The live preview is produced by a small neural surrogate model,
not a field solver. It was trained on 10,898 real NEC2 simulations
(a method-of-moments engine) — 5,898 from an optimization sweep and
5,000 from a uniform Latin-hypercube sample of the design space —
and reproduces the held-out test set with a coefficient of
determination of roughly R² 0.90–0.96
across the four metrics. Its purpose is instant exploration:
moving a slider re-evaluates the array in well under a millisecond,
which a full simulation cannot.
The surrogate is an approximation. For any design that matters,
treat these numbers as guidance and confirm them with a real solve in
NEC2 or openEMS.
Accuracy is best in the high-performance region and degrades toward
the edges of the sampled space.
The radiation pattern is an indicative azimuth cut constructed
from the predicted forward gain (main-lobe width) and front-to-back
ratio (back-lobe level). It is a visualization aid, not a complete
far-field radiation pattern; side-lobe structure and elevation
behaviour are not modelled.
Inputs are nine geometric parameters in wavelengths (λ): element
lengths L_ref, L_drv, L_d1–L_d3 and
inter-element spacings s_ref, s_d1–s_d3.
Outputs are forward gain (dBi), front-to-back ratio (dB) and the
feed-point impedance R + jX (Ω).