Reconfigurable beam shaping with rotating phase masks using diffractive neural networks

Diffractive neural networks (DNNs) as physical representations of artificial neural networks have demonstrated remarkably flexible potential as optical systems for laser beam shaping but dynamic reconfiguration typically relies on active spatial light modulators (SLMs). We present a method to train DNNs so that rotating one or more passive phase masks switches the device among multiple, completely distinct beam-shaping functionalities, enabling dynamic beam control without SLMs. The method can also be realized experimentally with reflective DOEs produced with a unique direct laser writing approach.