Organizations: Department of Mechanical Engineering, Northwestern University, Evanston, 60208, IL, USA. · Graduate School of Artificial Intelligence, Pohang University of Science and Technology, Pohang, 37673, Republic of Korea. · Department of Mathematics, Lehigh University, Bethlehem, 18015, PA, USA. · Department of Electrical and Computer Engineering, Seoul National University, Seoul, 08826, Republic of Korea. · Department of Mechanical Engineering, Pohang University of Science and Technology, Pohang, 37673, Republic of Korea. · Department of Chemical Engineering, Pohang University of Science and Technology, Pohang, 37673, Republic of Korea. · Department of Electrical Engineering, Pohang University of Science and Technology, Pohang, 37673, Republic of Korea. · POSCO-POSTECH-RIST Convergence Research Center for Flat Optics and Metaphotonics, Pohang, 37673, Republic of Korea.
In multiscale systems, nonlocal interactions among spatially distributed constituents can unlock collective, potentially emergent behavior irreducible to individual constituent responses. Beyond characterization, nonlocality in multiscale systems is increasingly viewed as a design element across fracture mechanics, wave physics, and disordered systems. Existing nonlocality theories provide rigorous descriptions and mechanistic insight, but often require prescribed analytic forms, range, symmetries, or periodicity. Machine learning offers a data-driven route to learn interactions, yet principled approaches that tie the network design to physical theory remain nascent. Here we present domain-decomposition attention neural operator (DANO), which learns configuration-dependent interactions in heterogeneous systems governed by linear field equations from field data alone. Through a case study in large-scale optical metasurfaces, we demonstrate DANO's full-field prediction under geometric heterogeneity and across drive conditions, small-to-large domain transfer, and multifunctional inverse design, with learned interactions validated through physics-based interrogation. As an AI-for-Science tool, it supports interaction discovery and scalable, customizable design of nonlocal multiscale systems, including virtual-reality optics.
Figures & tables
Figure 1: Nonlocal constituent interactions and windowed prediction of a heterogeneous metasurface. Schematics and fields are illustrative. a , Nonlocal interactions among distributed constituents in polymer-like networks, active particle assemblies, disordered wave media and optical metasurfaces. In each panel, the interaction of one constituent (black) with the others extends beyond its nearest neighbors (blue, darker for a stronger interaction). Dotted circles mark an interaction range, drawn with the same radius in every panel. Arrows give the propulsion directions of the active particles. b , Windowed prediction of the response field across a large metasurface using the domain-decomposition attention neural operator (DANO). Tan squares mark context windows, each containing an interior prediction core (red). The same DANO uses the geometry and drive within each window to predict the response field on its core. Partition-of-unity (PoU) assembly combines the core predictions into a field over the full array.
Figure 2: Field-mediated interactions among constituents and their representation in DANO. Box colors connect the physical operations in a with their learned counterparts in b : light blue, the drive and propagation between constituents; red, constituent response and its configuration-conditioned representation; orange, mapping to observation points; green, the scattered field. a , In electromagnetic scattering, the background field Eb and fields propagated from other constituents form the exciting field Eexc . Each constituent responds through T , and Gout maps those responses to the scattered field Es . b , For one prediction window, the geometry path encodes the constituent configuration independently of the drive. The field path repeatedly transports states between constituents, reinjects the drive and applies a configuration-conditioned response. A kernel-sum decoder maps the resulting states to the scattered field at observation points. Red arrows indicate operations that are complex-linear in the field. × 8 and × 2, numbers of stacked layers.
Figure 3: Full-field prediction accuracy across configurations and domain sizes. a , In-distribution prediction on held-out random layouts under plane-wave illumination at θ=0∘ , 20∘ and 40∘ (rows). Columns show the layout (black, SiN pillars), the ground truth from finite-difference time-domain simulation, the domain-decomposition attention neural operator (DANO) prediction and their error (DANO minus ground truth) for Re Es,x , normalized by the root-mean-square incident-field amplitude. b , Structured-layout transfer to a 33×33 hyperbolic metalens with a 10-µm focal length, from a layout family absent from training. Rows show the same incidence angles, and columns follow a . c , Domain-size transfer from the 33×33 -site training size to 64×64 and 128×128 sites at normal incidence, using one independently drawn layout at each size. The blue frame marks the training-array size; blue squares connected by dotted guides mark its 9.9-µm width for scale. d , Normalized mean squared error (NMSE) of the assembled field against the ground truth for one layout at each array size. e , Mean test NMSE of DANO and the neural Green’s operator (NGO), multiple-input operator network (MIONet), Fourier neural operator (FNO) and U-Net baselines, all trained on the same windows. Group labels give training corpus → test illumination, with pw for plane waves and spk for speckle. Each bar is one training run evaluated on 213 plane-wave or 600 speckle test cases.
