Photonic power splitter design based on deep learning and gradient descent method

Increasing demand for high density and broadband photonic integrated circuit (PIC) has motivated designs for photonic devices with high performance and compact footprint. However, the number of parameters in traditional photonic device designs is limited by the working principles for such devices, which often results in a perceived trade-off among device performances, such as bandwidth, efficiency, and footprint. Nonlinear optimizations such as direct binary search (DBS) and genetic algorithms (GA) have been explored in photonic designs, yet they have drawbacks such as slow convergence time in the range of 96-140 hours. By contrast, designs based on deep learning model relates device performances (output) and device parameters (inputs) via data-driven methodology, which enables arbitrarily large number of design parameter space that may overcome the perceived trade-off in traditional photonic designs. This work proposes a design methodology based on a combination of deep learning model and gradient descent method for photonic power splitter with arbitrary splitting ratio. Using pixel-based device geometry, a deep learning model relating the geometric parameters and device spectral performance is first established. Afterwards, a figure of merit based on a target splitting ratio is optimized through gradient descent method to yield the corresponding pixel-based device geometry. We demonstrate this method in the design of photonic power splitter with splitting ratio from 0.25 to 10, with insertion loss between 0.5dB to 1.16dB, and device size smaller than 16 μm2. In comparison with the genetic algorithm and direct binary search method, our proposed method is much faster in terms of convergence time.