RapidRF at MWJ RF & Microwave Summit: AI in RF - May 2026

Ajouté : 2026-05-23

[00:00:01]Hello everyone. It's a real pleasure to be back at the microwave journal RF and microwave summit and especially here at this AI in RF session which feels more relevant today than ever before. My name is Edward. I'm co-founder and CTO of Rapid RF, a company we started because we generally believe we are at an inflection point in how RF circuits get designed. Rapid RF builds AI native tools for automated RFIC design. Our goal is is straightforward. Make the artificial design engineer reality.

[00:00:37]Fully automated from specs to table and free engineers to focus on what truly matters. That's a bold claim. So before I tell you where we are going, let me show you what we have already built.

[00:00:54]And I don't say that lightly. Since 2022, we have been doing exactly that.

[00:01:00]Designing real circuits on real processes with real silicon coming back from the lab.

[00:01:07]We started with something ambitious. The first fully automated terraertz tape out an LNA and a power amplifier in silicon germanium designed by our system measured in the lab working in 2022. As three, we closed our first commercial partnership which told us that what we were building had real industrial value.

[00:01:30]Mid 2024 is where things got serious. We ran a direct industrial benchmark against expert RF designers at one of our partners. Our system designed gas LNAS each in under uh a day. Human experts needed over 10 days each. Same targets, better performance, 10 times faster with our AI. January 2025, the first ever state-of-the-art K and Ka band power amplifiers on gun were developed with the help of AI. Over 10 watts of output power, power added efficiency above uh 40% each designed in one or two days.

[00:02:10]By March 2025, our partner was running the platform fully autonomously. driver amplifiers designed without our engineers in the loop. And mid 2025, we launched our online platform live accessible to our customers generating tape ready LNAS and drivers uh in gas in under one hour which brings us to today and what I'm about to show you uh is the next step. So where does it leave us? We have proven that AI can design RF circuits, real ones, on real processes, faster than human experts and with competitive performance. But if you look at this slide, the left column is what exists. The right column is what doesn't. No one today has a fully autonomous loop, natural language to tape out without a human intervening at every stage that does not exist yet. And that gap is not a small one. It's a difference between a powerful tool and a truly autonomous design system.

[00:03:17]Now to be fair, LLMs are already doing useful things in EDI environments. They can set up simulation configurations, organize workspaces, generate plots, and navigate tool interfaces. They can read net lists, interpret schematics to a degree, and even generate simple schematics from explicit instructions.

[00:03:38]But here's where it breaks down. Ask an L&M to design a specific ARF circuit, a real one with targeted specs, and it will fail. Every time this has been demonstrated repeatedly, it gets worse for layout. And even the most promising agentic workflows that do handle layout and EM simulation. They rely on template knowledge or very explicit step-by-step instructions. They are not reasoning from physics because LLMs operate at the language level. They orchestrate, they describe, they navigate, but they do not understand what happens inside a transistor at 28 GHz. For a truly agentic design system, there is a layer that is still missing. A layer that operates directly in the ARF domain at the physics level. That missing layer is what we built. Let me tell you a story that made headlines earlier this year. A man in Australia, not a scientist and not a doctor, used ChetPut to develop a personalized cancer vaccine for his dog and it worked. Now the headline gives all these credit to Chetchip. But if you look at what actually happened, ChapT planned the workflow. It searched a literature. It translated complex biology into actionable steps. That part is genuinely impressive. But the actual heavy lifting predicting the three-dimensional protein structure of the tumor anti-gene that was alpha fold a specialized AI that operates not at the language level but the physical and molecular level an AI that encodes deep domain knowledge that no language model can replicate. Without alpha fold chatpt's plan would have gone absolutely nowhere. And that distinction matters because it maps perfectly onto RF design. LLMs are extraordinary at reasoning, planning, and orchestrating.

[00:05:35]But they speak in tokens. They have no physical intuition. They cannot tell you what happens inside a transistor at the quantum level or how a matching network behaves at millimeter wave frequencies.

[00:05:47]RF design like protein folding needs a domain AI, a layer that works directly in the physics that doesn't describe circuits but designs them. So what makes RF design so fundamentally different?

[00:06:02]Look at this biasing, stability, current density, noise figure, impedance matching, nonlinear behavior, layout parasitics, fabrication tolerances, EM coupling. These are all not concepts that you look up. They are all constraints you feel simultaneously in every decision you make. An experienced RF engineer doesn't think about these at one at a time. They hold all of them in their head at once and they navigate that space through years of earned intuition. That is exactly what a language model cannot do. It is stoastic by nature. It has no simulation ground truth. It cannot guarantee convergence.

[00:06:43]And if you ask uh it to design a circuit and the result is wrong. There's no physical telling uh physics telling it why. You simply cannot fix wiped circuits. Physics doesn't care how confident the model was. Circuit simulation is the ground truth. Always has been. And any AI that wants to design our circuits seriously has to build on that foundation not around it.

[00:07:09]We built that AI. Please let me introduce you to Heavy Mind, named after Olivia Heavyside, the engineer who gave us the mathematical foundations of transmission line theory, the man who made modern RF engineering possible in the first place. We thought the name was fitting. Heavymind is our answer to the missing layer. It is not a chatbot. It is not a co-pilot. It is not a wrapper around any existing EDA tool. Heavymind is a domain AI that operates directly in the RF design space. It encodes the physical intuition that RF engineers spend years developing and it applies that intuition to generate and optimize circuits from the ground up. The key insight behind heavymind is this generation and optimization are not two separate problems. They are one and simulation is not a verification step that happens afterwards. It is embedded inside every single in iteration. That is what separates Heavymind from every agentic workflow you have seen so far.

[00:08:13]It doesn't orchestrate existing tools.

[00:08:15]It reasons in the RF domain and it builds classical circuit simulation as it is physical ground truth from design space exploration to design space intuition.

[00:08:29]Before we dive deeper into heavy mind, let me show you prime examples of current approaches in research. Design space exploration treats RF design problem as a discrete landscape navigated step by step, one decision at a time. Each state represents a specific circuit configuration and each action moves to an adjacent state, adjusting a bias point, scaling a transistor or shifting an impedance.

[00:08:58]A path through this design space is built incrementally, guided by how close each step brings the design to its targets. Reinforcement learning is a prime example of an AI approach that operates exactly like this. Many potential best fit designs go undiscovered. The AI simply never explores regions of the design space that could satisfy the targets. These unknown states may seem redundant at first glance, but they are precisely what a model needs for deep circuit intuition. Reinforcement learning can navigate under specified problems. But it pays a steep price. Long training times, brittle generations in failure as soon as a design problem shifts even slightly from what it was trained on.

[00:09:44]L&Ms, for example, offer telling contrast. They acquired emergent generalizable behavior through self-supervised learning on data alone.

[00:09:53]Reinforcement learning enters only at the very end as a fine-tuning layer, not the foundation. The implication is clear. We want RF design AI that truly generalizes. The bulk of the learning must come from the data itself.

[00:10:10]The second approach is a template burst circuit design. The architecture is fixed up front and automation is reduced to parameter tuning within a predefined topology.

[00:10:20]In the performance complexity space, this corresponds to a purely vertical search. The structural dimension is never explored and entire regions of potentially optimal design regions remain out of reach. A two-stage LNA stays a two-stage LNA. It cannot evolve into a combined driver or power amplifier stage. What is commonly used here are surrogate based transistor and component models to accelerate the optimization loop. These can be effective within their narrow scope. But it would be a stretch to call this AIdriven design. The AI has no agency over the design itself. It only tunes what the human architect already decided.

[00:11:04]This is where heavy mind changes the picture entirely. Rather than following discrete steps through a predefined topology, Heavy Mind builds a continuous intuition over the full design space.

[00:11:17]Every region is reachable and every point is a candidate.

[00:11:23]We are no longer constrained to vertical search within a fixed architecture.

[00:11:27]Heavy mind can move freely across the structural dimension navigating through toward the parita front rather than stopping at the nearest local optimum within the template. Architecture limits are overcome and RF design becomes what it always should have been a pure optimization problem.

[00:11:47]At the core of heavy mind is a closed loop architecture built around the three elements.

[00:11:54]the AI optimizer, a set of intuition blocks, and a highfidelity simulation environment. Heavy mind drives the loop.

