Arena Physica to Train Heaviside on AMD-Powered Lux Supercomputer

Arena Physica said in its announcement on November 12, 2026, that it will work with AMD to train future generations of its Heaviside electromagnetism model on the Lux supercomputer at Oak Ridge National Laboratory. The project builds on Heaviside-1, a 350-million-parameter model released in September 2026.

Arena Physica describes Heaviside as the first foundation model for electromagnetism and reports that the initial version was trained on 250,000 three-dimensional field simulations spanning 500 billion individual field samples. Those figures establish the baseline for the AMD work, which is a planned training program rather than the release of a Lux-trained successor.

Simulation Data Drives Compute Demand

Each training sample comes from a full-wave electromagnetic simulation, with Maxwell’s equations solved across a detailed 3D model of a structure. Producing the training corpus is therefore a compute-intensive part of model development, alongside the compute needed for the training run itself.

“These fields are highly sensitive to the details of a design, from small local features to the structure as a whole. Capturing that complexity takes a lot of high-fidelity data and a purpose-built model architecture,” said Michael Frei, chief research officer at Arena Physica.

Arena Physica connected that requirement to Practical Scaling Laws, a paper by Christopher M. Bryant and Hao Liu. The paper argues that the scaling laws guiding modern model training were calibrated for data-rich, single-epoch pretraining and identifies structural limitations outside that regime. Its proposed formulation decomposes loss into undercapacity, undertraining and overfitting terms while distinguishing unique examples from the total examples seen.

Once calibrated, the form supports a cost-aware allocation that shifts toward smaller corpora and more epochs as data becomes expensive. Applied to Heaviside, that framework presents an infrastructure trade-off between generating additional unique simulations and making repeated training passes over existing data.

Lux Supplies Large-Scale Training Capacity

The Oak Ridge Leadership Computing Facility’s Lux specification lists more than 4,000 AMD MI355X GPUs, 1.1 petabytes of high-bandwidth memory and 20 exaFLOPS of FP8 AI-training performance. Each MI355X has 288 GB of HBM3E memory and 8 TB/s of bandwidth. Allocations also include access to Orion, OLCF’s Lustre file system, with more than 600 petabytes of capacity.

Those figures describe Lux’s capacity, not a measured Heaviside speedup. Their relevance spans both sides of the workload: accelerator throughput for training, and memory and storage for the simulation corpus that feeds it.

Planned Models Target Chips and Communications

Arena Physica plans for later Heaviside generations to cover far-field radiation patterns, more complex geometries and wider frequency ranges, scaling up to larger components and full systems. Target applications include antennas for 5G, 6G and satellite communications, automotive radio-frequency sensing, and semiconductor packaging and signal integrity.

For semiconductors, the stated uses include high-speed SerDes design. Kumaran Siva, AMD’s corporate vice president of AI enterprise strategy, connected the work to 3D chip packaging and SerDes speeds climbing from 224 to 448 gigabits per second, saying those trends have made electromagnetic behavior one of the hardest problems in semiconductor design.

Arena Physica says Heaviside is designed to predict electromagnetic behavior orders of magnitude faster and more cheaply than traditional solvers and to enable autonomous inverse design. These are objectives for the model’s development, not performance measurements from the Lux collaboration.

Arena Physica plans to share results as Lux training progresses in the months ahead. Those reports will provide the substantive evidence for assessing whether Lux-scale training can deliver the wider frequency, geometry and full-system coverage the company has outlined.