Introducing Atlas RF Studio: Toward a Foundation Model for Electromagnetics

Every electronic signal you've ever sent or received traveled through circuitry whose geometry was painstakingly shaped to control electromagnetic fields. This kind of design underpins cell phones, satellite communications, radar, power delivery, high-speed data transfer, and the interconnects inside chips themselves. And the demands keep growing. Data rates double with each new generation of communication technology. Phased arrays are scaling from dozens of elements to thousands for 5G/6G and defense radar.

As clock speeds rise and wavelengths shrink, this design problem shifts from circuit topology (what connects to what) to physical geometry (the shape and layout of the materials themselves). At low frequencies, a wire is just a wire: all that matters is that it connects point A to point B. But when we enter the radio frequency (RF) regime where wavelengths approach the scale of the physical structures that make up the circuit (e.g., the trace width and spacing, the via diameter, the layer stackup, even the solder ball radius), everything starts exhibiting electromagnetic behavior that can't be ignored. A wire becomes a transmission line. A bend becomes a reflector. Two parallel traces become coupled antennas. The geometry is the circuit.

While the infrastructure of the modern world runs on electromagnetics, we're now faced with two major issues:

1. The workforce hasn't kept up. The set of professionals capable of designing high-speed RF circuitry has been shrinking for years. It takes a decade to develop this intuition, and the trend toward offshoring has made U.S. expertise even scarcer.

2. The tools haven't kept up. Today's electromagnetic simulators brute-force Maxwell's equations with numerical solvers that are slow, often prohibitively so. Worse, they learn nothing from past work. Institutional knowledge lives in textbooks and the minds of a few experts. None of it is captured in the tools themselves.

Imagine a world where electromagnetic design expertise was abundant and fast. An RF engineer could explore thousands of candidate designs in a single morning instead of spending weeks refining just one. A startup could design phased arrays that previously required defense-contractor-scale teams. Unwanted cross-talk between high-speed data center signal channels could be eliminated in seconds. Wireless charging could work at nontrivial distances. Beaming power from orbital solar arrays to the ground could become efficient enough to be useful. These applications aren't science fiction. They're just constrained today by the cost and scarcity of good electromagnetic design.

At Arena Physica, we believe AI can fundamentally change this, but an LLM alone can't get us there. This requires models that have internalized how fields propagate, reflect, couple, and interfere—models grounded in real physics, trained on massive amounts of electromagnetic data entirely unlike anything any LLM has seen before. What we're describing is a foundation model not for text, but for electromagnetics.

Today, we're sharing the first concrete step on that journey—the beta release of Atlas RF Studio. It's an interactive sandbox for AI-driven inverse RF design. Given specifications or requirements, Atlas RF Studio uses agentic workflows to generate, simulate, and iterate on candidate designs. The models behind it, Heaviside-0 and Marconi-0, are "step zero" on a much longer roadmap—this public sandbox is a preview of deeper capabilities under development we're excited to share soon.

The Electromagnetic Design Problem

At its core, every electromagnetic design problem is the same: given a structure's geometry, its material composition, and how it's excited, predict or control what happens to the electromagnetic fields. A filter passes certain frequency bands and rejects others. An antenna radiates energy in specific directions. A transmission line delivers power from one point to another with minimal loss.

There are two problems here, and they run in opposite directions.

1. The forward problem is characterization: given a physical structure, predict its electromagnetic behavior. What S-parameters does this filter produce? What radiation pattern does this antenna create? Commercial field solvers like HFSS solve the forward problem by brute-force numerical methods, and a single forward pass can take minutes to hours, sometimes even days.

2. The inverse problem is design: given a desired electromagnetic behavior, find a physical structure that produces it. I need a filter that passes this band and rejects that one. What geometry will do it? Today, engineers solve this through iteration: guess, simulate, adjust, repeat.

Both problems are bottlenecked by the same thing. An expert starts with a design from a textbook or past experience, runs a forward simulation, waits, looks at the results, adjusts a parameter, and runs it again. A single design might require dozens or hundreds of iterations, depending on the complexity of the structure and the desired accuracy.