[00:12:02]At each step, it proposes an explicit circuit architecture and component instance. Not just parameters, but structural decisions. The intuition blocks guide this process, encoding learn prior over the design space to focus exploration where it matters. The simulation environment in the center operates with fast highfidelity solvers.

[00:12:22]For nonlinear problems, AI assisted models from the RDK are brought in directly, keeping simulation speed high without sacrificing accuracy.

[00:12:33]Third party or open source solvers can be attached through Python interfaces, keeping the platform open and integratable into existing workflows.

[00:12:42]And the loop doesn't stop at the physical circuit model. EM co- simulation can be embedded directly into the layout optimization stage, enabling high density designs that maintain top-notch RF performance without manual back and forth between simulation domains.

[00:13:00]Every generation and optimization step is tracked and fed back to heavy as learnable signal. The system continuously refineses its own intuition from the designs it explores.

[00:13:13]Let's go with this example of a multistage amplifier. You start the design at one stage. Optimize. Target met. No, just add another stage.

[00:13:23]Optimize again. Target met. Yes.

[00:13:26]Amplifier is ready. Every stage added to the amplifier is optimized with heavy minds intuition domain at arbitrary complexity without architectural constraints. There is no ceiling on what the loop can produce. This is what makes RF design truly algorithmic. And once the design is algorithmic, it becomes the foundation for agentic AI. A system that does not just assist but drives the design process end to end. The same logic applies beyond amplifiers.

[00:13:55]Switches, filters, VCOs, any circuit type follows the same principle. The flow is universal. Customers will be able to define their own flows and inject their own design knowledge turning heavy mind into a platform that grows with their expertise. The result, RF design becomes an optimization problem from the very first decision.

[00:14:18]What you're seeing here is a live generation run. It shows a stageby-stage construction of a multi-stage amplifier.

[00:14:25]Each stage is added and optimized sequentially within Heavy Mind's intuition domain and then afterwards together. Following that, the area optimization DRC violations are resolved and the layout is compacted into a dense clean floor plan. This was not done in a week, not in a day. This entire workflow from blank canvas to compact DSC clean performance verified layout in under 30 minutes. The performance plots before and after they are nearly identical.

[00:14:55]This might seemed counterintuitive to experienced RF engineers because layout aware design exists for a good reason, but both passes are driven by the same heavy mind simply running a different optimization objective. Performance first then area. Heavy mind handles both and knows how to preserve one while solving for the other.

[00:15:18]Heavy mind is not limited to a fixed set of circuit types. The design space it navigates is defined by the problem not by the tool. Today the platform covers LNAS and driver amplifiers on gas. Our AI can navigate any design space circuit topologies component instances fractal and pixelated networks. The only boundary is a problem itself and we have a concrete announcement to make here.

[00:15:43]Gun HPS will be released at IMS 2026 in Boston in the coming weeks. Beyond amplifiers, the roadmap extensive switches, filters, VCOs, mixers, and custom topologies. Any circuit type that can be defined as a design space heavy mind can learn to navigate.

[00:16:03]RF circuits will no longer be defined by handdrawn schematics or topology files.

[00:16:08]They will be described at the system level exactly the way it they appear in a textbooks. An RF switch, a splitter at the input, two branches each with a transistor block followed by a matching network. A do amplifier, a splitter, carrier and peaking transistor pairs, individual matching networks and a power combiner at the output. These are precisely the highlevel representations that large language models reason in natively. Heavy mind closes the loop between that highle intent and fully optimized layout ready design. And just to give you a sense of where this is already heading. These block diagrams were not drawn manually. They were generated by code by simply asking it for the block diagrams of an RF switch or doy amplifier. The agentic workflow is not a vision. It is already here.

[00:17:01]AIdriven design automation does not mean the engineer loses control. It means the engineer chooses how much control to hand over and when to take it back.

[00:17:11]Today, Rapid RF operates at a fully automated pipeline. Specs go in, a DSC clean performance verified layout comes out in under an hour. They blackbox delivers. The next steps opens that black box. After each generation stage, the engineer will be able to inspect the parto optimal design candidates, selecting the tradeoff point that matches their priorities at the schematic layout and EM level transparency without sacrificing speed.

[00:17:41]And further down the road, the engineer will be have full control over optimization goals, constraint weights, and design priorities, turning heavy mind into a precision instrument that amplifies the engineer's own expertise.

