Metasurface-based Terahertz Three-dimensional Holography Enabled by Physics-Informed Neural Network

This paper proposes a physics-informed neural network (LM-PINN) that enables rapid, label-free, and universal design of high-quality terahertz 3D holographic metasurfaces, overcoming the speed and adaptability limitations of traditional iterative and existing deep-learning methods.

The Gist

The Big Idea: Teaching a Computer to "Dream" in 3D

Imagine you want to create a hologram—a 3D image floating in mid-air, like the Princess Leia message in Star Wars, but using invisible "Terahertz" waves (which are like super-powered radio waves).

In the past, designing the physical surface needed to create these holograms was like trying to solve a massive jigsaw puzzle while blindfolded. You had to guess, check, guess again, and wait hours or even days for the computer to figure out the right shape for every tiny piece.

This paper introduces a new "super-smart" designer called LM-PINN. It's a type of Artificial Intelligence that doesn't just guess; it understands the laws of physics. It can design a perfect 3D hologram surface in less than a second.

---

The Problem: The Old Way Was Too Slow and Rigid

Think of the old method (called the Gerchberg-Saxton algorithm) like a very stubborn sculptor.

• The Process: The sculptor tries to carve a statue. They look at the result, realize it's wrong, chisel a bit more, look again, and repeat.
• The Issue: If the statue is simple, it takes a while. If the statue is complex (like a 3D model of a plane), the sculptor gets stuck, takes forever, and often gives up or makes a blurry mess.
• The Rigidity: If you wanted to move the statue to a different spot in the room, the sculptor had to start from scratch and learn the new position all over again.

The Solution: The "Physics-Savvy" AI

The authors created a new AI called LM-PINN. Think of this AI as a master architect who has memorized the laws of gravity and light.

Instead of guessing and checking, this architect looks at the picture you want to see (the target hologram) and instantly knows exactly how to carve the surface to make it appear.

Here is how it works, broken down into simple parts:

1. The "Local Map" (Local Polynomial Fitting)

To design the hologram, the AI needs to know how a tiny piece of silicon (a "meta-atom") bends light.

• The Old Way: The computer would run a massive, slow simulation for every single tiny piece to see how it behaves.
• The New Way: The AI uses a "Local Map." Imagine you are trying to describe the shape of a hill. Instead of measuring every single grain of sand, you divide the hill into small neighborhoods. In each neighborhood, you draw a simple curve that fits the shape perfectly. The AI does this for the light-bending properties, making it incredibly fast and accurate without needing a supercomputer to run simulations every time.

2. The "Self-Taught" Student (Self-Supervised Learning)

Usually, AI needs a teacher with a stack of answer keys (labeled data) to learn. But in the world of Terahertz waves, getting "answer keys" is hard because the equipment is expensive and rare.

• The Trick: This AI is a self-taught student. It doesn't need a teacher. It looks at the picture you want (e.g., the number "2") and tries to build the surface. It then simulates the light passing through its design. If the result looks like the number "2," it gets a gold star. If it looks like a "7," it learns from its mistake and tries again. It keeps doing this until it gets it perfect, all without needing a pre-made dataset.

3. The "Universal Remote" (Distance Encoding)

This is the coolest part.

• The Old AI: If you trained an AI to make a hologram appear 3 inches away, and then you wanted it to appear 10 inches away, you had to retrain the whole AI. It was like having a remote control that only works for one TV channel.
• The New AI (Dist-LM-PINN): The authors gave the AI a "Universal Remote" feature. They taught it that "distance" is just another number to pay attention to. Now, you can tell the AI, "Make a hologram of a plane, but put it 3 inches away," or "Put it 20 inches away," or even "Make a 3D model that floats in mid-air." The AI doesn't need to retrain; it just adjusts its settings instantly.

What Did They Prove?

The team didn't just write code; they built the real thing.

1. They designed a silicon surface with thousands of tiny pillars using their AI.
2. They built it using lasers and etching machines.
3. They tested it with a special Terahertz camera.

The Results:

• Speed: The AI designed the surface in less than 1 second. The old method took minutes or hours.
• Quality: The images were sharp and clear. The old method produced blurry, broken images.
• Versatility: They used the same trained AI to make:
  • Simple numbers (2, 4, 7, 8).
  • Images at different distances.
  • A complex 3D model of an airplane.
  • A multi-focal lens (like a camera lens that focuses on two things at once).

The Bottom Line

This paper is like upgrading from a hand-cranked calculator to a smartphone.

Before, designing 3D holograms for Terahertz waves was slow, difficult, and limited to simple shapes. Now, thanks to this "Physics-Informed" AI, we can design complex, high-quality 3D holograms instantly. This opens the door for real-time 3D displays, better security scanners, and advanced medical imaging that we can actually use in the real world.

Technical Summary

1. Problem Statement

The design of terahertz (THz) holographic metasurfaces, particularly for complex 3D applications, faces significant challenges:

• Limitations of Traditional Algorithms: Conventional iterative methods like the Gerchberg-Saxton (GS) algorithm suffer from slow convergence, high computational costs, and stagnation as target complexity increases. They typically handle phase-only modulation, neglecting amplitude, which degrades imaging quality for 3D and multi-color displays.
• Limitations of Existing Deep Learning (DL) Models: While supervised deep neural networks (DNNs) offer speed, they act as "black boxes" lacking physical interpretability. More critically, they are constrained by the "one-model-one-scenario" limitation; a model trained for a specific diffraction distance or target type cannot generalize to new physical configurations without expensive retraining.
• Data Scarcity: Obtaining large-scale, high-quality labeled experimental datasets in the THz band is difficult due to hardware limitations, making supervised learning approaches less viable.

