Experimental demonstration of ultra-wideband tapers, splitters and crossings with sub-0.1dB loss through computationally efficient and data-driven eigenmode expansion

Abstract

The growing demand for custom photonic components with stringent performance metrics necessitates efficient design approaches [1-2]. We experimentally demonstrate three ultra-broadband, silicon-based photonic devices-waveguide tapers, splitters, and crossings-designed using a data-driven eigenmode expansion (EME) method [3]. For each device category, we first model light propagation in a given geometry using cascaded eigenmode scattering matrices. With this representation, our electromagnetic computations benefit from parallel data processing in GPUs, reducing individual simulation times to tens of milliseconds, while maintaining 3D-FDTD level of accuracy. We then utilize nonlinear optimization algorithms [4], and iteratively optimize the device geometry for near-lossless operation. Our first device is a 30µm-long waveguide taper with input and output widths of 0.5µm and 9.0µm, respectively (Fig 1a). The insertion loss through this device is measured to be 0.050dB at 1550 nm, by using the cutback method through up to 250 copies of this taper. The 2µm-long, 1×2 power splitter (Fig 1b) similarly demonstrates low-loss operation with an insertion loss of 0.083dB, measured using a similar cutback approach. Finally, the waveguide crossing (Fig 1c), designed within a 12×12 µm2 footprint, also achieves a low insertion loss of 0.051dB. All devices maintain insertion losses below 0.1dB across the 1500-1580 nm range, with over 250nm of 1dB-bandwidths as confirmed by 3D-FDTD simulations. Our data-driven eigenmode simulator also allows for the input/output waveguide widths and the device length to be user-specified as design hyperparameters, offering flexibility for a variety of on-chip applications. Moreover, the total computational design time for the taper, splitter, and crossing are 18s, 14s, and 5s, respectively, offering simulation speeds more than 100,000 times faster than conventional EME methods, and highlighting the extreme efficiency of the data-driven eigenmode expansion method. These demonstrations can unlock new opportunities for scalable, application-based photonic systems in areas like communication, sensing, and computing.