Source

engines/quan_loncar_v1/package/quan_loncar_optimizer/engine/quan_loncar_series_phase44/README.md · assembled 2026-07-29 15:57 UTC.

---

Phase 44: unified nanobeam autograd adapter

evaluate_nanobeam_autograd_state(...) evaluates the current corrected resonance engine and returns all scalar values plus a custom reverse-mode pullback. One NanobeamObjectiveCotangents object can weight resonance, total Q, physical or normalized effective mode volume, external linewidth, external Q, remaining-loss Q, and guided decay fraction. The pullback returns one gradient in the documented 45-coordinate gap/rx/ry ordering.

The interface is framework-neutral: it can be wrapped by JAX custom_vjp, a PyTorch autograd.Function, or used directly by SciPy optimizers. Neither JAX nor PyTorch is installed in the current environment, so this phase exposes and tests the mathematical VJP without adding a framework dependency.

Since Phase 46, the forward implementation shares one immutable base/plus/minus frequency batch between QNM/mode-volume and coupling. This halves the expensive continuum assemblies without changing this public API.