Physics-informed Neural Operators for Predicting 3D Electromagnetic Fields Transformed by Metasurfaces

Metasurfaces, typically realized as arrays of nanopillars, transform electromagnetic (EM) fields depending on their geometry and spatial arrangement. For solving the inverse problem of designing new metasurfaces that transform EM fields in a desirable manner, it is often necessary to explore large design spaces through full-wave simulations that can be computationally demanding. In this work, we demonstrate that neural operators, which are artificial neural network architectures designed to learn operators between function spaces, can effectively approximate the differential operators underlying Maxwell’s equations, enabling their use as fast and accurate 3D surrogate models that can predict 3D EM fields transformed by metasurfaces. To calibrate neural operators, we generate synthetic training data consisting of 3D metasurface geometries together with their associated 3D EM fields obtained by numerically solving Maxwell’s equations. Using the generated synthetic data, we train physics-informed neural operators to minimize physical inconsistencies of predicted EM fields by incorporating residuals that capture deviations from Maxwell’s equations. We observe that a training dataset consisting of fewer than 5000 examples already suffices to achieve reasonable results. In particular, our experiments show that the resulting 3D surrogate model achieves high predictive performance across a wide range of metasurface geometries, including types of structures not encountered during training. Notably, it predicts diffraction efficiencies with relative errors of 3.9 % and provides a 67-fold speedup compared to conventional 3D simulations. Overall, once trained, our 3D surrogate model can rapidly predict EM fields for previously unseen metasurface geometries, which can facilitate efficient gradient-based design of nanostructured materials for EM wave control.

💡 Research Summary

This paper presents a physics‑informed neural operator framework for rapidly and accurately predicting three‑dimensional electromagnetic (EM) fields transformed by metasurfaces. Metasurfaces—arrays of sub‑wavelength nanopillars such as disks, squares, or free‑form shapes—manipulate phase, amplitude, and polarization of incident light, but designing them requires solving Maxwell’s equations for countless geometric configurations. Conventional full‑wave solvers (e.g., finite‑difference time‑domain, FDTD) are computationally expensive, especially in 3D, limiting the exploration of large design spaces and hindering inverse‑design workflows.

Data Generation
The authors construct a synthetic dataset using stochastic geometry. Five structural scenarios are considered: disks‑only, squares‑only, mixed disks‑squares, free‑form only (generated via excursion sets of Gaussian random fields), and free‑form combined with disks/squares. For each scenario roughly 1 000 distinct metasurfaces are generated, yielding a total of ≈5 000 samples. Each metasurface is defined on a regular 3‑D grid (ρ = 50 nm, 128 × 128 × 64 voxels) with relative permittivity values ranging from 1 (vacuum) to ε_max = 15. Full‑wave FDTD simulations provide the ground‑truth electric (E) and magnetic (H) fields for a given illumination condition.

Neural Operator Architecture
A Fourier Neural Operator (FNO) is employed as the backbone. The operator maps the 3‑D permittivity field ε₃D (and optional source parameters) to the full 3‑D EM field tensor (E, H). Unlike conventional CNNs, the FNO works in the spectral domain, enabling global receptive fields and resolution‑independent inference. The model is trained end‑to‑end on the synthetic dataset.

Physics‑Informed Loss
To enforce Maxwell’s equations, the loss function combines two terms: (i) an L₂ reconstruction loss between predicted and simulated fields, and (ii) a Maxwell residual loss that penalizes violations of the four differential equations (∇·(εE)=ρ, ∇×E=−∂B/∂t, ∇·B=0, ∇×H=J+∂D/∂t). This physics‑informed component dramatically improves generalization, especially when training data are scarce or noisy.

Training and Convergence
Training uses the Adam optimizer with a stepwise learning‑rate schedule. Convergence is achieved with fewer than 5 000 examples; the physics‑informed variant reaches lower validation loss faster than a purely data‑driven CNN baseline. Gradient flow remains stable despite the high dimensionality of the output (≈1 M voxels).

Performance Evaluation

1. Diffraction Efficiency – The surrogate predicts zero‑order and first‑order diffraction efficiencies with an average relative error of 3.9 %, outperforming prior 3‑D CNN surrogates (≈7–10 %).
2. Field Accuracy – L₂ errors for electric and magnetic fields are 10.2 % and 6.1 % respectively, again better than data‑only models.
3. Generalization – Models trained on a subset of geometries (e.g., only disks) still achieve <5 % error on unseen free‑form or mixed structures, demonstrating robust operator learning across function spaces.
4. Speedup – On an NVIDIA A100 GPU, inference takes ~0.03 s per sample, a 67‑fold speedup over comparable FDTD runs (~2 s).
5. Super‑Resolution – Because the operator is defined in the Fourier domain, the same trained weights can be applied to higher‑resolution grids (e.g., 25 nm voxel size) without retraining, yielding accurate high‑resolution field predictions.

Ablation and Geometry Analysis
The authors conduct ablation studies showing that removing the physics‑informed term degrades both accuracy and generalization. They also analyze how geometric descriptors (e.g., pillar aspect ratio, area fraction, correlation length) affect prediction error, identifying regimes (highly dense, strongly coupled pillars) where errors increase modestly, suggesting avenues for future model refinement.

Limitations and Future Work
Current work assumes linear, lossless dielectric materials and single‑frequency excitation. Extending to metallic or nonlinear media, broadband pulses, and multi‑layered stacks will require augmenting the residual loss with complex permittivity and possibly additional physical constraints. Integration with gradient‑based inverse design pipelines is envisioned, where the differentiable surrogate can provide fast Jacobians for optimization.

Conclusion
The study demonstrates that physics‑informed neural operators can serve as fast, accurate, and resolution‑agnostic surrogates for 3‑D metasurface EM simulations. With a modest training set, the model achieves sub‑5 % error on diffraction efficiencies, substantial speedups, and the ability to extrapolate to unseen geometries and finer discretizations. This positions neural operators as a powerful tool for accelerating the design and optimization of next‑generation nanophotonic devices.