Facial expression is one of the most powerful, natural, and universal signals for human beings to express emotional states and intentions. There are many applications for facial expression recognition (FER) in human society such as healthcare. Thus, the importance of correct and innovative FER approaches in Artificial Intelligence is evident. However, commonly used methods suffer from a lack of classification generalization in FER. To tackle this problem, we propose a generic facial expression recognition network based on evolutionary neural architecture search, called ENAS-FERNet, which can automatically evolve neural network architectures using both laboratory-controlled and in-the-wild FER datasets. The experiments of ENAS-FERNet were carefully designed and compared with state-of-the-art (SOTA) methods in the case of training from scratch. In addition, we validated the interference resistance of ENAS-FERNet on the synthetic noisy FER dataset and analyzed the time consumption of ENAS-FERNet. Comprehensive experimental analysis and results show that the proposed ENAS-FERNet method achieves the most well-known results on the CK+, Affect-Net, and RAF-DB (10%) datasets, as well as competitive results on the JAFFE, RAF-DB, and RAF-DB (20%) datasets. The results of these experiments show that our ENAS-FERNet has good classification generalization capabilities on these challenging datasets.