A chip-scale nanophotonic integrator for optical signal processing and machine learning

Scaling of artificial intelligence models has motivated the development of physical neural network systems for computation, such as integrated photonic accelerators. A key requirement for these systems is low-power optical nonlinearities. One approach is to use bulk optical nonlinearities, at the cost of requiring high pump powers; an alternative paradigm is to implement optoelectronic nonlinearity through co-integration of optical cavities with on-chip electronics. Here, we leverage the latter approach to realize a chip-scale optical integrator in a silicon nanophotonic circuit. This device computes in the photonic domain the time integral of a stream of optical pulses and applies an inline nonlinearity. The result is output onto a new optical field, enabling fast, direct computation of inner products and nonlinear functions in the optical domain. We fabricate this device and experimentally demonstrate its operation at a pulse rate of 6.25 GHz with an ultra-low energy consumption of 7 fJ/pulse. Finally, we discuss future integration of these devices into optical hardware for signal processing and neural networks.