne scientific foundation model in July 2025 and expanded it this April into a system of eight domain-specific models spanning physics, mathematics, and aerospace. The models are already delivering meaningful engineering results: Huawei’s MindSpore Flow includes an airfoil flow-field model, built with COMAC, that returns predictions at 1/24 the cost of traditional CFD.
What the memo gets right
The memo directs agencies to propose investments in “domain-specific scientific foundation models that enable high-fidelity simulations of natural phenomena.” At Luminary we call this class of AI model a Large Physics Model (LPM): a model trained on large corpora of high-fidelity simulation, experimental, and operational data that learns the mapping from geometry and operating conditions to the full solution field and the ultimate performance of, say, a new weapon system. It is the scientific foundation model the memo describes, built for the physics that govern engineered systems. A conventional physics solver integrates the governing equations from scratch for every case: change the geometry or the boundary conditions and you pay the full computational price again. A trained LPM returns the flow field, thermal field, or structural response in a single forward pass, in less than a second, at every point on the geometry, and with quantified uncertainties. The economics invert: compute is invested once up front, in training, and amortized across every question the model answers afterward, possibly hundreds of millions of times for both direct and inverse problems. The memo also notes that these models “require curated scientific datasets and compute that no performer can assemble alone,” which is why federal coordination matters and why they should be built as shared, pre-competitive assets.
The memo pairs this ambition with discipline. It instructs agencies to fund research that uses AI “as a new instrument of scientific discovery, not merely as a tool to augment existing capabilities,” and to reject proposals “that apply AI for incremental gains or without clear justification for why the problem requires AI-specific methods.” This discipline is both welcome and absolutely necessary. The field does not need another round of hyped-up projects that wrap a regression model around an existing workflow and call it AI. The test should be whether the method changes what questions a scientist can afford to ask and what outcomes are derived.
The memo also calls on agencies to make internal scientific datasets AI-ready and to capture data “at an industrial scale,” including “experimental records, negative results, and operational data” that are routinely abandoned. The federal government sits on some of the most valuable physics training data ever produced: decades of wind-tunnel campaigns, flight-test archives, and national-laboratory simulation output, most of it trapped in formats and silos that make reuse impractical. Curating those archives into training corpora may be the highest-return R&D investment available to several agencies, because the data is already paid for. Continuously accumulating curated data to inform the training of new models and the re-training of existing ones is absolutely vital to our physics AI supremacy.
The most forward-looking directive asks agencies to invest in “autonomous control of large-scale experiments in which AI systems generate hypotheses, conduct experiments, interpret results, and iterate in real time.” In computational science this loop is already buildable. A trained model reports where its confidence is low or where an input falls outside its training distribution; those flags automatically commission new high-fidelity solver runs; the new data retrains the model. The system decides what to simulate next. As needed, the same system can direct the carrying out of targeted experiments to most effectively improve the predictive quality of the models and to manage the remaining risks and uncertainties. What today takes a committee of experts months of judgment becomes a continuous process that runs overnight.
Underneath all of this, the memo prioritizes foundational research in “the thermal, fluid, and mechanical sciences” and in “the modeling, simulation, and verification on which fusion energy, advanced manufacturing, and space systems depend.” These and others are the fundamental disciplines of physics that determine whether hardware works. They have declined as a share of the federal research portfolio for an extended period, as the memo notes, and the correction is overdue.
Who this impacts
Six agencies clear the memo’s $3 billion R&D threshold on OMB’s federal crosscut: the Departments of War, Energy, Commerce, and Agriculture, plus NASA and NSF. Each has a concrete place to start.
The Department of Energy’s Fusion Science & Technology Roadmap targets the demonstration of commercial fusion power by the mid-2030s. Divertor and plasma-facing-component design is a coupled thermal-structural-fluid problem where each candidate configuration is expensive to evaluate; an LPM trained across a broad design space would allow engineers to sweep through thousands of configurations before committing to hardware, and would flag the regimes where high-fidelity simulation alone will suffice, or where additional experiments and simulations will be needed. On the fission side, advanced reactor developers and their regulators repeat the same classes of core and coolant thermal-hydraulics analyses on every design; a shared model trained on national-laboratory simulation archives would compress both design iterations and the analytical basis for licensing review. No agency is better positioned to lead: DOE owns the leadership-class computing and decades of simulation output the memo says these models require.