Figure 4: Physics-based interrogation of the learned response operator. Panels a – c vary the drive at fixed layout. Panels d – f change both side lengths of one pillar by ±20 nm at fixed drive. Distances in a , b and d – f are in units of the constituent spacing p=300 nm. a , Angular contrast of the cross-polarized ( x→y ) response to a localized diagnostic drive, defined as response magnitude divided by its angular mean at each radius. Left, DANO; right, a multipole model of pillar radiation. Maps are means over 71 validation windows. b , Radial decay of the co-polarized ( x→x ) response to the same drive. Blue circles and black line show the DANO and multipole-model medians across the same 71 windows, respectively. c , Commutation residual of the learned drive-to-field operator with the phase-ramp operator Mk , which multiplies a field by eikxx . The free-space wavenumber is k0 . Shading, beyond the plane-wave training coverage. d , Pillar footprints (top) and configuration-response magnitudes (bottom, common logarithmic scale) for disordered, uniform and metalens layouts at normal incidence. Responses are averaged over 361, 361 and 233 perturbed sites, respectively. Crosses mark perturbed sites; small circles, lattice sites; dashed circles, the 10% radii in e . e , Radial profiles of d , normalized at the perturbed site, with an isolated pillar on the bare substrate as a reference. Dashed line, 10%. f , Truncation error ϵ(R) , the relative L2 norm of the response omitted beyond radius R . Black dashed curve, drive response. Gray dashed lines, 30% and 10%.
Figure 5: Large-scale inverse design and scaling of full-field evaluation. a , Far-field intensity from two 512×512 -site layouts (153.6 µm) designed to display the POSTECH seal at θ=0∘ and the Northwestern logo at 20∘ (rows). The displayed region follows the zero order at each angle. Columns show the target, the local design evaluated by the local model and by the ground-truth solver (finite-difference time-domain simulation), and the domain-decomposition attention neural operator (DANO) design evaluated by DANO and the ground-truth solver. The local model assigns each site the angle-dependent transmission of a periodic unit cell. Each column is normalized independently. b , Time per evaluation versus the number of meta-atoms. Open stars show DANO times on larger arrays not simulated with the ground-truth solver. Ground-truth solver and DANO times were measured on a 64-thread node and one GPU, respectively. Dark filled circles and open squares mark full-wave solvers and learned surrogates in the literature, and open triangles mark the sizes of fabricated devices, with reference numbers beside them. Dashed lines connect the times from this work, and dotted lines extrapolate power-law fits to them. The gray band spans the interval between the largest array with whole-array ground truth and the smallest fabricated device. The top axis gives the side length of a square array at a 300-nm pitch. Insets show the hologram in a and two fabricated devices: a 300-mm wafer-scale metalens array with a scanning electron micrograph of its nanostructures (ref. 25 ; scale bar, 1 µm) and an astronomical imager built around a 100-mm metalens (ref. 64 ). Insets in b reproduced from refs. 25 and 64 , with permission from Springer Nature and under a CC BY 4.0 license, respectively.
Extended Data Fig. 1: Full-wave data generation and window-level illumination diversity. a , Simulation stack with a total-field/scattered-field (TFSF) source plane and box, observation plane zobs and perfectly matched layer (PML) boundaries. Vertical distances are not to scale. b , Real part of the background field Eb for one speckle realization and a plane wave at θ=20∘ , each normalized by its 99.5th-percentile magnitude. Solid squares mark the three 1.24 -µm windows of d (scale bar, 2 µm). c , Background-field power spectrum averaged over 300 speckle realizations (logarithmic color scale). Dashed and dotted circles, air and substrate light cones; open circles, the 33 plane-wave settings. d , Background fields in the three windows of b . e , 99%-energy dimension, the number of modes holding 99% of the background-field energy in local windows, versus the number of ground-truth simulations. f , Normalized Karhunen–Loève mode spectra of speckle background fields in fixed windows and in windows with random offsets. Dotted lines, the 99%-energy dimensions.
Extended Data Fig. 2: Phase mismatch between the local phase approximation and the ground truth. One held-out, fully occupied random layout under normal-incidence illumination, evaluated on the observation plane. a , Phase of the transmitted field Ex from the local phase approximation (LPA, computed by a local-phase lookup, left), the ground truth from finite-difference time-domain simulation (middle) and domain-decomposition attention neural operator (DANO, right). The black line marks the row examined in b . b , Phase along the marked row. c , Distributions of circular phase error relative to the ground truth.