This is not just slow. It is fundamentally limited by human intuition. Every existing simulator is a deductive tool: it applies Maxwell's equations to a specific structure and computes the result. The general understanding, the intuition for why a geometry behaves a certain way and what to try next, still comes entirely from the engineer. A designer can only explore a tiny fraction of the possible design space. Consider a simple two-layer circuit where each pixel in a 64x64 grid on the signal layer can be either metal or dielectric. That's already $2^{64 \times 64} \approx 10^{1233}$ possible configurations for a single, small component. The entire history of human RF design has explored a vanishingly small fraction of this space. What becomes achievable if we could expand that exploration by orders of magnitude?

We're building something different: a model that reasons inductively, learning general electromagnetic principles from millions of examples. That learned understanding is what lets it explore designs no engineer would think to try.

Why a Foundation Model?

The history of machine learning offers a useful analogy. Before large language models, the NLP field was dominated by specialized models: one for translation, one for sentiment analysis, one for spam detection, one for summarization. Each was trained on task-specific data and could only do its one thing.

Foundation models changed this. By training on enough data at enough scale, language models began to exhibit emergent capabilities. They didn't just memorize patterns; they learned the structure of language itself. A single model could translate, summarize, answer questions, and write code, because it had developed a general enough understanding of the underlying primitive: language.

We believe the same scaling laws and generalization dynamics apply to electromagnetics. The primitive here is not words but electromagnetic fields and the geometries, materials, and excitation conditions that shape them. A model trained on enough field data, across enough designs, should begin to internalize the physics of electromagnetism at a level that enables genuinely new capabilities.

This is not a surrogate model. The distinction matters. Today's ML-for-physics landscape is full of surrogate models: fast approximations of a specific simulator for a specific class of problems. Kriging-based surrogates for on-chip spiral inductor optimization. Surrogates for microwave filter design. CNNs trained to predict patch antenna S11 from pixelated geometries. Neural network surrogates for multi-port RF passives scoped to specific structure families and frequency ranges. Each is useful in its lane, but none "understands" electromagnetism. This is the same pattern that dominated pre-LLM NLP, where every task got its own bespoke model.

The core limitation of surrogates isn't any single factor. It's the combination of narrow datasets, task-specific architectures, and small scale that keeps them from learning general physics. And the training signal matters enormously. Take the S-parameter prediction problem for example. S-parameters are a reduced description of the system: they tell you what a structure does at its ports, but not how or why. The geometry-to-S-parameter mapping is many-to-one. Many different structures can produce identical S-parameters, which means a model can find shortcut rules that fit the training data without learning the actual physics. Within the training distribution, these shortcuts work. Outside it, they can fail, not by gracefully degrading, but by producing confidently wrong answers. The model's internal rules have no physical basis to generalize from. This is the ceiling of S-parameter-only training: it works until it doesn't, and the boundary is unclear.

S-parameter-focused modeling follows the same logic that produced pre-LLM natural language processing: constrain the problem, narrow the use case, optimize within the box. But for us, fields are the fundamental quantities. They are what Maxwell's equations actually govern. Training on fields themselves forces the model to learn the physics that produces S-parameters, rather than learning to approximate the mapping directly. S-parameters, radiation patterns, current distributions, and every other quantity of interest all follow from the fields. A model that understands fields understands the cause—everything else is effect.

This is the thesis behind our foundation model approach. We are not building a collection of surrogates. We are building a single foundation that learns electromagnetism from the ground up—an inductive model with a deep enough command of the physics to accurately characterize brand new designs, not just interpolate between designs it has already seen.

The Data Factory

Training a foundation model requires data at a scale, diversity, and quality that has never existed for electromagnetics. While there has been some recent effort to build collections of large-scale physics datasets like The Well, the domain of electromagnetics has largely been left behind. There is no "internet of EM data" we can draw from, so we had to build it ourselves. Our "data factory" produces two categories of training data: simulated and measured.

Simulated Data

The bulk of our training data comes from simulation: we design structures, then simulate their behavior with traditional slow but accurate field solvers. This process starts with expert-created design templates: canonical RF designs authored by experienced engineers who have decades of institutional knowledge about how these systems behave. Our Chief Scientist Harish Krishnaswamy and Distinguished Engineer Arun Natarajan are among the experts whose knowledge informs this work. These templates are the building blocks of real-world RF design: filters, couplers, matching networks, and more.

Each template is a seed. Through procedural generation, we create thousands of variations from a single template, varying dimensions, proportions, and parameters while preserving the essential design intent. This is how a handful of expert designs becomes millions of training examples.

Each variation is a plausible design that a human engineer might explore, anchored to the physics that makes the template work. Including this data ensures that even before our model learns to characterize wildly unfamiliar designs, it has enough understanding to be useful for a wide range of real-world applications.

But a dataset can't just be big—it has to be diverse if we want to encourage generalization. A model trained only on structures that look like real designs will learn the statistics of those designs without necessarily learning the physics that makes them work. So alongside the expert templates and their procedural variants, we include a large collection of random designs: geometries with no design intent behind them, sampled broadly through organic procedural generation across the space of possible geometric and material configurations. These designs force the model to confront electromagnetic behavior in regions of the design space that no expert would visit, building a broader physical understanding from which the expert-seeded knowledge can generalize.

The models in our first release are trained on 3 million simulated designs across 25 expert templates mixed with random structures, amounting to a total of over 20 years of combined simulation time. We're aiming to get to 100 years in the next few months, and 1,000 years by the end of 2026.

Measured Data

Simulation can scale almost infinitely with compute, but it will never fully capture what happens when a design meets reality—manufacturing tolerances, material variations, connector parasitics, measurement coupling—all the complex effects that push real-world behavior outside the scope of what a traditional idealized simulator could ever capture. Closing the sim-to-real gap means doing the physical work: design a structure, send it out for fabrication, wait for it to come back, calibrate a vector network analyzer, connect it with precision fixturing, measure S-parameters across a wide frequency range, and record the results in a format the training pipeline can ingest.

By fabricating and measuring new designs continuously in our lab, we can steer our model closer to reality. Traditional simulators can model known physical effects like surface roughness and process variations, but each solve starts from scratch. They learn nothing from past work. A learning system can continuously incorporate real-world data in a way that numerical solvers fundamentally cannot, capturing and compounding the knowledge that is currently trapped out there in the real world.

The Models

Atlas RF Studio is powered by two models that mirror the two fundamental problems of electromagnetic engineering. Both develop an understanding of how geometry and electromagnetic behavior relate, but they apply this understanding in opposite directions.

• We call the forward model Heaviside-0, after Oliver Heaviside, who reformulated Maxwell's equations into the compact vector form that made modern electromagnetic analysis possible. Heaviside was the great characterizer. He gave engineers the mathematical language to describe what fields do. Our forward model carries that legacy. Given a circuit layout, it predicts S-parameters in milliseconds rather than the significant compute time required by a numerical solver. This is what makes rapid design iteration possible.

• We call the inverse model Marconi-0, after Guglielmo Marconi, who invented and built the first practical radio systems through relentless experimentation, turning desired wireless behavior into physical reality. Our inverse model does the same thing. Given a target S-parameter response, it generates a physical geometry that produces it. This is the creative half, the one that proposes designs a human might never think to try.

These two models work in tandem. Marconi-0 proposes a collection of designs that it thinks will produce your target behavior, and Heaviside-0 quickly checks whether the proposed designs exhibit that behavior. Based on that feedback, Marconi-0 can then iteratively propose new designs that are more likely to produce the desired behavior, and Heaviside-0 can then characterize them again to see if they are any better.

With these processes combined in a tight feedback loop, the system can rapidly optimize the design for the target. Both speed and quality are critical in this optimization loop. The faster the models are, the more iterations can be run in a given amount of time. The better the models are, the more effective each individual iteration is in achieving the target.