2. Methodology

The authors propose a Physics-Informed Neural Network (PINN) framework called LM-PINN (Local Polynomial fitting and Multi-plane wave propagation PINN), which evolves into Dist-LM-PINN via distance encoding.

Core Architecture

The framework consists of four synergistic modules:

1. Input Module: Accepts target holographic volumes (2D or 3D).
2. Network Module (Y-net): A Y-shaped convolutional neural network that takes the target information and outputs geometric parameter maps (length $l$ and width $w$) for the metasurface meta-atoms.
3. Physical Module (Differentiable Surrogate):
   • Local Polynomial Fitting: Instead of using a separate forward prediction network or computationally heavy Full-Wave simulations (FDTD) during training, the authors use a differentiable local polynomial fitting model. This maps the meta-atom geometry ($l, w$) to its complex-amplitude response (phase and transmission) with high accuracy (NMSE < $10^{-6}$).
   • Angular Spectrum Method (ASM): A differentiable wave propagation algorithm used to simulate light propagation from the metasurface plane to multiple target planes ($z = d_1, d_2, \dots$).
4. Output Module: Generates the final reconstructed holographic volume.

Training Strategy: Self-Supervised Learning

• No Labeled Data Required: The network is trained using a self-supervised strategy. The loss function compares the reconstructed volume (generated by the physical module) with the target volume.
• Loss Function: A composite loss function minimizes the difference in image shape (using MSE and Normalized Pearson Correlation Coefficient) and maximizes light efficiency.
• Distance Encoding: To overcome the "one-model-one-scenario" limitation, the input includes distance encoding. The model is trained on single-plane data with random diffraction distances ($3\text{--}20$ mm). This allows the network to learn the mapping between the target electric field (which encodes distance information) and the metasurface structure, enabling zero-shot generalization to new distances and 3D configurations without retraining.

3. Key Contributions

• LM-PINN Framework: A novel self-supervised, physics-informed architecture that eliminates the need for paired input-label datasets and replaces computationally expensive forward simulators with differentiable local polynomial fitting.
• Universal Generalization (Dist-LM-PINN): By integrating distance encoding, a single trained model can design metasurfaces for varying diffraction distances, different 2D/3D targets, and multi-focal lenses, breaking the rigid constraints of previous DL models.
• Complex-Amplitude Control: Unlike phase-only algorithms, LM-PINN directly outputs geometric parameters ($l, w$), enabling independent control of both phase and amplitude, which is crucial for high-fidelity 3D holography.
• Experimental Validation: The method was validated not only through FDTD simulations but also via the fabrication and experimental testing of silicon-based metasurfaces using a THz Quantum Cascade Laser (QCL).

4. Results

• Imaging Quality:
  • Single-Plane: LM-PINN significantly outperformed the GS algorithm. For target "2" and "4", LM-PINN achieved NPCC scores of 0.971/0.964 vs. GS's 0.778/0.804, and PSNR of 17.9/19.6 dB vs. 10.8/13.2 dB.
  • Multi-Plane & Dual-Plane: LM-PINN successfully reconstructed identical images at three different distances and distinct images ("8" and "7") at two distances. In contrast, the GS algorithm failed to maintain pattern continuity in the dual-plane scenario.
  • 3D Holography: The model successfully designed metasurfaces for complex 3D models (e.g., an aircraft), reconstructing 15 cross-sections with high fidelity.
• Experimental Performance: Fabricated metasurfaces demonstrated high-fidelity reconstruction of the digits "2" and "4" at 2.58 THz. Experimental metrics (e.g., NPCC ~0.95) closely matched simulations, with minor degradation attributed to background noise and optical path errors.
• Computational Efficiency:
  • Inference Speed: LM-PINN inference takes < 1 second (approx. 0.5s on GPU, ~1.1s on CPU) for a 512×512 array.
  • Comparison: This is orders of magnitude faster than the GS algorithm, which requires thousands of iterations and takes significantly longer (e.g., ~14.9s for 512×512 on CPU).
• Versatility: A single Dist-LM-PINN model trained on 64×64 MNIST data successfully generalized to:
  • Different target distances (3mm to 20mm+).
  • Larger resolutions (512×512, e.g., Shanghai Jiao Tong University emblem).
  • Multi-focal metalenses.
  • Complex 3D volumetric holograms.

5. Significance

This work establishes a robust, universal framework for the inverse design of THz holographic metasurfaces. By combining physics-informed constraints with deep learning, it solves the critical bottlenecks of speed, data dependency, and generalizability.

• Real-Time Applications: The sub-second inference time paves the way for real-time, dynamic 3D holographic displays in the THz band.
• Scalability: The ability to design large-area (512×512) and complex 3D structures without retraining makes the technology scalable for practical deployment in communications, security, and augmented reality.
• Methodological Impact: The strategy of using local polynomial fitting as a differentiable surrogate and distance encoding for generalization offers a new paradigm for AI-driven photonic device design beyond just THz holography.