Complementary Reasoning in Meta-Optical Networks

• Complementary reasoning is an approach that alternates between unconstrained learning and physically grounded projections to ensure efficient meta-optical neural network training.
• It employs spatial decoupling by solving localized PDE subproblems, which drastically reduces computational complexity compared to full-scale simulations.
• The SP2RINT framework exemplifies this method by bridging digital model training with hardware constraints to achieve fabricable and scalable meta-optical devices.

Complementary reasoning is an optimization strategy employed to reconcile physically motivated constraints with flexible, data-driven parameterization in meta-optical neural network (DONN) training. By combining unconstrained learning phases in a relaxed parameter space with periodic, physically grounded projections, complementary reasoning ensures both learning efficiency and physical realizability. Its application is central to recent advances in meta-optical hardware optimization, where direct physical simulation is often intractable but physically infeasible solutions are common if training ignores physics entirely. The SP2RINT approach exemplifies complementary reasoning through its spatially decoupled, progressive inverse design framework for scalable PDE-constrained learning (Ma et al., 23 May 2025).

1. PDE-Constrained Learning in Meta-Optical Neural Networks

Meta-optical neural networks (DONNs) are hybrid analog–digital systems in which electromagnetic wave propagation through stacked metasurfaces acts as the main computational substrate. Each metasurface layer is characterized by a permittivity distribution, $$\varepsilon = (\varepsilon_1, \ldots, \varepsilon_K)$$, governing a transfer matrix $T_i(\varepsilon_i)$ implicitly defined by Maxwell’s equations: $$A(\varepsilon_i) h_i = b(x_i), \qquad A(\varepsilon_i) := \nabla \times (\varepsilon_i^{-1} \nabla \times) - \omega^2 \mu_0 \varepsilon_0$$ The end-to-end network response is

$$h_{\text{out}} = U_{K+1}\, T_K(\varepsilon_K)\, U_K\, \ldots\, T_1(\varepsilon_1)\, U_1\, x$$

where $U_i$ are diffraction operators. The training task is formulated as a stochastic PDE-constrained optimization: $$\min_{\varepsilon, w}~ \mathbb{E}_{(x, y)} \left[ \mathcal{L}\left(g_w(|U_{K+1}\prod_{i=1}^K T_i(\varepsilon_i) U_i x|^2), y \right) \right]\quad \text{subject to}~A(\varepsilon_i) h_i = b(x_i)$$ This setting enforces physical implementability via Maxwell’s constraints while seeking digital-task-optimal performance.

2. Relaxation to Banded, Freely Trainable Representations

Direct end-to-end training of DONNs parameterized via $\varepsilon$ is computationally prohibitive due to the necessity of solving high-dimensional Maxwell PDEs at each iteration. Complementary reasoning introduces a relaxation: metasurface layers are reparameterized by banded transfer matrices $\tilde{T}_i \in \mathbb{C}^{n \times n}$, collected as $\tilde{T} = (\tilde{T}_1,\ldots, \tilde{T}_K)$, and trained as unconstrained weights: $$\tilde{T}_i \approx T_i, \quad \text{subject to}~T_i~\text{implementable}$$ Bandedness is physically motivated by the near-field diffraction limit: $T_i(\varepsilon_i)$0 In this relaxed phase, gradients are computed as in standard neural network training (i.e., via backpropagation), with no PDE solves required.

3. Complementary Alternation: Progressive Physical Projections

SP2RINT alternates between two complementary training modes:

• Unconstrained relaxed optimization: $T_i(\varepsilon_i)$1 and $T_i(\varepsilon_i)$2 are updated via gradient descent to optimize the network objective, maintaining rapid learning unconstrained by the difficult PDE constraints.
• Inverse-design projection: At scheduled intervals, each relaxed transfer matrix $T_i(\varepsilon_i)$3 is “projected” to a physical implementable $T_i(\varepsilon_i)$4 by solving: $T_i(\varepsilon_i)$5 using adjoint-based gradient descent. Soft binarization with an annealed sharpness schedule $T_i(\varepsilon_i)$6 is used to prevent local minima and progressively enforce physical constraints.

