Metasurface-based Terahertz Three-dimensional Holography Enabled by Physics-Informed Neural Network
Pith reviewed 2026-05-16 17:39 UTC · model grok-4.3
The pith
A physics-informed neural network designs terahertz metasurfaces for 3D holography that generalize across distances after one training.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
LM-PINN performs end-to-end inverse design of metasurfaces by embedding physical wave propagation models into the network architecture, allowing self-supervised optimization that yields metasurface phase profiles capable of reconstructing complex 3D holographic targets with improved fidelity and broad generalization to new physical parameters.
What carries the argument
LM-PINN, a neural network that integrates local polynomial fitting to approximate phase distributions and multi-plane wave propagation to simulate diffraction, trained self-supervised to learn the mapping from target fields to metasurface designs.
Load-bearing premise
The self-supervised loss based on local polynomial fitting and multi-plane wave propagation fully captures the actual physical behavior of the metasurface-wave interaction.
What would settle it
Measure the holographic reconstruction error of a fabricated LM-PINN-designed metasurface at an untrained distance and compare it to the network's prediction and to designs from iterative methods.
read the original abstract
Artificial intelligence has revolutionized optical device design, overcoming the efficiency bottlenecks of traditional methods. For holographic metasurfaces, conventional iterative algorithms suffer from time-consuming iterations and convergence stagnation, especially as the complexity of 3D target fields increases. While recent deep-learning-based algorithms have improved the trade-off between speed and image quality, most existing models remain constrained by predefined physical scenarios (e.g., fixed distances), limiting their adaptability in dynamic practical applications. To address these challenges, we propose a physics-informed neural network (PINN) based on local polynomial fitting and multi-plane wave propagation (LM-PINN) for the rapid design of terahertz 3D holographic metasurfaces. By leveraging a self-supervised training strategy, LM-PINN eliminates the need for labeled datasets, enabling direct end-to-end mapping from target holographic patterns to the metasurface structures. Both simulated and experimental results demonstrate that LM-PINN-designed metasurfaces offer higher imaging quality than traditional iterative algorithms. Crucially, by incorporating a distance encoding process, a single trained LM-PINN generalizes effectively across diverse physical configurations, including varying diffraction distances and distinct 2D or 3D targets, eliminating the necessity for retraining. Furthermore, the inference process of LM-PINN typically takes less than 1 second, providing a multifold speed advantage over traditional algorithms. Consequently, this strategy offers a robust and universal framework that paves the way for high-quality, real-time, and large-scale 3D holographic technologies.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper introduces LM-PINN, a physics-informed neural network using local polynomial fitting and multi-plane wave propagation for end-to-end design of terahertz metasurface-based 3D holograms. It employs self-supervised training to map target fields directly to metasurface phase profiles without labeled data, claims superior imaging quality over iterative algorithms in both simulation and experiment, and asserts that a single trained model generalizes across varying diffraction distances and 2D/3D targets via a distance-encoding process, with inference under 1 second.
Significance. If the generalization and physics-embedding claims hold, the work would represent a meaningful advance in adaptive, real-time metasurface holography by removing the need for per-scenario retraining or slow iterative optimization, enabling practical deployment in dynamic THz imaging applications where fixed-distance DL models fail.
major comments (2)
- [Abstract and results] The central generalization claim (single trained LM-PINN works across diffraction distances and distinct targets) rests on the distance-encoding process plus self-supervised loss; however, no quantitative bounds on extrapolation error or out-of-distribution tests beyond the (presumably narrow) training range are provided, leaving open the possibility that phase errors degrade performance below iterative baselines for unseen distances.
- [Methods] The self-supervised training strategy (local polynomial fitting combined with multi-plane wave propagation) is asserted to accurately embed the underlying physics without labeled data; yet the manuscript provides no direct validation (e.g., comparison of predicted fields against full-wave FDTD simulations at specific points) that the loss captures metasurface scattering and propagation effects for complex 3D targets.
minor comments (2)
- [Figures] Figure captions and axis labels should explicitly state the quantitative metrics (e.g., PSNR, SSIM, or normalized MSE) used to claim 'higher imaging quality' so that the improvement over iterative baselines can be assessed directly.
- [Methods] The training distance range and the precise form of the distance-encoding input (concatenation, embedding layer, etc.) should be stated explicitly in the methods to allow reproducibility of the generalization experiments.
Simulated Author's Rebuttal
We thank the referee for the constructive feedback on our manuscript. The comments highlight important aspects of the generalization claims and the validation of the physics-informed loss. We address each point below and will incorporate revisions to strengthen the presentation of results.
read point-by-point responses
-
Referee: [Abstract and results] The central generalization claim (single trained LM-PINN works across diffraction distances and distinct targets) rests on the distance-encoding process plus self-supervised loss; however, no quantitative bounds on extrapolation error or out-of-distribution tests beyond the (presumably narrow) training range are provided, leaving open the possibility that phase errors degrade performance below iterative baselines for unseen distances.
Authors: We agree that explicit quantitative bounds on extrapolation would better support the generalization claim. Our current experiments demonstrate effective performance across a range of distances and targets using the distance-encoding mechanism, but we acknowledge the value of out-of-distribution testing. In the revised manuscript, we will add new results showing error metrics (e.g., RMSE and phase error) for diffraction distances outside the training distribution, along with direct comparisons to iterative algorithms to confirm that imaging quality remains superior. revision: yes
-
Referee: [Methods] The self-supervised training strategy (local polynomial fitting combined with multi-plane wave propagation) is asserted to accurately embed the underlying physics without labeled data; yet the manuscript provides no direct validation (e.g., comparison of predicted fields against full-wave FDTD simulations at specific points) that the loss captures metasurface scattering and propagation effects for complex 3D targets.
Authors: We appreciate this suggestion for additional validation. The self-supervised loss is constructed directly from the multi-plane wave propagation model and local polynomial fitting to enforce physical consistency without requiring labeled data. To address the request for direct validation, the revised manuscript will include comparisons between the fields predicted by the trained LM-PINN and full-wave FDTD simulations for representative complex 3D targets, quantifying the agreement in both amplitude and phase at selected planes. revision: yes
Circularity Check
Low circularity: self-supervised physics embedding in LM-PINN does not reduce to fitted inputs or self-citations
full rationale
The derivation chain relies on a self-supervised loss that directly incorporates local polynomial fitting and multi-plane wave propagation operators to enforce diffraction physics during training. This embeds the forward model as an independent constraint rather than deriving the metasurface mapping from pre-fitted parameters or author-specific prior equations. The distance encoding is presented as an input feature enabling generalization, not a definitional tautology that forces the output. No load-bearing step (e.g., the end-to-end mapping or the claimed superiority over iterative algorithms) reduces by construction to the training inputs; the central claim retains independent content verifiable against external wave-propagation benchmarks. Minor self-citation risk is possible in the broader literature but is not load-bearing here.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.