Answers are grounded in this paper and cite their sources.
Deep inverse models (DIMs) can circumvent the simulation bottlenecks of traditional optical metasurface design, but they have predominantly been applied to low-dimensional targets and isolated meta-atoms characterized under the local phase approximation (LPA). Their efficacy on high-dimensional near-field targets produced by strongly coupled, heterogeneous structures remains largely uncharacterized. We benchmark modern forward emulators and DIMs from three model families on full-wave simulations of a heterogeneous nanopillar array that captures inter-element near-field coupling directly. The preferred model depends on the candidate budget: conditional invertible neural networks (cINNs) give lower design error at one to two orders of magnitude lower cost, while the neural adjoint attains the lowest field error once ∼10 2 or more candidates are searched, an advantage that holds under full-wave verification. We then introduce a multi-wavelength framework that conditions a single emulator on a continuous wavelength descriptor, enabling inversion under varying degrees of spectral specification, and show that additional bands separate near-degenerate solutions that a single band cannot distinguish.
No takes yet. Share an insight, caveat, or question.
Mick et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: