New 'Dimension Expansion' method cuts nanophotonic simulation costs by 50%

Designing nanophotonic structures like metalenses or beam splitters is notoriously difficult: the design space is huge, the relationship between structure and performance is nonlinear, and each electromagnetic simulation is computationally expensive. Existing deep-learning approaches rely on precomputed datasets or libraries of optimized structures, limiting scalability for continuous inverse-design tasks. To solve this, a team led by Shuo Huang and Chia Wei Hsu developed the Dimension Expansion Network (DEN)—a fully unsupervised framework that removes the need for any pre-generated data.

DEN addresses the fundamental mismatch between low-dimensional design objectives (e.g., a desired focal length) and high-dimensional photonic structures. It projects those compact targets into structured, high-dimensional conditioning representations before generating the device geometry. The entire network is trained end-to-end using differentiable electromagnetic simulations. When tested on free-form metalens design, DEN achieved focal intensities matching adjoint-based optimization while using approximately 50% fewer simulations. It also generalized across tens to thousands of focal targets within a shared region. For asymmetric Y-splitter design, DEN accurately produced arbitrary power-splitting ratios using just 21 training targets and demonstrated robust broadband performance—a dramatic improvement in data efficiency.