Maxwell's equations, solved in a browser tab

NobuRF is a 3D field solver for antennas. It runs on your GPU through WebGPU (tested in Chrome and Edge). On my laptop's GTX 1660 Ti, a 5.8 GHz patch antenna (66 × 66 × 48 cells at 0.5 mm, run until the field energy has fallen 50 dB) solved in 1.9 s, about 3 s from click to results.

Figure 1. The vertical electric field inside the substrate of a 5.8 GHz inset-fed patch, over the 2.85 ns the solver simulates after the feed is driven. 64 × 64 × 48 cells at 0.5 mm, 3,000 time steps. Screen recording from the browser.

How close is it to a reference solver?

I ran the same patch in openEMS, an independent open-source FDTD code, with the same 50 Ω port at the end of the feed line, both on graded meshes with 0.25 mm cells at the copper, and openEMS meshed the way its authors advise for thin copper (grid lines a third of a cell inside each edge). The resonant frequencies agree within 1 %: 5.30 against 5.34 GHz for the edge-fed patch, 5.47 against 5.47 GHz with a 4 mm inset feed. The inset dip reaches −21.5 against −20.4 dB, and the impedance there differs by 2.1 Ω in R and 1.6 Ω in X (44.3 + j5.6 against 42.2 + j4.0 Ω).

On a uniform 0.25 mm grid, which snaps the edges of the inset notch, the inset dip is still too deep (−37 dB). Comparisons with lines on the drawn copper edges read both codes about 1 % low and hide that shared error; the validation report has the full set.

S11 against frequency for NobuRF and openEMS. Edge-fed: NobuRF dips at 5.30 GHz to −7.9 dB, openEMS at 5.34 GHz to −8.2 dB. Inset-fed: both dip at 5.47 GHz, NobuRF to −21.5 dB, openEMS to −20.4 dB. NobuRF openEMS Edge-fed 0 −2 −4 −6 −8 −10 4.5 5.0 5.5 6.0 6.5 7.0 7.5 5.30 / 5.34 GHz frequency (GHz) S11 (dB) Inset feed, 4 mm 0 −5 −10 −15 −20 −25 4.5 5.0 5.5 6.0 6.5 7.0 7.5 5.47 / 5.47 GHz frequency (GHz) S11 (dB) S11 against frequency for NobuRF and openEMS. Edge-fed: NobuRF dips at 5.30 GHz to −7.9 dB, openEMS at 5.34 GHz to −8.2 dB. Inset-fed: both dip at 5.47 GHz, NobuRF to −21.5 dB, openEMS to −20.4 dB. NobuRF openEMS Edge-fed 0 −2 −4 −6 −8 −10 4.5 5.5 6.5 7.5 5.30 / 5.34 GHz frequency (GHz) S11 (dB) Inset feed, 4 mm 0 −5 −10 −15 −20 −25 4.5 5.5 6.5 7.5 5.47 / 5.47 GHz frequency (GHz) S11 (dB)
Figure 2. S11 of the same patch in NobuRF and openEMS v0.37, both on graded meshes with 0.25 mm cells at the copper (openEMS with its advised lines a third of a cell inside each copper edge) and the same 50 Ω port at the feed end. Note the different vertical scales.

Most of the other checks in the validation report pass. A closed metal cavity resonates within 0.23 % of theory even on a coarse 1 mm grid, a plane wave crosses the grid at the speed it should, and a metal short reflects everything at the right phase. The open boundary reflects −84 dB of a wave hitting it head-on, and doubling the air around the board changes none of the results. What doesn't pass yet: the coarser 0.5 mm grid the live demo uses reads 1.7 to 2.4 % high in frequency. Not checked yet: the boundary at oblique angles, and a real board on a VNA.

The validation report has the method, the scripts, every check and its figure.

From one sentence to an antenna

For the Kimi K3.1 Pro hackathon on November 21 and 22, the solver becomes a tool for an AI agent. You type one line, such as "a 2.4 GHz patch on a 50 mm FR-4 board". Kimi, running inside OpenClaw, reads it, sets up the design, runs NobuRF solves, changes the geometry between them, and stops when S11 meets the goal it read or when it runs out of solves. You can type a correction while it works, and it reads that with its next solve.

I'll give the same sentence to the same model on Cerebras and on a GPU-based provider, one lane after the other on the same solver, and publish the logs. Each lane's clock starts at its first model request, and the logs keep model time and solver time apart. Each lane ends on the 3D radiation pattern of the antenna it settled on.

The results will be posted here after the runs.

S11 after each of five edits. At 5.8 GHz it goes from −2.4 dB at the start to −19.9 dB after the trim slot. 0 −5 −10 −15 −20 −25 4.5 5.0 5.5 6.0 6.5 7.0 7.5 5.8 GHz target start −2.4 dB after 5 edits −19.9 dB frequency (GHz) S11 (dB) S11 after each of five edits. At 5.8 GHz it goes from −2.4 dB at the start to −19.9 dB after the trim slot. 0 −5 −10 −15 −20 −25 4.5 5.5 6.5 7.5 5.8 GHz target start −2.4 dB final −19.9 dB frequency (GHz) S11 (dB)
Figure 3. A scripted five-edit tuning run on the demo grid, from before the agent work: two inset changes, a shorter patch, another inset change, and a 0.3 mm trim slot. S11 at 5.8 GHz goes from −2.4 to −19.9 dB on the 0.5 mm demo grid. The gray curves are the starting shape and the first four edits, the orange one the fifth. Treat the depth as approximate: the frequency of the dip is the reliable part.

Why I built it

Commercial 3D EM solvers are licensed per seat, at prices set for companies. Students lose access when they graduate, and small hardware teams often tune on the bench instead.

The physics has been public since Maxwell published it in 1865, and the GPU in an ordinary laptop is now fast enough to run it. I wanted to see how much of the expensive part a browser could do, and to have it say plainly where it is wrong.

What's missing

Since the first version the solver has gained a CPML boundary, a graded mesh, conformal curved copper, radiation patterns from a near-to-far-field transform, a modeled SMA feed, copper loss that grows as √f with foil roughness, and Dk/Df tables for laminates. It stops once the field energy has died down. Multi-port S-parameters have a first check: a two-port low-pass filter from Sheen et al. (1990) rolls off within 1.0 % of the measured curve, while their four-port branch-line coupler does not match yet. Conductor loss is checked on a cavity and a microstrip resonator, not yet on a patch. Gerber, DXF and STEP import work but are not yet validated. Still to do, in the order I plan to work on it, each with a validation figure:

  1. Microstrip mode ports with de-embedding.
  2. Vias and thin-wire feeds.
  3. Larger grids through f16 storage and multiple GPUs, for 2.4 GHz boards and arrays.
  4. A fabricated board measured on a VNA.

Run it

The workbench runs in this browser tab: nothing to install, and your project stays on your machine. Without WebGPU it solves on the CPU, which is much slower.

To run it from the source instead, with the agent bridge, you need Node.js 20.19 or later, and a browser with WebGPU, such as Chrome or Edge, for GPU solves.

git clone https://github.com/Llompi/NobuRF-Public
cd NobuRF-Public
npm install
npm run dev

Then open http://localhost:5173. On a Windows laptop with two GPUs, start Edge with --force_high_performance_gpu, or Windows hands the work to the integrated one.

Found a wrong answer? Open an issue with the geometry.