Fig. 1. Schematic diagrams of working principles. (a) Operating principle of device at high temperature; (b) operating principle of device at low temperature; (c) target spectrum of device at high and low temperatures

Fig. 2. Schematic of multilayer film system for self-adaptive thermal control, with each layer of material selected from the material library and related thickness optimization performed

Fig. 3. Schematic of global optimization network

Fig. 4. Problems that occur when networks use different loss functions. (a) Problem of network when Loss=(1-Δε)+θα is used for optimization; (b) problem of network when Loss=1/Δε+θα is used for optimization

Fig. 5. Influence of parameter β on network performance. (a) Ideal value of probability matrix P; (b) value of probability matrix P when material does not converge; (c) influence of parameter β on number of convergent iterations and network loss; (d) working process of network when β=0.95

Fig. 6. Results of optimized spectral emissivity and field distributions. (a) Spectral emissivity when number of layers is 10; (b) field distribution in high temperature state corresponding to Fig. 6(a); (c) spectral emissivity when number of layers is 60; (d) field distribution in low temperature state corresponding to Fig. 6(a)

Fig. 7. Comparison of different network optimization modes. (a) Histogram of loss of 100 devices generated from neural network when number of layers is 10; (b) variation of loss when number of layers is 10 for different optimization methods; (c) histogram of loss of 100 devices generated from neural network when number of layers is 60; (d) variation of loss when number of layers is 60 for different optimization methods