Fig. 1. Concept of pluggable diffractive neural networks (P-DNN) for multiple tasks. P-DNN is composed of common layers (marked in red) and classification layers (marked in blue and green, respectively). The recognition of handwritten digits and fashion datasets can be achieved by switching plugins of classification layers. When parallel light is encoded as a specific input and passed through a two-layer pluggable D²NN, the light can be focused on a specified region of the detection plane to achieve classification.

Fig. 2. Flowchart of multi-task P-DNN design. The information of the input object is encoded into the amplitude channel, which propagates in free space. The propagated complex field is multiplied by the phase at each layer before being passed to the next layer. The network parameters are optimized according to the mean square error (MSE) of the output field energy. The sigmoid function is used to constrain the phase of each neuron. In the first training, the parameters of the common layer and the classification layer need to be trained simultaneously. In the subsequent training of other tasks, only the parameters of the classification layer need to be optimized. MS: metasurface

Fig. 3. Design of the nanostructure based on geometric phase principle. (a) Schematic of an amorphous silicon nanorod fabricated on a glass substrate, where P_x and P_y are periods in the x and y directions, H is the height, and L and W are the length and width, respectively. (b) Amplitude map of the circular transmission coefficient in cross- and co-polarization for different geometry sizes. (c) Schematic diagram of the deflection angle of the nanofin. (d) Relationship between the rotation angle of the nanofin and the additional phase.

Fig. 4. Experimental setup and the SEM images of the metasurface. (a) Schematic of the experimental setup for observing the object classification. P: linear polarizer, QWP: quarter waveplate, MS: metasurface, MO: microscope objective. (b) The SEM images of the metasurface in the top and side view, respectively. A large pixel consisting of a 10 × 10 array of nanofins is marked by red dotted lines.

Fig. 5. Simulation and experiment result of num-P-DNN. (a) Training accuracy is 92% after training 10 epochs. (b, c) Simulation and experimental results of confusion matrix for handwritten digits. The Simulation and experimental results test accuracy is 91.8% and 91.3% respectively, which is obtained by dividing the sum of elements on the main diagonal of confusion matrix by the sum of all elements. (d) Handwritten digital input images were encoded into amplitude channel. (e, f) Output energy distribution maps of handwritten digits in simulations and experiments. (g) The energy distribution of handwritten digits experimental results and simulation results. ΔE represents the difference between the percentage of maximum and second maximum energy.

Fig. 6. Simulation and experiment result of fashion-P-DNN. (a) Training accuracy is 91% after training 5 epochs. (b, c) Simulation and experimental results of confusion matrix for fashions classification. The simulation and experimental results test accuracy is 90.2% and 90% respectively, which is obtained by dividing the sum of elements on the main diagonal of confusion matrix by the sum of all elements. In the simulation and experimental results, the test accuracy reaches 90.2% and 90%, respectively. (d) Fashion input images were encoded into amplitude channel. (e, f) Output plane energy distribution maps of fashions in simulations and experiments. (g) Energy distribution percentage of experimental and simulated results of fashions. ΔE represents the difference between the percentage of maximum and second maximum energy.