Wavelength-multiplexed massively parallel diffractive optical information storage and image projection

Abstract: We introduce a wavelength-multiplexed massively parallel diffractive information storage platform composed of dielectric surfaces that are structurally optimized at the wavelength scale using deep learning to store and project thousands of distinct image patterns, each assigned to a unique wavelength. Through numerical simulations in the visible spectrum, we demonstrated that our wavelength-multiplexed diffractive system can store and project over 4,000 independent desired images/patterns within its output field-of-view, with high image quality and minimal crosstalk between spectral channels. Furthermore, in a proof-of-concept experiment, we demonstrated a two-layer diffractive design that stored six distinct patterns and projected them onto the same output field of view at six different wavelengths (500, 548, 596, 644, 692, and 740 nm). This diffractive architecture is scalable and can operate at various parts of the electromagnetic spectrum without the need for material dispersion engineering or redesigning its optimized diffractive layers. The demonstrated storage capacity, reconstruction image fidelity, and wavelength-encoded massively parallel read-out of our diffractive platform offer a compact and fast-access solution for large-scale optical information storage, image projection applications.

• The paper introduces a deep learning-optimized diffractive architecture that stores and retrieves over 4,000 unique image patterns via wavelength multiplexing.
• It demonstrates high image fidelity (PSNR >48 dB), minimal crosstalk, and robustness to material dispersion in both simulations and experimental setups.
• The study highlights tradeoffs between diffraction efficiency and reconstruction quality, while also enabling security enhancements through wavelength and spatial keying.

Wavelength-Multiplexed Massively Parallel Diffractive Optical Storage and Image Projection: Technical Overview

Introduction and Motivation

This work presents a wavelength-multiplexed, massively parallel diffractive optical storage and image projection architecture wherein large numbers of unique image patterns are stored and retrieved by exploiting deep-learning-optimized multi-layer dielectric diffractive surfaces. Scaling optical information storage density and access speed is a critical challenge as demands grow in both conventional and ML-driven applications. The shortcomings of magnetic and other traditional storage media—including finite throughput, limited lifetime, and high cost—have increased interest in advanced optical solutions capable of high-density, parallel, and fast-access operations. However, state-of-the-art optical solutions such as holographic and metasurface-based memory remain bottlenecked by crosstalk, fabrication complexity, limited multiplexing bandwidth, and device dispersion.

The central innovation investigated is a compact, multi-layer diffractive system optimized using deep learning for phase profile synthesis, enabling the storage and retrieval of over 4,000 distinct images/patterns, each “addressed” by a unique illumination wavelength. The system operates over the visible spectrum, achieves high fidelity (output PSNR >48 dB), and shows minimal crosstalk. Unlike previous approaches, the design is robust to material dispersion and can be physically realized using diverse optical media.

System Architecture and Optimization

The platform comprises $K$ cascaded dielectric diffractive layers, each finely structured at the wavelength scale, forming a collective optical medium. Each image from a predetermined set is indexed by an illumination wavelength; the optical field generated by that wavelength is modulated sequentially through the multi-layer stack to reconstruct the corresponding image at the output focal plane. The diffractive features—each with trainable thickness and lateral placement—are optimized via gradient-based backpropagation to minimize MSE over all wavelength-channel/image pairs.

Key architectural parameters include:

• Number of layers ($K$)
• Number of diffractive features per layer ($N$)
• Number of stored patterns ($N_w$)
• Output FOV resolution ($N_o$)

The main optimization strategy employed a custom loss, aggregating MSE over all channels and leveraging stochastic gradient descent with dynamic channel-swapping: to maximize fidelity, images with higher reconstruction difficulty (e.g., high texture) are mapped to shorter wavelengths, which yield higher phase resolution per feature.

Numerical and Experimental Results

Storage Density and Image Quality

Simulations with $N_w = 4096$ (visible spectrum: 400–750 nm, $K = 8$ layers, each with $800 \times 800$ features) indicate faithful recovery of stored patterns (mean PSNR = 45.29 dB, improving to 48.01 dB with channel swapping) and uniformity across wavelength channels (SD drops $\sim$2.7-fold).

Scaling analyses further reveal a direct tradeoff between the number of trainable features and the maximum number of retrievable images. Doubling $N$ relative to $K$0 (i.e., $K$1) effectively sustains high fidelity even with increased channel/pattern counts.

Robustness to Material Dispersion

Statistical comparison between implementations using actual N-BK7 dispersion versus idealized dispersion-free materials (constant refractive index) shows no significant difference in PSNR (peak >75 dB), confirming that performance is phase-profile-dominated rather than dependent on dispersion engineering. This implies broad material and spectral portability.

Output Efficiency Versus Fidelity

Optimizations without efficiency constraints generally yield low diffraction efficiencies (<0.01%). By introducing efficiency penalties, the architecture can be flexibly tuned to achieve target output powers, with the expected tradeoff: high efficiency (up to ~10%) slightly degrades image quality (PSNR >20 dB with $K$2).

Architectonic Constraints: Phase Quantization

Limited bit depth for phase modulation (due to fabrication/SLM constraints) directly impacts fidelity: for $K$3 bits, PSNR falls to ~25 dB if trained under the same quantization. Training with target quantization during optimization mitigates this degradation.

Experimental Demonstration

Physical realization is demonstrated on a 2-layer SLM-based setup storing six MNIST patterns at six distinct wavelengths (500–740 nm). In-situ learning overcomes system misalignments and hardware nonidealities, significantly improving both correlation metrics (PCC) and output images relative to naive transfer from simulation.

Systematic Analyses: Tolerance and Security

Investigations into channel spacing and robustness show:

• As channel spacing ($K$4) narrows, inter-channel crosstalk increases, setting a practical limit near the line widths of high-quality lasers ($K$510 pm) for high fidelity.
• Tolerance to wavelength drift (tens of pm) and inter-layer misalignments (tens of nm lateral, ~200 nm axial with “vaccinated” training) can be explicitly engineered via stochastic augmentation during optimization.
• Security implications are notable: correct wavelength and input aperture are necessary “keys,” and an additional diffractive “key-layer” mechanism can be included for further encryption.

Implications and Prospective Developments

This methodology establishes a new direction for scalable, ultra-fast, and energy-efficient analog optical storage and projection. The demonstrated platform leverages deep learning for dense encoding in the spectral domain, is independent of fine-tuned material properties, and is amenable to stacking with orthogonal multiplexing (e.g., by angle, polarization, or spatial shift), further scaling parallelism and capacity. The analog nature enables sub-picosecond retrieval latency.

Prospective advances include:

• Development of hybrid multiplexing schemes for terabit-level storage density
• Integration with high-density, re-writable SLM or printable diffractive layers
• Snapshot parallel readout (using filter arrays or virtual demosaicing) for high-throughput retrieval
• Application-specific storage for security, private communication, or class-conditional projection

Conclusion

This work establishes a scalable, wavelength-multiplexed diffractive optical information storage and image projection architecture, optimized via deep learning, that demonstrates the storage and retrieval of thousands of distinct patterns with high fidelity and minimal crosstalk. The design is robust to material dispersion, tolerates realistic fabrication and alignment errors, and admits tuning between power efficiency and reconstruction quality. The platform is adaptable across broadband spectral regions and advocates for emergent applications in rapid, secure, and high-density optical information storage and all-optical image projection, with implications for both practical systems and the theoretical limits of analog optical computation and memory (2604.02624).