Evaluation

Evaluating models like Heaviside-0 and Marconi-0 is an open problem. These are the first models with these particular capabilities, so there are no established baselines or benchmark suites to draw from. We can't point to a leaderboard and say "we're state of the art," because the leaderboard doesn't exist yet.

When people ask "why can't I just get ChatGPT to do this?"—we understand the impulse. Frontier LLMs are good at so many things that it is natural to assume they might handle this too. But generating precise numerical answers for electromagnetic tasks is far outside the scope of what LLMs have been trained to do, so we should not expect them to perform well.

To verify this claim empirically, we set up a small evaluation using 100 designs drawn randomly from our full validation set of 30,000 designs (held out during model training). We assessed both Anthropic's and OpenAI's frontier models against the forward and inverse tasks on those 100 designs, testing each in default and extended thinking modes. We tuned our prompts to get the models to perform as well as we could on these tasks.

They don't come close. Across both the forward task (predicting S-parameters from geometry) and the inverse task (generating geometry from target S-parameters), frontier LLMs are both slower and substantially worse than our purpose-built models. See below for more details.

Forward: Heaviside-0

Quality

We evaluated the forward model's ability to predict S-parameters by tracking 3 key metrics: weighted-MAE of the magnitude, MAE of the phase, and RMSE of the combined real and imaginary parts of the S-parameters. Since it is infeasible to manufacture all test cases for validation, we measure performance relative to a commercial solver.

What's immediately clear from these results is that frontier LLMs are incapable of achieving the performance of Heaviside-0, and Heaviside-0 achieves a magnitude weighted-MAE well under 1 dB (for context, the range that RF engineers typically care about spans roughly 20-30 dB, so <1dB is a very strong result). It is interesting that extended thinking does allow the frontier LLMs to perform slightly better than their base behavior, but the performance gap to our model is still substantial.

Speed

Our forward model's primary advantage over alternative approaches is speed.

On these examples, a commercial numerical field solver takes roughly 4 minutes per example. In contrast, Heaviside-0 produces S-parameter predictions for an input board in 13 milliseconds. Even more, since our model leverages GPU parallelism, when given a batch of 1024 boards, it can generate predictions at an average of 0.3 ms per board. This is an 18,000x to 800,000x speedup over a traditional numerical solver.

With extended thinking, the frontier LLMs are even slower than a solver (and not nearly as accurate), and burn through so many tokens that they are effectively unusable for this task.

A speedup like that doesn't just save time. It transforms what's possible. Having realtime feedback on a design change is unprecedented in this field. An engineer who could previously evaluate a hundred designs in a workday can now explore a million. Previously inaccessible regions of the design space become reachable.

Inspecting the Mind of the Model

Heaviside-0 is a transformer-based neural network, trained primarily to understand how different geometries and materials produce different S-parameters. As the inputs pass through the layers of the model architecture, they are transformed into increasingly abstract representations that capture features of the input data essential for the task at hand. To gain some intuition for what those representations look like, we used UMAP to probe the network right before it makes its final S-parameter predictions:

In some ways, the model organizes the space similarly to how we might as humans, often clustering designs from the same template family together. But in other ways, it doesn't. Two designs that look nearly identical—differing only slightly in geometry or material properties—can end up in completely different regions of the embedding space because those small changes dramatically alter the electromagnetic behavior. Conversely, designs that look nothing alike may cluster together because they produce similar behavior. This disconnect between appearance and electromagnetic function is precisely what makes RF design hard for humans, and it's precisely what the model can learn to see past. Stay tuned for a future post diving much deeper into this topic.

Generalization from fields

Today's model is trained mostly on S-parameter data, mixed with a smaller amount of field data. However, even with the small amount of field data we used, we've seen that fields encourage the model to learn representations that improve S-parameter prediction. To quantify this effect, we ran a small ablation study. Our base dataset contains 3 million designs labeled with S-parameters. We then augmented this with a supplementary dataset of full-wave field data for 10,000 of these designs—just 0.3% of all the designs in the total dataset.

