Self-updating AutoML model for building properties prediction: A case study on building mass

Abstract

Building properties such as mass and natural period are intrinsic characteristics of an existing structure and are critically important for assessing its safety and stability. To accurately predict these properties, many researchers have developed data-driven machine learning (ML) models for estimating natural period and damping ratio. However, studies focusing on building mass/weight remain limited. Moreover, existing ML models usually result from one-time training and require specialized computer skills for development, and thus lack the ability to update automatically when new data are available. To address these gaps, this study proposes a generalizable model for predicting building properties using AutoGluon, an automated machine learning (AutoML) framework, with a focus on building mass prediction. