Far-field optical classification of subwavelength objects

Object detection requires localizing and classifying the size and shape of an unknown object. Here we show that artificial-intelligence-enabled analysis of light scattered on objects that are not resolvable by conventional microscopy can be used for their shape classification. In a proof-of-principle experiment, we demonstrate classification with ∼90% accuracy for objects of unknown subwavelength dimensions in the range from λ/6 to λ/2 (where λ is the illumination wavelength) belonging to one of five shape classes. The method can be scaled to applications across the entire electromagnetic spectrum and used in a variety of tasks, such as the detection and study of biological particles, environmental sensing, and device diagnostics.