Self-adaptive thermal control devices have become the research focus due to their adaptive characteristics. However, on one hand, the special spectral requirements lead to a complex and time-consuming design process, and on the other hand, the device performance needs to be optimized to meet special application scenarios. To this end, we propose a deep-generation network model to perform complex optimization tasks. Unlike traditional approaches relying on dataset updates, our model integrates a generated neural network with the transfer matrix method (TMM), which generates the expected multi-layer structure and automatically optimizes the material type and the thickness of each layer using the gradient information provided by TMM.
Firstly, a neural network for global optimization is devised to intricately design the structure of photonic devices. The optimization network consists of a residual generation network and an electromagnetic solver TMM. The residual generation network obtains the refractive index and thickness of the material. The TMM solver is employed to derive the spectrum of the generated structure and compute the loss function for reverse parameter updates until the network converges. Secondly, the material categories are constrained, and the material optimization space is limited to a finite number of material properties in the specified material library. We adopt a reparameterization technique to relax the refractive index to a continuous value and restrict it to a specified position on the continuous interval with network updates. A hyperparameter is adopted to regulate the sharpness of the softmax function, thereby limiting the contribution of various materials in the material library to the specified layer. The influence of different loss functions and hyperparameters on network optimization is studied, the loss function is customized, and the best hyperparameters are selected to ensure that the network meets the requirements. Finally, a deep neural network model is utilized to optimize an adaptive thermal control device based on phase change material vanadium dioxide. The structures of 10-layer and 60-layer films are optimized, and their spectral and field distributions of the structure at high and low temperatures are studied to assess the performance.
The proposed global optimization network model eliminates the need for a dataset and can simultaneously optimize the design of material types and thicknesses. We employ 12 materials from the material library to automatically design and optimize multi-layer film devices for adaptive thermal control on a 500 nm Ag substrate. Firstly, a 10-layer adaptive thermal control device is optimized, and the film structure is shown in Table 1. The solar absorption ratio of this device is 0.19, and the difference in high- and low-temperature emissivity is 0.79. For thin films in a high-temperature state, the electric field intensity decreases monotonically along the incident direction. Due to the top-down absorption of the thin film at this time, almost no interference between the incident and reflected waves can be observed. For thin films in a low-temperature state, the entire film system becomes semi-transparent, and strong interference between the incident and reflected waves can be observed. Increasing the number of film layers to 60 can improve device performance, which leads to a solar absorption ratio of 0.17 and 0.82 respectively (Fig. 6). When the number of membrane layers is 10, the traditional neural network's loss value continuously decreases and stops decreasing after 20 optimizations, falling into the local optimal solution and causing the gradient to disappear. Meanwhile, the global optimization network exhibits a spike in the loss value attributable to varying initial points in each optimization run, which makes the structures deviate from local optimal solutions. As the number of membrane layers increases to 60, the global optimization network yields more instances where the results diverge from local minimum values. This characteristic enables the network to effectively explore global optimal solutions and mitigates the risk of the network converging to local optimal solutions (Fig. 7).
We develop a global optimization network framework for designing optoelectronic devices with complex multi-layer film structures. The network solves the material classification problem by adopting probability matrices, and residual modules in the network are also leveraged to make optimization easier. As a validation and demonstration of network optimization capabilities, we adopt this method to design an adaptive thermal control device based on vanadium dioxide. This structure can automatically turn on and off radiative cooling according to environmental temperature without any additional energy input. Meanwhile, it yields excellent performance with a high-temperature solar absorption ratio below 0.2, a high-temperature emissivity greater than 0.9, and an emissivity difference greater than 0.8. Compared with traditional optimization algorithms, neural networks search for the optimal solution with high degrees of freedom and faster speed in searching for optimal solutions, underscoring the practicality of this method in complex design tasks. The results suggest the versatility of this method in designing various optoelectronic systems and highlight the potential extension of this approach to 3D photonic structures using trained neural networks, which offers possibilities for more intricate photonic device design and effective material design in diverse fields.
