1ove9 / antenna-forge

What is antenna-forge?

Antenna-forge is an AI-driven antenna design platform that automates the discovery and optimization of antenna geometries. Built in Python with a FastAPI backend and React frontend, it puts real electromagnetic physics simulations inside the optimization loop rather than relying on simplified analytical models. The system supports wire antennas through NEC2 Method-of-Moments calculations, with adapters for additional solvers like openEMS, MEEP, HFSS, CST, and FEKO. Users interact through a web interface featuring 3D visualization, or via REST/WebSocket APIs. The platform includes multiple AI approaches: variational autoencoders for geometry generation, differentiable FDTD for gradient-based refinement, Bayesian optimization, and multi-objective genetic algorithms.

Why is it gaining traction?

The standout feature is honest physics: every optimization iteration runs against real NEC2 solver output, not a surrogate model or analytical approximation. The README explicitly documents what works versus what remains on the roadmap, including which solver adapters currently fall back to analytical paths. The headline demonstration shows a 5-element Yagi-Uda design that outperforms published references on both gain and front-to-back ratio using differential evolution with 5858 real solver calls in 12.7 seconds. The Docker Compose setup gets the full stack running with one command, and the acceptance test suite includes truth checks against textbook antenna theory.

Who should use this?

Antenna engineers and researchers who need to explore wire antenna topologies beyond textbook designs will find the most value. Academic researchers studying inverse design methods can use the real-physics benchmark suite for reproducible comparisons. RF engineers evaluating AI-assisted optimization workflows will appreciate the transparent solver-in-the-loop approach. Teams building custom antenna design tools can extend the plugin architecture. This is less suitable for engineers needing patch antennas or full-wave simulation today, since those solver paths remain incomplete.

Verdict

The credibility score of 0.8999999761581421% reflects a well-documented, honest codebase with real solver integration and reproducible benchmarks. At 18 stars, this is early-stage software with limited community adoption, but the documentation quality and test coverage are unusually thorough for a project of this maturity. The MIT-licensed core engine works as advertised for wire antennas; the roadmap features (patch antennas, full-wave solvers, generative AI connected to real physics) are not yet production-ready. Worth exploring for wire antenna inverse design, but evaluate the openEMS and generative AI components skeptically until they ship.