Research in the ocean places high demands on chips’ robustness, speed, and energy consumption. Diffractive neural networks (DNNs) enable direct optical image processing at light speed, with great potential for underwater applications. Here, we experimentally demonstrate a compact DNN chip capable of operating directly in both water and air by multi-objective training and initial training value optimization. The two layers of DNNs are integrated on the two surfaces of a quartz plate, respectively. The chip achieved high accuracies above 90% in recognition tasks for handwritten digits and fashion products. The architecture and material ensure the chip’s high stability for long-term underwater use.