|
Recent developments in computational inverse design offer the promise of significantly reducing the footprint and allowing complex optical functionality in silicon photonic components as compared to existing conventional building blocks. However, reliable fabrication of such components is one of the major bottlenecks in its widespread adoption. A common characteristic of such designs is the presence of small features that have meaningful impact on the optical performance. Current approaches to tackle this problem consider designing for robustness, such as by co-optimizing for over- and under-etched geometries at the design stage. This is often followed with a design-for-manufacturing optimization step to meet specifications of a foundry such as minimum feature size and curvature radii. Those approaches often incur additional significant computational costs as well as a reduction in peak optical performance. In this work, we highlight our recent progress to bridge the gap between inverse design methods and their ability to deliver reliable and manufacturable designs. We observe that the so-called parameterized shape optimization methods are more likely to produce robust designs for certain classes of components, as showcased in integrated mode converter designs. For components that benefit from topological inverse design such as wavelength demultiplexers, we propose a new optimization penalty that naturally leads the optimizer towards more robust designs. In a new research direction, we also consider improving fabrication reliability by the development and use of data-driven predictive models for fabrication. Leveraging deep learning tools, we present prediction and correction models that improve fabrication outcomes for a variety of components made at an e-beam prototyping foundry.
|
