In this paper, we propose an efficient optimization approach for multi-objective optimization problems (MOPs). The proposed utilizes an autoencoder-based surrogate model and gradient descent to search for Pareto optimal solutions. Evolutionary multi-objective optimization algorithms excel in finding diverse non-dominated solutions but often suffer from excessive execution times, particularly for real-world applications. Surrogate models offer a promising alternative by reducing the computational cost of objective function evaluations, and as such, have become the focus of increasing research interest. However, challenges persist in achieving accurate approximations and reducing surrogate model training times, limiting their practicality. To address this, we propose a method that uses an autoencoder with residual connections as surrogate models, introducing active learning to reduce sampling costs. The problem is then decomposed into multiple single-objective problems using weighted sum decomposition, and the solutions are optimized based on the approximate gradients established by the surrogate model. Evaluation experiments using the ZDT series and a human-powered aircraft design problem demonstrate that the proposed method effectively reduces sampling and training time, while achieving better approximation accuracy compared to existing surrogate methods. Furthermore, it achieves Pareto optimal fronts more quickly, while maintaining solution accuracy comparable to other representative evolutionary multi-objective optimization algorithms.