Software- and hardware-accelerated inverse design of photonic devices

Co-Packaged Optics offer superior energy efficiency and data density compared to pluggable and onboard optics. As optical engines become more tightly integrated with switch silicon, new challenges emerge in thermal management, yield and reliability – highlighting the need for innovative photonic devices. Inverse design using the adjoint method enables efficient exploration of high-dimensional parameter spaces and is particularly well suited for multi-objective optimization of photonic structures with complex geometries. We accelerate the inverse design process through both software and hardware improvements. On the hardware side, we leverage the high memory bandwidth of GPUs by porting an open-source inverse design toolkit to GPUs. Single-GPU FDTD simulations achieve up to ~100× speedup over CPU implementations, consistent with memory bandwidth advantages. Multi-GPU FDTD simulations demonstrate over 95% parallel efficiency on an NVIDIA DGX-2 server equipped with 16 NVLink-connected Tesla V100 GPUs. As device complexity increases and hundreds of design parameters are needed to define intricate geometries, gradient computation emerges as a major bottleneck – often surpassing FDTD simulation time. To overcome this challenge, we develop an Autograd-based gradient computation algorithm tailored for complex geometries, achieving approximately a 60× speedup in gradient computation. We demonstrate the enhanced simulation and design capability through three representative examples: (1) a polarization-splitting grating coupler with record-high coupling efficiency; (2) a single-polarization grating coupler with wide bandwidth covering the entire CWDM band; and (3) a rib-waveguide ring coupler with 2× reduced sensitivity to etch-depth variation.