FrontierAbsorbBench v0.1: An Auditable Protocol and Software Framework for Broadband Absorber Optimization
FrontierAbsorbBench (FAB) v0.1 introduces an auditable protocol and software framework for comparing optimization agents and traditional algorithms on a frozen broadband electromagnetic absorber design problem: 0.1-18 GHz, incidence angles 0-80 degrees, and TE/TM polarization, under explicit material, geometry and resource constraints. The implementation includes a complex128 stratified reflection solver, joint material-data gates, weighted metrics, workload accounting, checkpoint nominations, offline adaptive verification, baseline optimizers, a restricted JSON agent interface, and Windows migration scripts. This is a methods and software preprint released BEFORE runtime validation. No numerical tests, Windows/CUDA validation, baseline experiments, or LLM experiments have been executed for this release. No rankings, performance gains, experimentally fabricated designs, or physical-limit breakthroughs are claimed. Three authored causal synthetic fixtures and twelve development task templates are provided. The physical leaderboard is disabled because full-band measured material spectra and independently calibrated taper reduced models are not available. General 3D structures are explicitly unsupported. Performance engineering includes independent-process CPU suites, bounded BLAS/OpenMP threading, batched NumPy/PyTorch evaluators, CUDA execution, frequency chunking, and streaming layer recursion in complex128. The target-machine benchmark script measures throughput, parity and memory after validation. No speedup or target-hardware performance has been measured for v0.1. The deposited source archive contains the implementation, unexecuted test suite, task/configuration files, provenance audit, reproducibility instructions, and manuscript source. The doctoral thesis used as background is cited but not redistributed. Paper license: CC BY 4.0. Software license: MIT. Authored synthetic fixtures: CC0 1.0. AI assistance was used for implementation and manuscript drafting; responsibility for subsequent validation and interpretation remains with the named human author.
Authors
- xun zhe zhong
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-15
- DOI
- https://doi.org/10.5281/zenodo.24246579
- Primary Topic
- Advanced Antenna and Metasurface Technologies
- Type
- preprint