In the past three decades, artificial photonics devices have made remarkable achievements in the fields of super-resolution, biosensing and optical communication. The designs of traditional photonics devices are usually realized by analyzing physical models and establishing numerical simulation methods. However, the structural design based on the numerical simulation method largely depends on the empirical model, and a large number of parameter combinations need to be calculated in the process of structural optimization, so it can only get suboptimal results in a limited parameter space. With the continuous improvement of computer computing capability and the enrichment of computer algorithms, the inverse design of photonics devices can effectively solve the above obstacles. The inverse design method can not only find the optimal structure distribution in a broader parameter space, but also design irregular structures that cannot be directly designed by human brain, which makes the performance of inverse designed photonics devices closer to the limit. This review first introduces the three common methods of photonic device inverse design and then introduces several important applications based on inverse design methods in detail. Common inverse design methods can be divided into gradient descent algorithm and genetic algorithm. Gradient descent algorithm uses gradient information to guide the optimization of structure. Topology optimization is a commonly used algorithm in gradient descent algorithm, which can optimize the material distribution in a given design area according to the given objective function and constraint function. The gradient of objective function in topology optimization is usually calculated by adjoint method. Genetic algorithm looks for the global optimal value by simulating the evolution process of “survival of the fittest”. The algorithm has four main steps: initial population guess, crossover, mutation and selection. By iterating the above four steps for a certain number of times, the whole population can evolve in the desired direction. The deep learning model can effectively extract features from a large amount of data and has been proved to characterize the physical relationship between photonic structures and their optical properties. The trained deep learning model can replace the electromagnetic simulation process and greatly reduce the design time of photonics devices. Based on the above characteristics, these algorithms have been successfully used in the design of photonic crystal, metagrating, metasurface and other optical devices. However, inverse design methods are not omnipotent. They have different limitations. For example, gradient descent method usually converges to the local optimal value, so it needs a better initial starting point. Genetic algorithm has a large amount of calculation and is sensitive to parameter setting. Deep learning needs data sets, and the performance of the network is affected by super parameter setting. Therefore, when selecting the inverse design method, it is necessary to evaluate the calculation time and optimization dimension of the physical model. Intelligent algorithm combines nano photonics with computational science, and provides researchers with an efficient new design method. In the future, the difficulty of optical field manipulation can be simplified by intelligent inverse design methods, so as to achieve the purpose of designing high-performance optical devices from the demand.