Deep learning co-design helps scientists project 28-layer 3D images without crosstalk

Engineering researchers at the University of California, Los Angeles (UCLA) have developed an advanced three-dimensional (3D) image projection system capable of displaying 28 distinct depth layers simultaneously in a single optical exposure.

Led by Professor Aydogan Ozcan from the UCLA Samueli School of Engineering and the California NanoSystems Institute (CNSI), the project introduces a compact framework designed to enhance next-generation holographic displays, medical imaging, and virtual reality interfaces.

The research details a hybrid digital-optical architecture that resolves persistent issues of visual distortion and cross-talk in dense 3D imaging.

Traditional multi-layer volumetric projection methods often suffer from image degradation when multiple focal planes are closely spaced. As these layers converge, the light fields often bleed into one another, reducing depth clarity, causing severe blurriness, and diminishing visual sharpness for the viewer.

Deep learning and light programming

To overcome these limitations, the UCLA research team utilized deep learning to co-optimize a digital computational encoder alongside a passive, physical optical decoder.

The system functions by running target visual data through a digital neural network that incorporates explicit depth and coordinate instructions. This network compresses the multi-layered structural information into a singular, unified phase pattern that represents the entire 3D volume.

When light passes through the system, it hits a series of structurally optimized diffractive surfaces that act as an analog decoder. These surfaces physically manipulate the light wave, routing specific components of the image to their exact pre-assigned depth planes.

This precise “light programming” allows the system to suppress data leakage between adjacent layers, maintaining clean visual separation even when planes are separated by distances near the scale of a single light wavelength.

Scalability and experimental validation

Through numerical modeling, the team demonstrated that the framework scales effectively to accommodate complex volumetric scenes split into 28 independent axial slices.

Additionally, the system features a dynamic adjustment capability, allowing operators to alter the target depth positions of the projected images on demand without modifying the core physical architecture.

To confirm the practical viability of this digital-optical pipeline, the researchers built a physical, two-plane hardware prototype that uses a single-layer optical decoder operating in the visible light spectrum.

Experimental measurements confirmed that the projected light distribution closely matched both the target designs and computational simulations. The experimental setup significantly outperformed unassisted free-space optical systems, validating the real-world stability and accuracy of the design.

Future tech integration

The compact layout of this encoder-decoder architecture provides an energy-efficient foundation for high-resolution volumetric imaging. Beyond immediate integration into near-eye augmented reality (AR) and virtual reality (VR) headsets, the technology holds potential applications in multi-depth microscopy, real-time 3D medical visualization, and optical computing.

Moving forward, the research team aims to expand the capabilities of this framework by exploring multispectral operations to support full-color projections, multi-perspective holography, and the integration of physically fabricated multi-layer decoders suitable for commercial manufacturing form factors.

The research was first published in the journal Light: Science & Applications.

Tags: