Inverse design for passive devices · rfsurrogates
Filter Surrogate
Edge-coupled microstrip band-pass filter, 3–7 GHz, RO4350B (εr 3.66, h 0.508 mm): a textbook coupled-line model plus a network trained on 261 Palace FEM solves. It gives |S21| and |S11| in about 2 ms in your browser. Simulated only.
f0: centre of |S21|². BW: widest band within 3 dB of the peak; "split" when the passband has two peaks. Return loss: worst |S11| between the outer reflection zeros. Spread: seed-to-seed deviation of |S21| in the passband.
Same rectangles as the mesh generator. The pulses show |S21| and |S11|, not fields. The blue faces are the wave ports. Drag to turn; keys: arrows, + and −, 0 to reset, E to explode.
Inverse design
Set f0, bandwidth and minimum return loss. The page tries 256 geometries in the box and refines the best eight. It takes about 2 s.
Click a row to load it.
How good is the surrogate?
Error on 26 Palace designs that no model saw in training.
| model | passband |S21| error | f0 error | BW error |
|---|---|---|---|
| prior only (Pozar coupled lines) | 2.86 dB | 61 MHz | 114 MHz |
| prior + DeepONet, 5 seeds | 0.52 dB | 7.9 MHz | 10.8 MHz |
| DeepONet from scratch, 5 seeds | 0.32 dB | 4.2 MHz | 26.8 MHz |
BW error uses the 6 test designs with one passband.
Mesh. The training solves use a coarse mesh. On 10 designs solved again with a finer mesh, the coarse mesh is 1.03 dB off and the model is 0.91 dB off.
Prior. With 10 training solves, the prior gives 1.8 dB against 9.5 dB without it. From about 100 solves, the model without the prior is better.
Spread. The band is smaller than the true error. Use it as a lower bound.
Simulated against simulated, one substrate. Model card.
Palace-verified designs
The nominal and five test designs. Load one to show its Palace curve.
Verify in Palace
Run these commands from the repo root to solve this geometry in Palace 0.16. One solve takes about 11 min on 4 cores and 2 GB of RAM.
Contribute data
You can solve designs on your own machine and send the results. A checker tests each result before it is used. Data contributions open with v0.2: see CONTRIBUTING.md.
Links
- rfsurrogates on GitHub: code and data tools.
- Model card and weights (Hugging Face).
- Birdcage builder: another Palace front end.
- agent.rf: the lab hub.
- Palace: the open-source FEM solver.
References
- D. M. Pozar, Microwave Engineering, 4th ed., Wiley (2012), sec. 8.7, coupled line filters.
- M. Kirschning, R. H. Jansen, "Accurate wide-range design equations for the frequency-dependent characteristic of parallel coupled microstrip lines," IEEE Trans. MTT 32(1), 83–90 (1984).
- E. Hammerstad, Ø. Jensen, "Accurate models for microstrip computer-aided design," IEEE MTT-S Int. Microwave Symp. Dig. (1980).
- M. Kobayashi, "A dispersion formula satisfying recent requirements in microstrip CAD," IEEE Trans. MTT 36(8), 1246–1250 (1988).
- L. Lu, P. Jin, G. Pang, Z. Zhang, G. E. Karniadakis, "Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators," Nature Machine Intelligence 3, 218–229 (2021).
- Palace: 3D finite element solver for computational electromagnetics, AWS Labs (open source).