In order to reduce the contribution of the building sector to global greenhouse gas emissions and climate change, it is important to improve the building performance through retrofits from the perspective of carbon emission reductions. Data-driven methods are now widely used in building retrofit research. To better apply data-driven techniques in low-carbon building retrofits, a better understanding is needed of the connections and interactions in optimization objectives and parameters, as well as optimization methods and tools. This paper provides a bibliometric analysis of selected 45 studies, summarizes current research hotspots in the field, discusses gaps to be filled, and proposes potential directions for future work. The results show that (1) the building-performance optimization (BPO) process established through physical simulation methods combines the site, retrofit variables, and carbon-related objectives, and the generated datasets are either directly processed using multi-objective optimization (MOO) algorithms or trained as a surrogate model and iteratively optimized using MOO methods. When a sufficient amount of data is available, data-driven methods can be used to develop mathematical models and use MOO methods for performance optimization from the perspective of building carbon emission reductions. (2) The benefits of retrofits are maximized by holistically taking environmental, economic, and social factors into account; from the perspectives of carbon emissions, costs, thermal comfort, and more, widely adopted strategies include improving the thermal performance of building envelopes, regulating HVAC systems, and utilizing renewable energy. (3) The optimization process based on data-driven methods, such as optimization algorithms and machine learning, apply mathematical models and methods for automatic iterative calculations and screen out the optimal solutions with computer assistance with high efficiency while ensuring accuracy. (4) Only 2.2% and 6.7% of the literature focus on the impacts of human behavior and climate change on building retrofits, respectively. In the future, it is necessary to give further consideration to user behaviors and long-term climate change in the retrofit process, in addition to improving the accuracy of optimization models and exploring the generalization and migration capabilities of surrogate models.

Latin hypercube sampling (LHS) has become the mainstream sampling method for global sensitivity analysis (GSA) in building performance simulation (BPS) owing to its efficient space-filling characteristics. However, a structural mismatch between the diversity of GSA methods and their specific sampling requirements leaves building researchers without systematic decision criteria. This study focuses on three LHS-compatible GSA methods—the variance-based RBD-FAST (random balance design Fourier amplitude sensitivity test), the rank–regression-based SRRC (standardized rank regression coefficient), and the distribution-based PAWN (moment-independent measure)—and systematically compares their performance and complementary characteristics in the BPS context. Using a typical rural residential building in North China as a case study, we conduct a multi-indicator sensitivity analysis based on 1000 LHS samples. The results demonstrate that the three methods are fully consistent in identifying dominant parameters. For linear output indicators (electricity consumption and carbon emissions), the SRRC and the RBD-FAST exhibit strong rank correlation (Spearman $\rho$ > 0.90), validating the reliability of the LHS-GSA framework. For nonlinear outputs (cost and thermal comfort), the inter-method correlations diverge significantly, with moderate consistency between the SRRC and PAWN on the cost indicator ($\rho$ = 0.62) and weak correlation between the RBD-FAST and the other two methods ($\rho$ = 0.00–0.29), revealing differential adaptability of GSA methods to model complexity. This study provides a method-selection framework for LHS-based sensitivity analysis in BPS and recommends a multi-method cross-validation strategy to ensure robustness of analytical results.