Epoch-level system calibration further aligns the total cascaded physical response $T_i(\varepsilon_i)$7 to its target $T_i(\varepsilon_i)$8 by minimizing $T_i(\varepsilon_i)$9 with respect to $$A(\varepsilon_i) h_i = b(x_i), \qquad A(\varepsilon_i) := \nabla \times (\varepsilon_i^{-1} \nabla \times) - \omega^2 \mu_0 \varepsilon_0$$0.

A summary of the alternating workflow is as follows:

Phase	Variable Updated	Physical Simulation Required
Relaxed DONN Training (TrainStep)	$$A(\varepsilon_i) h_i = b(x_i), \qquad A(\varepsilon_i) := \nabla \times (\varepsilon_i^{-1} \nabla \times) - \omega^2 \mu_0 \varepsilon_0$$1, $$A(\varepsilon_i) h_i = b(x_i), \qquad A(\varepsilon_i) := \nabla \times (\varepsilon_i^{-1} \nabla \times) - \omega^2 \mu_0 \varepsilon_0$$2	No
Inverse-Design Projection (DesignStep)	$$A(\varepsilon_i) h_i = b(x_i), \qquad A(\varepsilon_i) := \nabla \times (\varepsilon_i^{-1} \nabla \times) - \omega^2 \mu_0 \varepsilon_0$$3, $$A(\varepsilon_i) h_i = b(x_i), \qquad A(\varepsilon_i) := \nabla \times (\varepsilon_i^{-1} \nabla \times) - \omega^2 \mu_0 \varepsilon_0$$4	Yes (patchwise)
System Calibration	$$A(\varepsilon_i) h_i = b(x_i), \qquad A(\varepsilon_i) := \nabla \times (\varepsilon_i^{-1} \nabla \times) - \omega^2 \mu_0 \varepsilon_0$$5	Yes (global, optional)

4. Spatial Decoupling: Local Atomic Patch Probing

To address the computational barrier of full-size transfer-matrix inference ($$A(\varepsilon_i) h_i = b(x_i), \qquad A(\varepsilon_i) := \nabla \times (\varepsilon_i^{-1} \nabla \times) - \omega^2 \mu_0 \varepsilon_0$$6 per metasurface), complementary reasoning motivates spatial decoupling. Each metasurface of $$A(\varepsilon_i) h_i = b(x_i), \qquad A(\varepsilon_i) := \nabla \times (\varepsilon_i^{-1} \nabla \times) - \omega^2 \mu_0 \varepsilon_0$$7 meta-atoms is partitioned into $$A(\varepsilon_i) h_i = b(x_i), \qquad A(\varepsilon_i) := \nabla \times (\varepsilon_i^{-1} \nabla \times) - \omega^2 \mu_0 \varepsilon_0$$8 overlapping atomic patches of size $$A(\varepsilon_i) h_i = b(x_i), \qquad A(\varepsilon_i) := \nabla \times (\varepsilon_i^{-1} \nabla \times) - \omega^2 \mu_0 \varepsilon_0$$9. Rather than $$h_{\text{out}} = U_{K+1}\, T_K(\varepsilon_K)\, U_K\, \ldots\, T_1(\varepsilon_1)\, U_1\, x$$0 full-size PDEs, only $$h_{\text{out}} = U_{K+1}\, T_K(\varepsilon_K)\, U_K\, \ldots\, T_1(\varepsilon_1)\, U_1\, x$$1 small PDEs are solved per forward or adjoint pass:

