Deep learning meets nanophotonics: A generalized accurate predictor for near fields and far fields of arbitrary 3D nanostructures

TL;DR

Conventional nanophotonic simulations can be time-consuming, while prior neural-network approaches were typically tailored to specific inverse problems and geometries. The paper develops a 3D convolutional predictor of internal fields for arbitrary 3D nanostructures, then derives near- and far-field observables from those predictions. It reproduces diverse plasmonic and dielectric effects with few-percent errors and accelerates predictions by several orders of magnitude, while remaining limited by training conditions and extrapolation risk.

• Problem

  Conventional numerical simulations can take hours or days, and prior neural-network approaches generally required separate designs for specific inverse problems and geometric models.

• Method

  A fully convolutional 3D neural network is trained to predict the coupled-dipole representation of internal fields in arbitrary-shaped nanostructures.

• Results

  The predictor reproduces complex near- and far-field effects in plasmonic and dielectric nanostructures without effect-specific training, with secondary quantities derived at few-percent uncertainty and predictions 3 to 5 orders of magnitude faster than simulations.

• Takeaways & Limitations

  A single generalized predictor can support rapid modeling and potentially universal nano-photonic inverse design across varied structures and optical effects.

• Takeaways & Limitations

  The network is limited to training conditions, requiring separate training when the structure model, material, or illumination configuration changes, and performance is reduced outside the known parameter space.