Explainable AI for inverse-design of optical network on chip routers

Abstract Inverse design uses advanced AI capabilities to autonomously create the best geometries and configurations for an optical network-on-chip router. Here, we use reinforcement learning to facilitate the design process, where the system learns over time what architectures best suit the target criteria (latency, signal loss, area, throughput) to achieve the best functioning router. However, while reinforcement learning is an effective means of achieving desired output, the reinforcement learning created is often complex and non-interpretable for human engineers. Therefore, we use explainable artificial intelligence methods to make the final interpretations more interpretable and justifiable. Explainable artificial intelligence helps to explain why each move was made during the design process. Thus, reinforcement learning-inverse-designed optical network-on-chip routers will perform better under desired metrics with explainable artificial intelligence providing human designers a level of explainability and justification for human trust and verification for artificial intelligence generated things.