MetaDiT: High-Fidelity Metasurface Design

• MetaDiT is a generative framework for metasurface inverse design that integrates a Diffusion Transformer with a contrastively pretrained spectrum encoder to produce all design parameters from full-resolution spectra.
• It employs in-context token-level conditioning and coarse-to-fine guidance to capture fine spectral details, achieving a remarkable MAE reduction to 0.0801 compared to baselines.
• Extensive ablations demonstrate that full-parameter generation and high-resolution spectral conditioning yield diverse, accurate metasurface designs, advancing inverse design methodologies.

Searching arXiv for the MetaDiT paper and nearby related work identifiers. MetaDiT is a generative framework for high-fidelity metasurface design that targets two limitations identified in prior machine learning-based inverse design methods: restriction to generating only a subset of design parameters, and reliance on heavily downsampled spectral targets. It is presented as a Diffusion Transformer-based system conditioned by a robust spectrum encoder pretrained with contrastive learning, with the stated aim of exploring a large, unconstrained design space while precisely capturing the physical relationships between material parameters and high-resolution spectral responses (Li et al., 7 Aug 2025). In the reported formulation, MetaDiT generates all relevant structural and material parameters and conditions directly on the full high-resolution scattering spectrum, thereby enabling fine-grained constraints in high-degree-of-freedom metasurface design (Li et al., 7 Aug 2025).

1. Problem setting and design objective

Metasurfaces are described as ultrathin, engineered materials composed of nanostructures that manipulate light in ways unattainable by natural materials (Li et al., 7 Aug 2025). Within this setting, the inverse-design problem is to synthesize a metasurface unit cell whose electromagnetic response matches a desired target spectrum. The paper frames the central difficulty as the need for a generative model that can simultaneously search a large, unconstrained design space and maintain fidelity to high-resolution spectral constraints (Li et al., 7 Aug 2025).

The representation used in MetaDiT makes this objective explicit. Each metasurface unit cell is specified by a binary pattern matrix of size $64\times64$, indicating the presence or absence of meta-atoms, together with three continuous design parameters: atom refractive index $r_{\text{atom}}$, atom thickness $h_{\text{atom}}$, and lattice constant $l_{\text{lattice}}$ (Li et al., 7 Aug 2025). The target scattering spectrum is represented as $S \in \mathbb{R}^{301\times2}$, corresponding to real and imaginary components sampled at 301 frequency points (Li et al., 7 Aug 2025). The paper identifies the use of the full $301\times2$ spectrum as a key distinction from approaches that downsample spectral supervision (Li et al., 7 Aug 2025).

A central claim of the method is “full-degrees-of-freedom design”: MetaDiT generates all relevant structural and material parameters rather than fixing some values in advance (Li et al., 7 Aug 2025). This suggests that the framework is intended not merely as a pattern synthesizer, but as a unified conditional generator over geometry and material variables.

2. Representation and overall architecture

MetaDiT maps the unit cell into a 3-channel image $U \in \mathbb{R}^{3\times64\times64}$, where each channel corresponds to one of the three continuous parameters and is spatially organized by the binary pattern matrix (Li et al., 7 Aug 2025). The architecture is then divided into three components: a spectrum encoder, a Diffusion Transformer backbone with fine-grained conditional injection, and classifier-free guidance during sampling (Li et al., 7 Aug 2025).

The spectrum encoder takes the target spectrum $S \in \mathbb{R}^{301\times2}$, linearly projects it to $X_S \in \mathbb{R}^{301 \times D_S}$, and processes it using both sequence self-attention and channel self-attention (Li et al., 7 Aug 2025). The sequence path models dependencies along the frequency axis, while the channel path models couplings between the real and imaginary channels (Li et al., 7 Aug 2025). The output is a sequence of spectrum tokens used for conditioning the generative model (Li et al., 7 Aug 2025).

The generative backbone is a Diffusion Transformer. Its conditioning mechanism is explicitly coarse-to-fine. At the fine level, projected spectrum tokens are concatenated with image tokens and processed jointly by self-attention, enabling token-level interaction between spectral features and material-design features (Li et al., 7 Aug 2025). At the coarse level, pooled spectrum information is added to the diffusion timestep embedding and injected through adaptive LayerNorm modulation (Li et al., 7 Aug 2025). The paper characterizes this as simultaneous use of global and fine spectrum features for hierarchical guidance (Li et al., 7 Aug 2025).

Classifier-free guidance is implemented by randomly dropping the spectral condition during training and combining conditional and unconditional predictions at sampling time (Li et al., 7 Aug 2025). This places MetaDiT within the standard conditional diffusion paradigm while preserving the paper’s emphasis on strong spectrum-conditioned generation (Li et al., 7 Aug 2025).

