In order to further reduce the subjectivity of network design and improve the ability of model feature extraction, an ultrasonic detection method for weld defects based on neural network architecture search is proposed. Through the designed multi-level and multi-branch search space and an untrained architecture search and evaluation method, an efficient defect classification network was automatically constructed to complete the task of weld defect classification. Experiments were carried out on a self-constructed data set, and compared with the manually designed model, the classification accuracy of defect types reached 95.26% when the number of parameters was only 7.3 M. Compared with the model constructed using neural network architecture search, the proposed method can reduce the searching time to 8.29% of the baseline model while weighing multiple conflicting objectives, which proved the efficiency and effectiveness of the proposed method.