[Paper] Phase-multiplexed optical computing: Reconfiguring a multi-task diffractive optical processor using illumination phase diversity

Source: arXiv - 2512.06658v1

Overview

A new optical‑computing architecture lets a single diffractive processor perform hundreds of different linear transformations simply by changing the phase pattern of the illumination light. By encoding each desired operation in a distinct 2‑D “phase key,” the same hardware can be re‑programmed on the fly, opening the door to versatile, low‑power optical accelerators for signal processing and machine‑vision tasks.

Key Contributions

• Phase‑multiplexed reconfigurability: Introduces a method to switch among (T) different complex‑valued linear mappings using only the illumination phase, without swapping physical components.
• Compact monochrome design: Achieves the same functionality as wavelength‑multiplexed systems while staying in a single wavelength band, dramatically reducing chromatic aberrations and fabrication complexity.
• Scalable network sizing: Shows that a diffractive layer with (N = 2 \times T \times N_i \times N_o) optimized features can faithfully implement any of the (T) transformations for arbitrary input fields.
• Large‑scale proof‑of‑concept: Demonstrates numerically that (T = 512) distinct complex transformations can be realized with negligible error.
• Error reduction vs. wavelength multiplexing: Quantifies how phase‑multiplexing yields lower transformation errors, enabling larger‑scale optical processing on a single chip.

Methodology

1. Network architecture: A single planar diffractive optical element (DOE) sits between an input aperture (pixelated field (N_i)) and an output plane (pixelated field (N_o)).
2. Phase keys: For each target transformation, a pre‑computed 2‑D phase mask (the “phase key”) is placed on a spatial light modulator (SLM) or similar device that illuminates the input aperture.
3. Joint optimization: Using gradient‑based wave‑propagation simulations, the DOE’s surface relief is co‑optimized together with all (T) phase keys so that, when a particular key is displayed, the DOE demultiplexes the input and reproduces the corresponding linear mapping at the output.
4. Complex‑valued linear mapping: The system is modeled as a matrix (H_t \in \mathbb{C}^{N_o \times N_i}) for each task (t). The training loss penalizes the difference between the actual output field and the desired (H_t \cdot)input for every task and a diverse set of random inputs.
5. Numerical validation: Simulations with realistic diffraction physics (Fresnel propagation, quantized phase levels, and noise) verify that the same DOE can switch among 512 tasks with mean‑squared error well below (10^{-4}).

Results & Findings

• High fidelity across tasks: The average normalized mean‑square error (NMSE) across the 512 transformations was < 0.001, indicating near‑perfect reconstruction of arbitrary complex inputs.
• Phase‑key orthogonality: Distinct phase keys produce well‑separated output fields, confirming effective multiplexing without cross‑talk.
• Comparison to wavelength multiplexing: For the same number of tasks, phase‑multiplexed designs showed a 3–5× reduction in NMSE, mainly because they avoid chromatic dispersion and the need for broadband diffractive designs.
• Scalability: The required number of diffractive features grows linearly with both the number of tasks and the input/output resolution, suggesting predictable hardware scaling.

Practical Implications

• Reconfigurable optical accelerators: A single DOE can act as a programmable linear operator for tasks such as Fourier transforms, convolution kernels, or matrix‑vector multiplications—key building blocks in optical neural networks and signal‑processing pipelines.
• Low‑power, high‑throughput inference: Since the computation is performed at the speed of light with negligible electronic overhead, latency‑critical applications (e.g., real‑time video analytics, lidar preprocessing) can benefit.
• Simplified hardware stacks: By staying monochrome, the system avoids expensive multi‑laser sources, dichroic optics, and complex alignment procedures required for wavelength‑division schemes.
• Rapid prototyping: Updating the phase key on an SLM is essentially a software change, enabling on‑the‑fly deployment of new algorithms without fabricating a new diffractive element.
• Integration pathways: The approach can be combined with existing silicon‑photonic or metasurface platforms, allowing compact on‑chip implementations for edge devices.

Limitations & Future Work

• Physical implementation challenges: The study is currently numerical; real‑world fabrication tolerances, SLM refresh rates, and phase quantization could introduce additional errors.
• Scalability trade‑offs: While the feature count scales linearly, very large (T) or high‑resolution inputs may demand impractically fine diffractive patterns or large apertures.
• Dynamic range and noise: The system assumes ideal coherent illumination; incoherent or partially coherent sources may degrade performance.
• Future directions: Experimental validation, exploration of hybrid amplitude‑phase keys, adaptive learning for on‑device calibration, and integration with electronic control loops to close the loop for closed‑form optical computing pipelines.

Authors

• Xiao Wang
• Aydogan Ozcan

Paper Information

• arXiv ID: 2512.06658v1
• Categories: physics.optics, cs.NE, physics.app-ph
• Published: December 7, 2025
• PDF: Download PDF