3. Spectrum encoder and contrastive pretraining

The spectrum encoder is one of the method’s principal innovations. It is described as “Seq+Channel Attentive” and contrastively pretrained (Li et al., 7 Aug 2025). Its encoder layer is given by

$$\begin{aligned} &X_l^{(1)} = X_l + \text{Attn}_\text{seq}(\text{Norm}(X_l)) \ &X_l^{(2)} = X_l^{(1)} + \left[ \text{Attn}_\text{chan}(\text{Norm}(X_l^{(1)})^{\mathsf{T}}) \right]^{\mathsf{T}} \ &X_{l+1} = X_l^{(2)} + \text{FFN}(\text{Norm}(X_l^{(2)})) \end{aligned}$$

where the two attention operators capture complementary structure in the spectrum: long-range dependencies across frequencies and interactions between real and imaginary components (Li et al., 7 Aug 2025).

Pretraining aligns spectral and structural representations using a CLIP-style contrastive objective. A ViT encoder processes the material image, and paired spectrum and structure embeddings are encouraged to be close in the joint representation space (Li et al., 7 Aug 2025). The loss is stated as

$r_{\text{atom}}$0

with $r_{\text{atom}}$1 and $r_{\text{atom}}$2 denoting $r_{\text{atom}}$3-normalized material and spectrum embeddings respectively (Li et al., 7 Aug 2025).

The reported ablations assign a decisive role to this component. Removing the pretrained spectrum encoder and replacing it with a plain MLP projector changes MAE from $r_{\text{atom}}$4 to $r_{\text{atom}}$5 and AAE from $r_{\text{atom}}$6 to $r_{\text{atom}}$7, corresponding to a relative degradation of $r_{\text{atom}}$8 in the table presented in the paper (Li et al., 7 Aug 2025). The paper interprets this as evidence that the encoder learns rich, physically meaningful spectral features (Li et al., 7 Aug 2025).

4. Diffusion Transformer conditioning and generative process

The Diffusion Transformer backbone uses both token-level and pooled spectral information. At the fine level, spectrum tokens and image tokens are concatenated and passed through full self-attention:

$r_{\text{atom}}$9

followed by extraction of the updated image-token subset,

$h_{\text{atom}}$0

This mechanism is described as in-context learning over the combined token sequence (Li et al., 7 Aug 2025). The paper explicitly contrasts this with explicit cross-attention and reports that the in-context variant performs better (Li et al., 7 Aug 2025).

At the coarse level, the spectrum tokens are pooled,

$h_{\text{atom}}$1

and combined with the timestep embedding,

$h_{\text{atom}}$2

to modulate the network through adaptive LayerNorm (Li et al., 7 Aug 2025). The reported ablation “w/o Coarse Condition” worsens MAE from $h_{\text{atom}}$3 to $h_{\text{atom}}$4 and AAE from $h_{\text{atom}}$5 to $h_{\text{atom}}$6, which the paper uses to support the value of multi-granular control (Li et al., 7 Aug 2025).

The diffusion objective follows the standard denoising formulation. Given structure $h_{\text{atom}}$7, the forward process generates

$h_{\text{atom}}$8

with $h_{\text{atom}}$9, and the model is trained using

$l_{\text{lattice}}$0

During sampling, classifier-free guidance is applied via

$l_{\text{lattice}}$1

These equations make clear that MetaDiT is not only a conditional generator but a spectrum-conditioned diffusion system with conditioning injected at multiple resolutions (Li et al., 7 Aug 2025).

5. Evaluation protocol, metrics, and benchmark results

The paper evaluates MetaDiT using Mean Absolute Error (MAE) and Accumulated Absolute Error (AAE), with AAE defined as

$l_{\text{lattice}}$2

It also introduces $l_{\text{lattice}}$3 as the maximum AAE over $l_{\text{lattice}}$4 independently generated designs for the same target:

$l_{\text{lattice}}$5

The stated purpose of $l_{\text{lattice}}$6 is to evaluate both diversity and consistent accuracy (Li et al., 7 Aug 2025).

