Fig. 1. Bilayer DNN chip integrated on a quartz substrate. (a) A schematic diagram of the chip capable of operating in both air and water. Optical images enter the DNN, and the final recognition results are reflected on the output layer through the distribution of light intensity. (b) The schematic diagram illustrating the propagation of light in the chip. (c) A network description of the physical computation process of the DNN chip. Dataset images are generated at the input layer and then propagate through two diffractive layers with optical operation based on coherent superposition. In our experiment, the diffractive layer is designed with binary phase modulation. (d) The digital image of the DNN chip. Scale bar, 5 mm.

Fig. 2. The simulation results of the DNN chip. (a) The training process diagram of the DNN chip. (b)–(e) Simulated outputs of the DNNs and corresponding normalized light intensities in the 4 circled regions. (b) Task: handwritten digital recognition. Medium: air. (c) Task: fashion product recognition. Medium: air. (d) Task: handwritten digital recognition. Medium: water. (e) Task: fashion product recognition. Medium: water.

Fig. 3. Characterization and fabrication error analysis of the chip for fashion product recognition. (a) The obtained phase map of the DNN after training. (b) The optical images of the two surfaces of the DNN chip captured by a 4f optical system. Scale bars, 1 mm. (c) The simulated impact of phase modulation errors caused by etching on the accuracy of the DNN. (d) and (e) The simulated impact of alignment errors caused by double-sided photolithography on the accuracy of the DNN working in (d) air and (e) water, respectively.

Fig. 4. Experimental results of the DNN chip. (a) Experimental optical setup. HWP: half-wave plate, PBS: polarizing beam splitter, BS: beam splitter, CCD: charge-coupled device, DMD: digital micromirror device. (b) The digital image of the DNN chip working in water. (c) Confusion matrices for handwritten digital recognition in air and water. (d) Confusion matrices for fashion product recognition in air and water. (e)–(h) Recorded outputs of the DNNs and corresponding normalized light intensities in the 4 circled regions. Scale bars, 2 mm. (e) Task: handwritten digital recognition. Medium: air. (f) Task: fashion product recognition. Medium: air. (g) Task: handwritten digital recognition. Medium: water. (h) Task: fashion product recognition. Medium: water.