agent.rf

Ben Hardy · RF / EM lab

agent.rf the RF lab

Learn RF by doing. Each station on the bench opens an interactive tool or demo: Maxwell's equations, RF fundamentals, a Smith chart game, a VNA, RF calculators, an EM solver guide, an MRI coil builder. Behind the glass is the agent lab, where AI agents run the solvers and are scored against benchmarks.

  • Live
  • Building
  • Coming soon

Hover or Tab to a station · click or Enter to open

SOLVER LIBRARY ∇·D = ρ ∇·B = 0 ∇×E = −∂B/∂t ∇×H = J + ∂D/∂t c = 1/√(μ₀ε₀) BLACKBOARD PA DIE DRC · LVS rules PDK CABINET S11 · 1–6 GHz dB(S11) VNA · 2-PORT 12 VNA BENCH LC LPF · 0402 PORT TUNING RF BENCH · mismatch RF BENCH 3 T BIRDCAGE MRI CORNER CALCULATORS CALCULATORS RUN QUEUE em-benchmark ILLUSTRATION FEMFDTD ELMERthermal rf-agent-bench ILLUSTRATION BENCHMARKS ILLUSTRATION · NOT A REAL RUN $ agent run <task> layout layout-agent-tools solve palace-agent-tools tune circuit-agent-tools check passivity, reciprocity log rflab-core grade rf-agent-bench review a human signs off AGENT TOOLKITS WORK WITH ME WORK WITH ME

The front room

Learn by doing

Interactive tools and demos that run in your browser.
Simulation only unless a station says measured.

  • Solver library · live

    EM solver guide

    Which solver for which problem: twenty devices with real Gmsh meshes, an in-browser method-of-moments solve checked against a Python reference, and data for 62 tools.

  • Blackboard · live

    Maxwell by Heart

    Learn to write Maxwell's four equations from memory, with an interactive field for each law.

  • PDK cabinet · coming soon

    Inside a PDK, and an RF PA from scratch

    What a process design kit contains, and a series that takes a 28 GHz PA on the open IHP SG13G2 PDK from transistor model to full-wave EM of the die.

    • What is inside a PDKcoming soon
    • RF PA from scratch, seriescoming soon
  • VNA bench · live

    Smithy's Match, and a virtual VNA

    A Smith chart matching game, and a simulated two-port VNA with a real error model that you calibrate yourself, then break on purpose.

  • Port-tuning bench · coming soon

    Port tuning for layout parasitics

    Why a laid-out LC low-pass detunes, and how to fix it with one EM solve: put a port in every footprint, then retune the parts in circuit simulation.

    • Port tuning walkthroughcoming soon
  • RF Bench · live

    RF Bench: interactive RF fundamentals

    Change a part, watch the response: mismatch with a standing wave and Smith chart, resonators and Q, matching, splitters, PA load lines, coupled resonators and filter order. Predict the number first, then drill.

  • MRI corner · live

    Birdcage coil builder

    How a 3 T birdcage coil works: build it in Gmsh, solve its modes in Palace, find the uniform k = 1 pair and tune it onto the Larmor frequency.

  • Calculator drawer · live

    Visual RF calculators

    Skin depth, coax, microstrip, waveguide and path loss, each drawing its fields as you type. The microstrip page runs a live field solve, within 0.36 % of Hammerstad–Jensen over 20 cases.

Behind the glass

The agent lab

The front room is for people. The lab behind it is where AI agents work: they set up and run EM and circuit solves through tool servers, queue jobs, log every run, and are scored against benchmarks with known answers. A person reviews before anything is built.

  • Agent toolkits · building

    Tools agents call

    MCP servers that let agents drive open solvers: set up, lint, run and check Palace FEM models; build layouts; tune circuits. Each tool returns checks, not just numbers.

  • Benchmarks · building

    Known answers to score against

    Field solvers against reference solutions, error versus unknowns, time and memory (phase 1: Palace and MEEP, five cases, plus Elmer thermal cases), and a task benchmark that scores agents on RF design jobs.

    • em-benchmarkbuilding
    • Elmer thermal casesbuilding
    • rf-agent-benchbuilding
  • Run queue · building

    Every run, logged

    A job queue and run database: what each agent ran, on which solver, at what cost, and whether it passed. Later, a read-only dashboard of runs and benchmark results.

    • rflab-core, run DB and job queuebuilding
    • Dashboard of runs and resultscoming soon
    • The closed loopbelow

The closed loop

Agents run the design and solve steps. Graders and measurement decide what is true.

  1. Spectargets, limits, process
  2. Agent designlayout, parameters
  3. EM + circuit solveGmsh, Palace, HFSS, surrogates
  4. Gradersgolden cases, convergence, second solver
  5. Human reviewbefore money is spent
  6. FabPCB, MPW, prototype
  7. MeasureVNA, load-pull, bench
  8. Calibratemodels, surrogates, rules

grader fails → redesignmeasured data → update models

Work with me

Verified AI design loops for RF hardware

I set up agent-driven simulation and design flows inside RF hardware companies, and the checks that show when a result is wrong, so faster design does not mean less trust. In short: an autonomous RF lab of agents, like the one behind the glass, built around your parts and your measurements.

Ben Hardy · PhD, MRI RF hardware · EM simulation application engineer · AI design-agent evaluation

What changed. A simulated design now costs hours, not weeks: open solvers (Gmsh, Palace, openEMS) and agents do the meshing, solving and sweeping. The bottleneck is now trust (is the answer right?) and hardware agreement (does the part match the model?).

  • An EM and circuit pipeline that agents can run headless, validated against your HFSS results and your measurements.
  • Graders and golden cases for each component family, with pass rules, and a second-solver check where it matters.
  • Surrogates with error bars over a stated design space, and a cost per design (unknowns, GB, minutes).
  • A closed loop from spec to measured part, with human review gates where a wrong answer is expensive.
Start here

Pilot 3–4 weeks · fixed scope

One component family. Open EM flow on your design, agent loop, grader and surrogate, validated against your data. Report with cost per design.

Build

Lab build 8–12 weeks

Several families, a solver farm, a golden-case library, fab-and-measure feedback, and a handover to your team.

Ongoing

Advisory monthly

Review of AI design claims, benchmark design, and tool and vendor choices.

Recent public work

Each one has a check that could have failed.

Port tuning, 5th-order LC low-pass, 0402, RO4003C2.21 → 2.91 GHzLayout parasitics pulled the cutoff low. One EM solve plus 0.2 s of tuning restored it; a second EM solve agreed within 0.005 dB.
3 T birdcage coil builder (Palace)16.6 → 3.7 GBPEC/PMC symmetry planes cut the solve 4×. A convergence study caught a 0.7 % offset.
Interactive EM solver guide20 devicesReal Gmsh meshes, an in-browser MoM solve checked against a Python reference, data for 62 tools.
Visual RF calculators≤ 0.36 %Live microstrip field solve against Hammerstad–Jensen over 20 cases; 33 automated tests.
28 GHz PA on an open PDK (IHP SG13G2)1.14 M unknownsOpen tools only: transistor model, load-pull, layout, full-wave EM of the die.

All results are simulation only unless measured.

Get in touch via benjhardy.github.io

Start with a pilot: one component family, fixed scope.