To evaluate, we held out two design templates entirely for "out-of-distribution" validation (the model never saw these templates during training), and randomly split the remaining designs into training and "in-distribution" validation sets. You can think of in-distribution validation as interpolation (predicting the behavior of new designs from familiar template families) and out-of-distribution validation as extrapolation (predicting the behavior of designs from template families the model has never seen).

On extrapolation, field data acted as a regularizer, reducing overfitting and modestly improving out-of-distribution performance:

On interpolation, the effect was far more dramatic—a 15% performance improvement:

Even though we observed larger benefit here on interpolation, the extrapolation task is worth discussing in more detail. A common criticism of ML models is that they can only remix what they've already seen—that they'll fail on anything truly new. And in a narrow sense, that's true: if you train a model purely on S-parameters, a brand new structure with unfamiliar S-parameter behavior is out-of-distribution, and the model has no recourse.

But fields may offer a way out. The S-parameters of a novel structure may be unlike anything the model has seen, but the physics that produces them is not new. Fields behave according to the same Maxwell's equations regardless of the structure's geometry. A model that has learned how fields work doesn't need to have seen a similar S-parameter curve before—because S-parameters are physically derived from fields, if it has learned an internal representation that allows it to predict fields, that representation should allow it to predict S-parameters as well. Learning fields turns S-parameter extrapolation into something closer to an in-distribution task.

We're seeing the earliest signal of this with just 0.3% field coverage. As we scale up the ratio of field data in future releases, we expect these effects—particularly the extrapolation benefit—to become substantially more pronounced.

Inverse: Marconi-0

Quality

In contrast to the forward problem, the inverse problem has no solver we can use as a baseline. Traditional solvers are fundamentally incapable of inverse design. However, we can still evaluate our model against the generative capabilities of frontier LLMs. To do this, we gave each model a target S-parameter response from the validation set and asked it to generate a circuit layout that produces it. We then evaluated the quality of the generated designs by running them through a commercial simulator to obtain their actual S-parameters, then tracked the same metrics as we did for the forward model. This answers the question "how well does each model generate designs that match the target S-parameter response?"

Recall that this evaluation is still constrained to the relatively simple problem of 2-layer designs that can be represented by a single 2D image and some metadata. We anticipate that the discrepancy between frontier LLM performance and Marconi will only widen when we move toward more complex 3D designs. Even if the LLMs somehow have strong enough reasoning capabilities to mathematically manipulate this complex structured data (which they likely don't), the physics of the problem will require increasingly complex calculations that render the problem effectively intractable for any language-based LLM.

Speed

At our current data and training scale, even though Marconi-0 does not immediately remove the need for human intuition in the design loop, the iteration speed it unlocks will accelerate the design process by orders of magnitude.

For reference, a good design with non-trivial EM behavior can take an experienced RF engineer weeks to design by hand. This is because each candidate design must be simulated with a traditional numerical solver, the solver is exceedingly slow, and very few candidate designs an engineer creates will meet the desired S-parameter behavior exactly.

Marconi-0 cuts down on that time by generating many decent-quality candidates in parallel, validates them all with Heaviside-0 to rank adherence to the target S-parameters, and returns the best candidate to the user (or even multiple candidates matching the target behavior, if desired).

Similar to how LLMs have a quality-vs-latency trade-off introduced by the amount of "thinking" they do, Marconi-0 can "think" for longer by searching over a larger space of candidates before responding. In addition to generating more candidates, Marconi-0 can perform iterative refinement: starting from a promising candidate, it spawns variations that attempt to close the remaining gap to the target S-parameters. Breadth finds good starting points and depth hones them.

From Target to Design

Marconi-0 is a generative model trained to produce designs that create a target S-parameter response. It works through a conditional diffusion process similar to modern image generation networks, starting from an unstructured canvas and progressively resolving it into a coherent design, guided by the target specifications.

If we watch the model's intermediate outputs as it works, we can see how a design emerges from pure noise:

Alien Structures

Marconi-0 can be conditioned to generate designs consistent with a specific template, but it is not fundamentally constrained by conventional design patterns. Perhaps the most exciting outputs are the designs that don't look like anything a human would draw. We've been experimenting with pushing the generative model into unfamiliar territory, conditioning on target responses and allowing it to explore geometries far from the expert-designed templates.

