MIT researchers use Ferveret cooling platform to advance AI work on nuclear materials

AI research may hit a physical limit long before it runs out of ideas: the amount of processing a lab can power and keep cool. Ferveret says MIT researchers are using its high-performance computing platform to develop artificial intelligence methods for studying nuclear materials, using its Adaptive Phase Cooling technology to work on sustained scientific workloads.

The October 22 announcement links two engineering challenges. Researchers need greater computational power to understand materials exposed to extreme conditions. Data center operators must squeeze more useful work out of limited electrical capacity. Ferveret’s proposal is that better heat removal can help both, without requiring a new generation of processors.

Why nuclear materials research needs more AI-based calculations

The work involves MIT’s Materials in Extreme Environments laboratory, led by Ericmoore Jossou, an assistant professor of nuclear science and engineering and electrical and computer engineering. The MIT faculty profile describes research that combines advanced microscopy and coherent X-ray imaging to examine structural properties and defects at nanoscale resolution.

Ferveret identifies two projects using its platform. One develops artificial intelligence approaches to analyze coherent X-ray measurements of nuclear materials. The other trains machine learning-based interatomic potentials, models that predict how atoms interact and help researchers simulate the behavior of materials under challenging conditions.

There is independent support for the research direction of the first project. MIT’s 2026 Nuclear Science and Engineering Research Expo lists a presentation by DongHun Ryu titled “Physically-Based Deep Generative Models for BCDI Phase Recovery.” BCDI stands for Bragg Coherent Diffraction Imaging, a technique also identified in Jossou’s faculty research description. This establishes a concrete search link beyond the cooling company’s announcement, although the event listing doesn’t rate Ferveret’s hardware.

Interatomic potentials face a different computational bottleneck. Instead of performing an expensive electronic structure calculation at each step of the simulation, a trained model approximates the relationship between atomic arrangements, energies, and forces. Research held in the MIT archive describes how machine learning force fields can extend the time and length scales of atomistic simulations while maintaining the accuracy of the benchmark calculations on which they are trained.

This accuracy remains conditioned by the model and the situations it has learned to represent. Ferveret’s announcement states that the funded research aims for simulations that approach the accuracy of density functional theory; does not provide an error table, published model, or numerical comparison that establishes this outcome for these projects.

How adaptive phase cooling removes heat

Ferveret’s description of the technology focuses on subcooled boiling, an approach inspired by research on the heat transfer of nuclear reactors. The heat on the surface of the chip forms vapor bubbles, but the surrounding liquid remains below boiling temperature. The company says small bubbles quickly break away and condense back into that cooler liquid, refreshing contact with the hot surface.

The difference is where the steam goes. In the saturated boiling immersion approach, Ferveret compares bubbles rising in a vapor space before condensing. His approach seeks to condense them locally within the liquid. The company says this supports ambient pressure operation and compact enclosures serving single servers.

The data center’s product page describes a 4U rack-mounted cooling unit called Lambert, a custom heat exchanger, and control software that regulates operating conditions. The intended implementation is a server-scale cooling system connected to the facility’s heat removal infrastructure, rather than a large shared soak tank.

For AI operators, the goal is to keep processors within thermal limits during extended workloads while reducing cooling costs. Cooler hardware can sustain higher operational performance when heat would otherwise limit it. The size of the advantage still depends on the processors, workload, power settings and cooling system used for comparison.

What the 15% and 35% performance figures actually mean.

The two main figures in the ad describe different measurements. Ferveret reports up to 15% more server-level compute compared to leading direct-to-chip cooling approaches, and up to 35% more usable compute when fabric-level improvements are included.

A March announcement from Ferveret provides more details on the first figure. Describes benchmarking with UCLA’s Intelligent Connectivity Laboratory using NVIDIA H200 GPUs, reporting a 15 percent improvement in computational efficiency measured in TFLOPs per kilowatt of server power. The release explicitly excludes facility-level power usage effectiveness from that server parameter.

This is helpful context, but it remains a company-issued account of the collaboration. The release does not provide a complete benchmark protocol or sufficient raw results to reproduce the comparison. Nor does a 15% improvement in floating point throughput per watt automatically establish a 15% improvement in training time or application output.

The larger figure combines server efficiency with a structure model. In Ferveret’s worked example, the comparison moves from a power usage effectiveness, or PUE, of 1.20 to 1.03, then applies the 15% server efficiency improvement. PUE is the total power of the facility divided by the power of the IT equipment; a lower ratio means less facility overhead for a given IT load.

Using the indicated inputs, the combined gain is approximately 34%: 1.15 multiplied by 1.20/1.03. Ferveret describes the result as about 35%. The arithmetic explains the scope of the request, but it’s a scenario based on the plant’s assumed efficiency, rather than a published measurement of a 35% speedup in MIT’s nuclear materials experiments.

Where trials end and opportunities begin

In the announcement provided, Jossou states that access to the platform allows for larger models, faster iteration, and continuous experiments. Such observations support the practical value of further access to computation. They don’t isolate how much improvement came from cooling, how much from hardware allocation, or how much faster a particular scientific task was performed.

The same distinction applies to sustainability. Ferveret describes a closed-loop system capable of rejecting heat through dry coolers without consuming water through evaporation. This is a statement about the cooling configuration. It should not be interpreted as eliminating the water footprint of electricity generation, equipment manufacturing, or any possible facility design.

The announcement also states that future research publications are expected to recognize Ferveret’s infrastructure contribution. Published methods and application-level results would help determine how the platform impacts reconstruction accuracy, simulation scale, training time, and energy per completed experiment.

For now, the strongest conclusion is the pointed one: Ferveret is providing infrastructure for a research program whose focus on artificial intelligence and nuclear materials is confirmed by MIT, while its broader efficiency claims require attention to measurement limits. The long-term opportunity is to further exploit scarce computing power, turning improvements in heat transfer into more scientific work within the same electricity budget.