Arena Physica Trains Heaviside on Lux Supercomputer

The electromagnetism foundation model will be trained on AMD-powered hardware at Oak Ridge National Laboratory.

Updated on Nov. 12, 2026 in Artificial Intelligence

Isometric-style editorial illustration of matte server cabinets in a long hall with overhead cooling ducts, rendered in slate, cream, and teal, evoking large-scale AI computing infrastructure.
Arena Physica will train its Heaviside electromagnetism model, a 350 million parameter system built on 250,000 3D field simulations, on the Lux AI supercomputer at Oak Ridge National Laboratory. AI Illustration. Upload story photo >

Arena Physica announced it will train the next generation of its Heaviside model on the Lux AI supercomputer. Heaviside-1, released in September 2026, is a 350 million parameter model trained on 250,000 3D field simulations.

Why it matters

Heaviside is described as the first foundation model for electromagnetism, and scaling its performance requires substantial compute. Training on Lux, which is powered by AMD hardware, reflects the growing demand for large-scale computing resources in specialized AI science applications.

Heaviside-1 contains 350 million parameters and was trained on 500 billion individual field samples generated by solving Maxwell's equations across detailed 3D models. Performance gains from the next generation have not been quantified, and results will be shared as training on Lux progresses.

The players

Arena Physica

Arena Physica is the company developing the Heaviside family of electromagnetism foundation models.

Lux AI supercomputer

Lux is an AI supercomputer located at Oak Ridge National Laboratory and powered by AMD hardware.

Oak Ridge National Laboratory

Oak Ridge National Laboratory is a U.S. Department of Energy research facility that hosts the Lux supercomputer.

AMD

AMD is a semiconductor company whose hardware powers the Lux AI supercomputer.

The details

Heaviside learns from full-wave electromagnetic simulations, with each sample solving Maxwell's equations across a detailed 3D model of a structure. The next generations are projected to extend to far-field radiation patterns, more complex geometries, and wider frequency ranges.

Timeline

  1. September 2026: Heaviside-1 model released.

  2. November 12, 2026: Arena Physica announced its Lux training partnership with AMD.

The Tech Race

Heaviside-1, released in September 2026, is the first foundation model for electromagnetism, and the planned training on Lux extends that work toward far-field radiation patterns and wider frequency ranges. The move places Arena Physica's scaling effort on AMD hardware at a national laboratory rather than on a conventional commercial cloud.

For most readers, the immediate effect is indirect, since Heaviside is aimed at engineers who design electromagnetic hardware rather than consumers. Faster simulation tools built on models like Heaviside could shorten development cycles for devices that use wireless and radio-frequency components.

The takeaway

Foundation models trained on physics simulations may reduce the time needed to design electromagnetic components, and readers following AI infrastructure can track results as Arena Physica shares training progress on Lux.

Further reading

Explore more Artificial Intelligence coverage to see how foundation models are being applied to scientific domains.

Source note: This article includes information reported by HPCwire.