While many of the generated candidates are too far outside of the familiar design space for Heaviside-0 to accurately characterize, some of these structures actually work. They meet their target S-parameter specifications despite having very alien-looking geometries. These structures are precursors to the kinds of designs that a scaled foundation model could produce routinely: geometries that exploit electromagnetic phenomena in ways human intuition would never suggest.

The Road Ahead

While Atlas RF Studio is our first public release, it's just a waypoint in our much larger mission—to create electromagnetic superintelligence. Today's release supports two-layer, 8mm x 8mm structures with ground vias and 3 dielectric choices. We chose this as our starting point because it encompasses a wide range of practical RF components while keeping the design space tractable enough for rigorous validation. But this is just the beginning. Here's where we're headed:

• Scaling the design space. We are actively expanding to more layers (4+), larger dimensions relative to the characteristic wavelength, more complex via structures, higher frequencies, and a wider range of dielectric and conductor materials. Each of these axes increases the space of designs the models can handle and brings Atlas RF Studio closer to the full complexity of production circuit design.

• Scaling up field-based training. Heaviside-0 already incorporates field data in training, but S-parameters still dominate the loss signal today. The full electromagnetic field distribution contains far richer information. We believe this is the key that unlocks real out-of-distribution generalization, and we've seen enough in early field-based experiments to know the direction is correct. The next step is scale: more field data, larger models, and longer training runs to let the physics fully saturate the representation.

• Shrinking to silicon. The physics of our models extends naturally from PCB structures down to integrated circuits. The core challenge is a training problem (different geometries, materials, and scales), not a fundamentally different modeling problem. We are actively working toward taping out our first AI-designed silicon in 2026.

• Enabling broader EM characterization. Beyond expanding structure complexity, we plan to expand our forward prediction capabilities from S-parameters to a wider space of electromagnetic phenomena, including near-field characterization, antenna radiation patterns, and current distributions.

• Moving toward multiphysics. Real-world electromagnetic systems don't exist in isolation. Temperature shifts material properties; mechanical stress distorts geometry. Our approach starts with electromagnetics, but the same foundation model principles extend naturally toward multiphysics—reasoning about the interactions between fundamental forces that real designs must contend with.

• Closing the loop with reality. As our fabrication pipeline scales, we will continuously incorporate real-world measurement data back into training. This creates a flywheel: better models produce more interesting designs, which get fabricated and measured, which produce higher-quality training data, which improves the models. We will also explore approaches to fuse sparse sensor data (e.g. measurements at only a few points) into the forward model at inference time. This will allow us to steer dense predictions (e.g. field values everywhere in space) to be consistent with the measurements we collect from real devices.

• Expanding from components to systems. Today's release targets individual components: filters, couplers, matching structures, etc. But real RF systems are assemblies of many interacting components, and the most interesting problems live at the system level. We will integrate our EM foundation models into larger simulation and design workflows, so that component-level EM intelligence can plug into full system-level analysis—bridging the gap between electromagnetic design and the broader electronic design chain.

A Foundation Model for Electromagnetic Superintelligence

If we can build a system that genuinely understands electromagnetics, one that reasons inductively about fields and geometries, the implications run deep.

Intuition for everyone. Real-time forward and inverse design capabilities could collapse the learning curve for the "dark art" of RF design. When you can tweak a geometry and see the electromagnetic impact immediately, cause and effect become visceral, not opaque. The intuition that used to require years of trial and error to build could be acquired in a fraction of the time.

Superhuman design. Beyond increasing accessibility, we arrive at the frontier: designs that surpass human capability, harnessing electromagnetic phenomena in ways that even our best experts have no intuition for. Many breakthrough applications are not limited by the physics, but by our ability to design structures that properly exploit that physics.

The path to electromagnetic superintelligence is long, but the signal is clear. We invite the RF community to try out our first step along this path, Atlas RF Studio in beta—push it to its limits, and tell us what you'd build with a foundation model for electromagnetics.

Come on this journey with us.