The Ozcan Lab at UCLA used artificial intelligence (AI) to create diffractive waveguides that can control light in ways that are challenging to achieve using conventional waveguides. Unlike traditional dielectric waveguides, which rely on the refractive index profiles of materials, the diffractive waveguides use cascaded phase modulation through spatially optimized, diffractive layers. The structured, cascadable layers are optimized by a deep learning algorithm to collectively sculpt and guide light as it propagates. In the work, the AI element serves to fine-tune the patterns on the surface of each layer of the waveguide to ensure that desired light modes pass through with minimal loss and high purity, while unwanted modes are filtered out. To demonstrate the versatility of their platform, the researchers designed diffractive waveguides that perform specialized functions, including mode filters that selectively transmit or block specific spatial and spectral modes of light, and mode-splitting waveguides that separate and multiplex different light modes into separate output channels. Illustration of diffractive waveguides. Courtesy of Ozcan Lab at UCLA. The researchers also developed bent diffractive waveguides to redirect mode propagation. Additionally, they created diffractive waveguides to control the polarization state of certain spatial modes. Using numerical testing, the researchers showed that the diffractive waveguides can transmit specific spatial modes, with high coupling efficiency and low propagation loss, through multiple cascades of the same diffractive unit. In addition to numerical analysis, they performed experiments in the terahertz (THz) spectrum to show that the waveguides can selectively allow certain spatial modes to pass while blocking others. The experimental results matched the team’s simulations, confirming the feasibility of the diffractive waveguide framework. The diffractive waveguides provide several advantages. Optimizing the training loss function of the diffractive waveguides can fine-tune them for different goals involving spectral, spatial, and polarization features, without the need for material dispersion engineering. This functionality supports the design of waveguides for specific tasks. Additionally, the design framework for the diffractive waveguides can be scaled up or down to accommodate various spectral ranges, including the visible and infrared. Transitioning between spectral bands does not require retraining or redesigning the diffractive waveguide unit. This scalability simplifies the design process and reduces the time and resources needed to implement the waveguides for different applications. According to the researchers, the waveguides can be extended to operate simultaneously at multiple wavelengths and multiple polarization or mode combinations. They can perform a different function at each wavelength channel and each polarization state or mode. The diffractive waveguides can function effectively in the air or immersed within various liquids or gaseous environments. This feature could broaden their potential use as sensors under various environmental conditions. The mode filtering and mode splitting functions are designed for both monochromatic and multiwavelength light. The waveguides can be cascaded to form any desired length. As a result of their versatility and functionality, the diffractive waveguides could allow for the development of highly customized photonic systems for applications such as mode-division or wavelength-division multiplexing, where precise light manipulation, beyond basic transmission, is critical. “Our diffractive waveguide framework reimagines how we can control light. Instead of being constrained by the physical properties of materials, we can teach a sequence of surfaces to guide light and perform complex optical tasks in a cascaded manner,” professor Aydogan Ozcan said. “This gives us a new toolbox, like an optical Lego set, to create highly versatile, task-specific waveguides that can be cascaded for a wide range of applications, from advanced optical communication systems to compact and sensitive sensors.” The research was published in Nature Communications (www.doi.org/10.1038/s41467-025-60626-3).