[00:17:54]The AI is never the ceiling. It is a foundation. At IMS, we are also introducing Circuit Studio, the pipelinedriven design environment that brings everything together. Circuit studio gives you a clear overview of your current design project at every stage. After generation, you pick from multiple Parto optimal candidates to find the circuit with the right performance complexity tradeoff for your needs. From there, manual adaptions are possible directly in our user-friendly lay layout editor with instant simulation feedback on every change.

[00:18:28]A circuit level optimization step provides further parameter tuning to push performance beyond the initial AI generation. A system compliance check ensures your circuit will be a stable at the operational level with spy circuitry accounted for. The AIdriven layout optimization then compacts your design and an optional automated bias routting finalizes the layout. The engineer can intervene at every step of the process.

[00:18:55]Every optimization run every heavy mind performs can be become part of your own proprietary model. Your proven architectural patterns, your design knowledge, your best performing circuits fed back into heavy mind as structural priors that improve every future run.

[00:19:12]This means customers can define their own automation flows for any circuit type, build their own intuition models for VCOs, switches, filters or anything else and benchmark their architectures directly against alternatives inside the optimizer. Your IP stays exclusively yours within your environment and for organizations where data sovereignity is non-negotiable, Heavymind will be available on premise announced at IMS 2026. the full capability of the platform running entirely within your own infrastructure. Your architecture knowledge encoded into AI, a competitive edge that grows with everyone and belongs entirely to you. And there's more coming soon. RF design automation is only as good as the model it builds on. Transistor models are the foundation of every simulation. And if they are inaccur inaccurate, no amount of AI optimization on top will fix it. This is where Dante comes in. Dante is our n AI native transistor modeling tool. It takes measured device data sparameters, IV curves, noise and fits a physically precise simulation ready model fully automated. The results plug directly into heavy mind simulation environment.

[00:20:28]Measured data to calibrated model to optimize circuit. One AI native stack end to end. This is what Dant will enable. Every time a chip comes back from the foundry, something valuable comes with it. Real world measurement data that tells you exactly where the model was right and where it wasn't.

[00:20:48]Dante closes this loop. Simulated predictions are compared against measured silicon results. The delta is extracted and the transistor model is updated automatically. The next design starts from a more accurate foundation than the last. This means your measurement data is not just a verification step. It is a training signal. Every tape makes your model more precise.

[00:21:12]This is the complete picture. Heavy mind handles RF circuit generation and optimization from specifications to tapered ready layout. Fully automated any circuit topology any complexity level. Dant provides the model foundation underneath. physically precise transistor models trained on your own measurement data continuously self-improving with every taper cycle.

[00:21:35]The platform connects into your existing workflows through a Python native API enabling integration with commercial EDA environments via their own interfaces and giving customers full freedom to run their own highfidelity verification and finalization steps with in-house models.

[00:21:52]And for teams that want to go further, open source FM and circuit solvers can be attached directly, keeping the platform open, flexible, and extensible from transistor models to taper. One stack, every layer, no gaps.

[00:22:09]We started this talk with circuits, real chips, real performance, real silicon.

[00:22:14]Everything since has been about how we get there faster, smarter, and without the bottlenecks that have defined RF engineering for decades. This is where it leads. An artificial RF design engineer that develops expert level solutions for any specification. One that collaborates with engineers, picking trade-offs, proposing architectures, learning preferences over time. One that scales from a single component to a complete RF system. And one that continuously improves with everyone, every customer, every silicon result that comes back from the foundry.

[00:22:49]And perhaps most importantly, when the AI handles the RF problem, the engineer is free to focus on what truly matters.

[00:22:58]Novel architectures, system level innovation. Ideas that never made it past the whiteboard because the design cycle consumed everything before them.

[00:23:07]This is not a distant road map. The foundation is built. The stack is running. The direction is clear. From design space exploration to design space intuition, the future of RF engineering is now.

[00:23:21]Thank you very much for your attention.

[00:23:23]We would like to invite you to visit us at IMS 2026 in Boston in the coming weeks. Find us at the booths 15046.

[00:23:32]We look forward to showing you what Heavymind can do in person. Visit us at rapidarf.ai or reach out directly at [email protected].

[00:23:41]Say I.