Abstract: To address the limitations of traditional materials in achieving both lightweight design and high-efficiency low-frequency sound insulation, this paper proposes an optimization design method for membrane-type metamaterials that integrates a forward prediction neural network with a multi-objective evolutionary algorithm. By establishing a high-precision surrogate model of the “structural parameters–sound insulation performance” relationship and coupling it with a generalized differential evolution algorithm, the method enables automatic identification of Pareto-optimal designs under multiple acoustic objectives. Furthermore, using this surrogate model, we design a membrane-type metamaterial exhibiting broadband low-frequency sound insulation performance, achieving an average sound transmission loss of 16.2 dB over the 100–500 Hz frequency range while maintaining a low surface density of only 0.23 g/m². This work advances, from a methodological perspective, the design paradigm of sound-insulating metamaterials—from empirical trial-and-error toward a synergistic optimization framework integrating neural networks and evolutionary algorithms—thereby offering a systematic solution strategy for the optimal design of complex vibro-acoustic systems.