Scaling photonic systems-on-chip production with neural networks

Abstract

We describe our use of deep learning to optimize the multi-dimensional parameter space of systems-on-chip as an important step towards the scalable production of photonic solutions and their widespread integration into high-volume applications. The challenges of transitioning between prototype and volume production are highlighted, and the suitability of deep neural networks for navigating the multi-dimensional design space of today’s photonic circuits is discussed. We adopt multi-path neural network architectures to reduce the computational requirements of model training and to mitigate the risk of overfitting. We demonstrate the use of a multi-path neural network to optimize the construction parameters of photonic designs in a high-volume production environment. Lastly, we discuss the advantages of using machine learning not only as a highly capable tool for navigating the multi-dimensional design space of complex systems-on-chip but also as an effective strategy for compensating for fabrication process non-uniformities that are undetectable by standard process metrology instruments.