The main benchmark table reports the following results (Li et al., 7 Aug 2025):

Model	# Param	MAE
Baseline (AVG)	-	0.5860
MetaDiff	32.56M	0.1861
MetaDiff-HR	33.41M	0.1315
DiT (single vec)	32.80M	0.1677
MetaDiT	32.57M	0.0801

A second view of the same benchmark, emphasizing integrated spectral error, is as follows (Li et al., 7 Aug 2025):

Model	AAE	AAE$l_{\text{lattice}}$7
Baseline (AVG)	352.74	352.7424
MetaDiff	112.06	170.63
MetaDiff-HR	79.14	100.25
DiT (single vec)	100.94	138.07
MetaDiT	48.25	58.80

The paper further reports $l_{\text{lattice}}$8 for MetaDiT, compared with $l_{\text{lattice}}$9 for MetaDiff-HR and $S \in \mathbb{R}^{301\times2}$0 for the DiT single-vector baseline (Li et al., 7 Aug 2025). It states that MetaDiT reduces error by $S \in \mathbb{R}^{301\times2}$1 relative to the DiT baseline and by $S \in \mathbb{R}^{301\times2}$2 relative to MetaDiff-HR (Li et al., 7 Aug 2025).

Qualitatively, MetaDiT is said to produce accurate structures and spectra, including fine features in the target spectrum, while also generating diverse viable structures for a given target (Li et al., 7 Aug 2025). Since these qualitative judgments are summarized rather than formally quantified beyond the reported metrics, a plausible implication is that the proposed conditioning scheme helps preserve spectral detail without collapsing diversity.

6. Ablations, comparative claims, and significance

The ablation study isolates three design choices: the pretrained spectrum encoder, coarse condition injection, and the use of in-context self-attention rather than explicit cross-attention (Li et al., 7 Aug 2025). The full default configuration yields MAE $S \in \mathbb{R}^{301\times2}$3 and AAE $S \in \mathbb{R}^{301\times2}$4; removing the pretrained encoder yields MAE $S \in \mathbb{R}^{301\times2}$5 and AAE $S \in \mathbb{R}^{301\times2}$6; removing the coarse condition yields MAE $S \in \mathbb{R}^{301\times2}$7 and AAE $S \in \mathbb{R}^{301\times2}$8; and replacing in-context attention with cross-attention yields MAE $S \in \mathbb{R}^{301\times2}$9 and AAE $301\times2$0 (Li et al., 7 Aug 2025).

The paper states that replacing in-context conditioning with explicit cross-attention increases parameters by $301\times2$1 while reducing performance (Li et al., 7 Aug 2025). It also reports that simply scaling backbone size produces only minimal improvements of $301\times2$2–$301\times2$3, and can reduce performance beyond a threshold (Li et al., 7 Aug 2025). These results are presented as evidence that conditioning design and representation learning matter more than vanilla model scaling in this application (Li et al., 7 Aug 2025).

Several comparative claims follow directly from these results. First, MetaDiT is positioned against methods that either generate only part of the design space or rely on lower-resolution spectral supervision (Li et al., 7 Aug 2025). Second, within the paper’s benchmark, the strongest numerical comparison is against MetaDiff-HR, which preserves high-resolution targets but does not match the reported spectral accuracy of MetaDiT (Li et al., 7 Aug 2025). Third, the DiT single-vector baseline serves to isolate the effect of richer conditioning: the improvement from MAE $301\times2$4 to $301\times2$5 suggests that the benefit is not merely due to using a diffusion transformer, but to the particular spectrum encoding and conditioning path (Li et al., 7 Aug 2025).

The paper characterizes MetaDiT as the first framework to combine a contrastively trained spectrum encoder, in-context token-level conditional injection, and a DiT backbone for fine-grained, high-resolution, all-parameter metasurface inverse design (Li et al., 7 Aug 2025). Because that claim is presented in the source as a summary contribution, it should be read as a contribution statement rather than an independently verified field-wide taxonomy.

7. Position within metasurface inverse design

MetaDiT is situated in a line of work that applies computational optimization, machine learning, and deep learning to automate metasurface design (Li et al., 7 Aug 2025). The paper’s specific critique of prior approaches has two parts: incomplete parameter generation and downsampled spectral targets (Li et al., 7 Aug 2025). MetaDiT addresses the first by generating all cell and geometric parameters, and the second by conditioning directly on the full scattering spectrum at 301 frequency points (Li et al., 7 Aug 2025).

Its technical identity is therefore defined by three linked properties. The first is full-parameter generation over a high-degree-of-freedom design space (Li et al., 7 Aug 2025). The second is fine-grained spectral conditioning through token-level fusion between spectrum tokens and image tokens (Li et al., 7 Aug 2025). The third is hierarchical guidance through simultaneous fine and coarse injection of condition information (Li et al., 7 Aug 2025). Together, these design decisions are presented as the basis for fine-grained constraints in metasurface inverse design (Li et al., 7 Aug 2025).

The paper also states that code and model weights will be open-sourced to facilitate future research (Li et al., 7 Aug 2025). This suggests an intended role for MetaDiT not only as a point method but as a reusable framework for subsequent work on high-fidelity generative design of photonic structures.