Extended Data Fig. 3: Full-field prediction on finer simulation meshes. The domain-decomposition attention neural operator (DANO) trained on a 20-nm mesh is evaluated on 12.5-nm and 9.375-nm meshes without retraining. a , One fixed, learned first-layer discrete–continuous (DISCO) kernel on its 60-nm-radius support (dashed circle), shown as a continuous function (left) and sampled at the grid points (dots) of the 20-, 12.5- and 9.375-nm meshes. b , One layout at normal incidence on the 20-, 12.5- and 9.375-nm meshes (rows). Columns show the ground truth from finite-difference time-domain simulation, DANO prediction and their error (DANO minus ground truth) for Re Es,x . The blue frame marks the 20-nm training mesh. c , Permittivity of one pillar from b on the three meshes. Gray grid cells show fractional pillar coverage. d , Mean normalized mean squared error (NMSE) over 50 layouts at each mesh spacing.
Extended Data Fig. 4: Independent window predictions and partition-of-unity assembly. a , Layout with two overlapping input windows (gold and teal). Inner squares mark the 1.24 -µm prediction cores, and outer squares include a 2.1 -µm context halo on each side. Dotted square, the permittivity crop of the gold window. b , Nine overlapping prediction cores, evaluated independently in their respective windows and separated for display before assembly. c , Validation core normalized mean squared error (NMSE) against context-halo width in constituent spacings, with one training run per width. Points are means over 600 speckle or 71 plane-wave cases. Vertical dashed line, the selected halo of 7 spacings (2.1 µm). d , Assembled-field error of the domain-decomposition attention neural operator (DANO), DANO minus the ground truth from finite-difference time-domain simulation. Grid lines mark core boundaries, and the black square marks the region enlarged in e . e , Assembled field in that region, with colored outlines marking overlapping prediction cores. Dashed horizontal line, the cut in f . f , Independent core predictions (top), partition-of-unity weights summing to one (middle) and the assembled field (bottom) along the cut. Shading, the overlap.
Extended Data Fig. 5: Speckle test error by source band. Errors of the domain-decomposition attention neural operator (DANO) on held-out speckle test cases. Source bands group the angular components of the speckle source by f=∣k∥∣/k0 , where k∥ is the in-plane wavevector and k0 is the free-space wavenumber. a , Normalized mean squared error (NMSE) split into f≤1 , f>1 and cross terms. b , Contribution of each of eight bands to the NMSE, normalized as in a , without cross terms between bands. c , NMSE for plane waves (pw), averaged over all incidence angles and at 40∘ , and for the f≤1 part of the speckle illumination (spk). d , NMSE of each band alone, normalized by the scattered-field energy of that band. Dotted lines in b , d , f=1 . Bands of unequal width are drawn at equal spacing. Error bars, 95% bootstrap intervals over 32 realizations with two layouts each ( a and spk in c ) or over 213 cases (pw mean in c ), and interquartile ranges over 12 realizations with one layout each ( b , d ).
Extended Data Fig. 6: Scattered-field prediction under Gaussian-beam illumination. One held-out layout is illuminated by normally incident, x -polarized Gaussian beams with waists w0 of 0.5, 1, 1.5 and 2 µm. The domain-decomposition attention neural operator (DANO), U-Net, Fourier neural operator (FNO), multiple-input operator network (MIONet) and neural Green’s operator (NGO) were trained on the plane-wave and speckle corpora. a , Rows show beam waist. Columns show the background field Eb,x , the ground-truth scattered field Es,x from finite-difference time-domain simulation, and each model’s prediction and error (prediction minus ground truth). Real parts are normalized by the peak background-field amplitude in each row. b , Normalized beam angular spectra against in-plane wavevector. Black triangles mark the plane-wave training settings. c , Prediction-core normalized mean squared error (NMSE) against beam waist for DANO and the four baselines.
Extended Data Fig. 7: Full-field prediction for uniform periodic arrays. Eight fully occupied 33×33 arrays, each repeating one pillar cross-section on every site, are illuminated by a normally incident plane wave, with one array per cross-section. a , Arrays with the smallest, median and largest normalized mean squared error (NMSE) of the prediction (rows), with the cross-section at left. Columns show the layout, the ground truth from finite-difference time-domain simulation, the domain-decomposition attention neural operator (DANO) prediction and their error (DANO minus ground truth), followed by DANO ×s , the prediction multiplied by the single complex number s that best fits the ground truth, and its error. All field maps show Re Es,x , normalized by the root-mean-square incident-field amplitude on one color scale. b , NMSE of the assembled scattered field for the eight arrays, before (open circles) and after (filled circles) multiplication by s . Dashed line, mean NMSE of the 21 normal-incidence plane-wave test cases on random layouts.
The Pennsylvania State University, Department of Electrical Engineering, University Park, PA 16802, USA · School of Electronics and Information Technology, Sun Yat-Sen University, Guangzhou 510275, China
Image Processing Group – Signal Theory and Communications Department, Universitat Politècnica de Catalunya, Barcelona, Spain. · Inorganic and Organic Chemistry Department and Institute of Theoretical and Computational Chemistry, Universitat de Barcelona, Barcelona, Spain. · Department of Physics and Materials Science, University of Luxembourg, Luxembourg City, Luxembourg.