At the Department of War, hypersonic vehicle development is gated by aerothermal predictions: boundary-layer transition, surface heating, and thermal-protection-system sizing, where every flight condition is its own expensive computation and flight test costs make broad exploration prohibitive. A large physics model spanning the Mach-altitude-attitude envelope, including control surface deflections, gives design teams multi-physics, envelope-wide answers in seconds and reserves ground and flight test for the highest-uncertainty operating conditions that constrain the design. The same holds for store separation and munitions-integration databases that today are rebuilt program by program, combination by combination.
The memo commits NASA to returning Americans to the lunar surface by 2028, a lunar base, and space nuclear power. Entry, descent, and landing analysis today means building an aerodynamic database point by point across Mach number, altitude, and attitude, with thermo-chemical non-equilibrium effects. An LPM covering that envelope turns database generation from a months-long campaign into a simple set of queries, and every mission thereafter inherits the model. Space nuclear systems add reactor thermal management in vacuum and partial gravity, a regime where experimental data will always be scarce and simulation-trained models carry most of the load.
Someone has to fund the science of these models themselves, and that work belongs to the National Science Foundation: approximation theory and error bounds for neural operators, physics-constrained training, uncertainty quantification that regulators and flight-clearance authorities will accept, and the open benchmark datasets that let the field measure progress the way shared corpora did for language modeling. The memo’s push for agile, mid-scale science through models like NSF’s X-Labs fits this need, which sits awkwardly between single-investigator grants and national facilities.
The Department of Commerce holds two distinct cases. The memo’s semiconductor mission names EUV-and-beyond photolithography and 3D advanced packaging. Thermal and mechanical behavior of stacked die, packages, boards, racks, and data centers is now a first-order design constraint, and NIST’s metrology mission generates exactly the characterization data such models train on and rely on. At NOAA, the case is already proven at planetary scale: published results in Science in 2023 showed a model trained on decades of re-analysis data outperforming the leading operational medium-range weather forecast on a large majority of verification targets. The question for NOAA is not whether the method works but how fast to operationalize it across its modeling portfolio, as well as how to make the predictions and the uncertainty estimates more accurate.
And USDA may be the agency most surprised to receive this memo, and that is precisely the opportunity. Agricultural engineering is dense with thermal-fluid problems: airflow and energy management in controlled-environment agriculture, post-harvest drying and cold-chain design, spray transport and drift. A focused investment in an LPM for one of these problems would be a credible, differentiated action plan.
In every case the pattern is the same: the agency owns the problem and much of the data; the methods exist; simulation data can now be generated at scale. What is missing is the decision to build a family of models as an institutional capability rather than rerun the analysis forever without learning from those one-off investments.
Where Luminary stands
This is the work Luminary was founded to do and that American frontier AI labs must focus on. We are the only American Physics AI company building the data, training infrastructure, and models necessary to solve these problems. We supply a platform to build, deploy, and operationalize Large Physics Models trained on simulation, experimental, and operational data, spanning fluid, thermal, structural, acoustic, and electromagnetic physics. Our models report uncertainty and detect out-of-distribution inputs, so a scientist knows when to trust a prediction and when to verify it. And we deploy where national security work actually happens: in the cloud, on-premises, and in air-gapped, ITAR-compliant environments.
The memo asks the right question of every agency: which scientific problems in your mission could a domain-specific Large Physics Model unlock? With our existing capabilities and the vast data sources available to these agencies, transformational science and engineering can take place. Luminary stands ready to partner with the agencies drafting these plans. To the teams working against the 90-day clock: we would welcome the conversation.
Statement from Juan J. Alonso, Co-founder and CTO of Luminary; Professor and Chair, Aeronautics and Astronautics Department, Stanford University