• Extract local permittivity patch $$h_{\text{out}} = U_{K+1}\, T_K(\varepsilon_K)\, U_K\, \ldots\, T_1(\varepsilon_1)\, U_1\, x$$2 for each atom $$h_{\text{out}} = U_{K+1}\, T_K(\varepsilon_K)\, U_K\, \ldots\, T_1(\varepsilon_1)\, U_1\, x$$3.
• Apply localized stimulus $$h_{\text{out}} = U_{K+1}\, T_K(\varepsilon_K)\, U_K\, \ldots\, T_1(\varepsilon_1)\, U_1\, x$$4; solve $$h_{\text{out}} = U_{K+1}\, T_K(\varepsilon_K)\, U_K\, \ldots\, T_1(\varepsilon_1)\, U_1\, x$$5.
• Insert central entries of $$h_{\text{out}} = U_{K+1}\, T_K(\varepsilon_K)\, U_K\, \ldots\, T_1(\varepsilon_1)\, U_1\, x$$6 into $$h_{\text{out}} = U_{K+1}\, T_K(\varepsilon_K)\, U_K\, \ldots\, T_1(\varepsilon_1)\, U_1\, x$$7.

Boundary patch fields are assumed near zero due to limited diffraction, minimizing artifacts. Patch solves cost $$h_{\text{out}} = U_{K+1}\, T_K(\varepsilon_K)\, U_K\, \ldots\, T_1(\varepsilon_1)\, U_1\, x$$8, so per layer cost is $$h_{\text{out}} = U_{K+1}\, T_K(\varepsilon_K)\, U_K\, \ldots\, T_1(\varepsilon_1)\, U_1\, x$$9 for fixed $U_i$0.

Adjoint-based optimization for inverse design uses only these patch solves, fully decoupling the full-size PDE into independently solvable subproblems.

5. Computational Efficiency and Scalability Implications

In naïve simulation-in-the-loop training, the computational cost for a $U_i$1-layer DONN with $U_i$2 epochs and $U_i$3 images is $U_i$4. The patch-based complementary reasoning approach reduces this to $U_i$5, with $U_i$6 epochs and $U_i$7 inverse-design projections per epoch. Since $U_i$8 and $U_i$9 are constant or small and inverse-design projections are infrequent, complexity is effectively linear in $$\min_{\varepsilon, w}~ \mathbb{E}_{(x, y)} \left[ \mathcal{L}\left(g_w(|U_{K+1}\prod_{i=1}^K T_i(\varepsilon_i) U_i x|^2), y \right) \right]\quad \text{subject to}~A(\varepsilon_i) h_i = b(x_i)$$0.

SP2RINT demonstrates a $$\min_{\varepsilon, w}~ \mathbb{E}_{(x, y)} \left[ \mathcal{L}\left(g_w(|U_{K+1}\prod_{i=1}^K T_i(\varepsilon_i) U_i x|^2), y \right) \right]\quad \text{subject to}~A(\varepsilon_i) h_i = b(x_i)$$1 acceleration compared to full simulation-in-the-loop for a 32-atom, 2-layer task. Scaling behavior is confirmed in runtime experiments: transfer matrix probing time is linear in $$\min_{\varepsilon, w}~ \mathbb{E}_{(x, y)} \left[ \mathcal{L}\left(g_w(|U_{K+1}\prod_{i=1}^K T_i(\varepsilon_i) U_i x|^2), y \right) \right]\quad \text{subject to}~A(\varepsilon_i) h_i = b(x_i)$$2 for SP2RINT, cubic for full FDFD. Trivial patchwise parallelism enables near-constant wall-clock time given sufficient hardware (Ma et al., 23 May 2025).

6. Significance for Physically Realizable Neural Hardware

Complementary reasoning, as instantiated in SP2RINT, resolves the conflict between fast, unconstrained model training and the necessity of final physical realizability by alternately optimizing in a relaxed space and projecting to a physical one. This dual space workflow bridges the gap between abstract digital models and hardware-constrained implementations, underpinning scalable, accurate, and fabricable meta-optical neural systems. The spatially decoupled, patchwise strategy, informed by the natural locality of field interactions in Maxwellian optics, enables the synthesis of complex, high-capacity optical networks without prohibitive computational costs.