Abstract: Orthogonal experiment was combined with BP neural networks to optimize welding process parameters of 0.2 mm GH4169 diaphragms by using microbeam TIG welding echnique. The BP neural networks model was trained by the results of orthogonal experiment, and the BP neural networks model described the relationship between welding joints’ properties (diameter, height, and tensile resistance) and welding process parameters, including peak current, background current, welding velocity, and pulse frequency. The results shows that according to the orthogonal experiment and the prediction of BP neural networks model, then small-step method is used for further searching, the best parameters are found when peak current is 11.6 A±0.2 A, background current is 4.3 A±0.1 A, welding velocity is 4.14 mm/s±0.1mm/s, and pulse frequency is 52 Hz±2 Hz. The xperiment validated that the test values of four samples are within the prediction range, and the tensile force is higher than that of previous experiment. It proves that the model has a high accuracy of prediction and the method can increase the efficiency of process design availably.