FoundaRF reuses AI training to speed up radio-frequency component design

FoundaRF addresses a practical bottleneck in radio-frequency and millimeter-wave chip design: predicting how a component’s physical layout will affect electrical signals. The researchers report improved prediction accuracy, reduced training-data requirements and faster automated design in tests spanning four semiconductor technologies.

The work was posted as an arXiv preprint on November 6, 2026. Its listing states that it was accepted for the proceedings of the 2026 IEEE/ACM International Conference on Computer-Aided Design. The results are the authors’ reported computational findings, not an independent assessment of industrial deployment.

Why reusing a model matters

Passive components, including inductors and transformers, help filter, match and condition signals. At high frequencies, their geometry and surrounding metal structures strongly influence their behavior. Designers therefore use electromagnetic solvers to evaluate layouts, a process the paper says can take minutes to hours for a single structure.

Machine-learning substitutes, or surrogate models, can make predictions faster. But a model trained for one component family or manufacturing process may work poorly on another. Building a replacement then requires more simulations to supply labeled training examples: layouts paired with their calculated electrical behavior.

FoundaRF aims to reuse both geometric knowledge and knowledge learned from those electrical responses.

Learning geometry before electrical behavior

The framework first converts chip-layout files into normalized images representing three important metal layers: a primary layer, a secondary layer and a ground or shielding layer. It keeps physical dimensions, connection details and other design information separately, rather than expecting an image alone to convey everything.

An encoder—the network section that extracts useful features—initially learns from layouts without electromagnetic simulation labels. Its training task is to reconstruct a distance field: a map showing each location’s distance from the nearest conductor boundary.

This encourages attention to features such as conductor spacing, widths and edges. Those features matter to electromagnetic behavior, giving the training objective a physical motivation. It does not mean that this stage calculates the full electromagnetic fields.

A decoder subsequently learns to predict broadband scattering parameters, usually called S-parameters. These describe how signals entering a component’s connection points are reflected or transmitted across frequencies. They provide a shared prediction target for components that otherwise need different simplified circuit descriptions.

A conditioning mechanism called FiLM adjusts the decoder’s internal features using the separately stored dimensions and design details. Small, task-specific output modules accommodate different connection counts and frequency ranges. For transfer to new tasks, the researchers reuse pretrained network sections and adapt the model with target-domain examples.

What the tests showed

The datasets covered transformers in indium phosphide and silicon-on-insulator technologies, T-coils in silicon-germanium, and inductors in gallium nitride. Electrical reference values came from electromagnetic simulations.

In layout-processing comparisons, the final method reduced mean absolute prediction error by up to a factor of 3.3 relative to the best tested single-input baseline. That maximum occurred for a parallel-transformer dataset in silicon-on-insulator technology; it was not the improvement achieved in every test.

For transfer experiments, the researchers pretrained on silicon-on-insulator and indium-phosphide data, then adapted to silicon-germanium and gallium-nitride components. Transferring both encoder and decoder worked better than transferring the encoder alone, supporting the importance of learning electrical responses as well as geometry.

In one gallium-nitride comparison, 25 labeled examples delivered accuracy comparable to training from scratch with 100 examples. This is the basis for the reported fourfold reduction in data requirements, rather than a universal saving across all tasks.

The team also used FoundaRF inside an automated search for two T-coil networks: a driver output and a receiver input protection network. The search balanced layout area, reflection and transmission. On four CPU cores, the authors report convergence within ten minutes and describe a speed advantage exceeding 100-fold over electromagnetic-simulation-in-the-loop optimization. That timing concerns optimization, not the entire process of generating training data and training the model. Final designs were checked with electromagnetic simulations.

Where the evidence stops

The demonstrations do not establish measured performance in fabricated chips. Nor do four technologies establish generality across all passive components: the study concentrates on magnetically coupled structures, while transmission-line networks and couplers remain proposed extensions.

The representation also simplifies the metal stack to three principal layers rather than modeling full three-dimensional structures. Physics-informed training provides useful guidance, but the reported tests do not establish that every prediction necessarily obeys all physical constraints.

Finally, the authors explicitly describe FoundaRF as a foundation-model framework, not a claim that its compact 5.2-million-parameter network is itself a large-scale foundation model. Its demonstrated contribution is more specific: reusable prediction machinery that can reduce simulation demands within the component families tested.