At present, buildings in arid and hot regions are facing severe challenges of indoor comfort improvement and carbon emission reduction, especially in rural areas. Multi-objective optimization could be an effective tool for tackling the aforementioned challenges. Therefore, this paper proposes a life-cycle optimization framework considering thermal comfort, which is beneficial to promoting residents’ motivation for low-carbon retrofit in arid climate regions. First, in response to the above problems, three objective functions are specified in the framework, which are global warming potential (GWP), life cycle cost (LCC), and thermal discomfort hours (TDH). To improve the optimization efficiency, this research uses Deep Neural Networks (DNN) combined with NSGA-II to construct a high-precision prediction model (meta-model for optimization) based on the energy consumption simulation database formed by the orthogonal multi-dimensional design parameters. The accuracy index of the modified model is R 2 > 0.99, cv(RMSE) ≤ 1%, and NMBE ≤ 0.2%, which gets rid of the dilemma of low prediction accuracy of traditional machine learning models. In the scheme comparison and selection stage, the TOPSIS based on two empowerment methods is applied to meet different design tendencies, where the entropy-based method can avoid the interference of subjective preference and significantly improve the objectivity and scientific nature of decision analysis. Additionally, sensitivity analysis is conducted on the variables, which supports guidance for practitioners to carry out the low-carbon design. Finally, the multi-objective optimization analysis for a farmhouse in Turpan is taken as a case study to evaluate the performance of the framework. The results show that the framework could significantly improve the building performance, with 60.8%, 52.5%, and 14.2% reduction in GWP, LCC, and TDH, respectively.