In recent years, perovskite solar cells (PSCs) have attracted much attention because of their remarkable advantages in power conversion efficiency and manufacturing cost. However, their complex physical mechanisms and numerous constraints pose challenges to experimental design, process fabrication, and comprehensive optimization strategies. Here, we carried out a series of multi-physical field simulations with the optoelectronic multi-physical field coupling model as the core, and studied the underlying physics and boundary conditions of the optoelectronic coupling model, and then obtained a large amount of data on the optical and electrical properties of PSCs. Based on these data, we established the machine learning models and neural network models for the micro physical quantities and macro photoelectric responses, which predicted the performance of PSCs with an error of less than 3% in a fast speed. Combined with the genetic algorithm, the model reversely optimized the structural parameters according to the given response curves to obtain the more efficient PSCs. This study effectively solves the problem that PSCs are difficult to optimize design due to complex photoelectric coupling mechanism, numerous physical property parameters and slow simulation speed, and provides a feasible path for rapid and intelligent design of photovoltaic devices.