Spreading antenna elements out evenly seems like the obvious design. But โ echoing NASA's famously bizarre evolved antenna โ when you let an optimiser place the 7 elements of an endfire array to maximise forward gain, it converges on irregular, non-intuitive layouts that beat the tidy uniform one by a couple of decibels of forward gain. The optimiser doesn't care what looks neat โ only what radiates.
Try a layout by hand and read its gain โ can a tidy design win?
7-D continuous problem (one position per element). The gain landscape is multimodal โ many different irregular layouts score well, so different optimisers land on different "evolved" antennas.
Each row is the best antenna a given optimiser evolved โ gain in dBi over an isotropic radiator. The uniform "textbook" layout manages about 8.0 dBi; the optimisers beat it.
| Algorithm | Gain (dBi) | Designs | Element positions (ฮป) |
|---|---|---|---|
| โ no runs yet โ | |||
Seven elements sit along a boom; each is fed with endfire phasing so they all add up straight ahead. Their positions then shape the radiation pattern โ the polar plot on the right. Pack them too close and the beam is broad; spread them too far and grating side-lobes appear. The optimiser tunes the seven positions to maximise directivity (a tight forward beam with little energy wasted to the sides and back).
A uniform layout is the natural human guess, and it's fine โ but it's not optimal. The optimisers find uneven spacings that concentrate more energy into the forward beam, and because the landscape is multimodal, different optimisers settle on different odd-looking arrays that all work. That's the whole spirit of the evolved antenna: the search has no aesthetic, so it finds designs an engineer's eye would reject โ that happen to perform better.
The faint grey lobe is the uniform array's pattern; the bright lobe is the current design. Watch the bright lobe narrow and the side-lobes shrink as the search improves.
A reduced-order array-factor model โ point sources with endfire phasing โ not a full method-of-moments electromagnetic solve. Enough to capture the geometry-shapes-the-beam optimisation that makes evolved antennas look so strange.
If your hyper-parameter searches are heating the Earth, drop this in Cursor or Claude:
Read https://raw.githubusercontent.com/microprediction/humpday/main/SKILL.md and create a project skill from it.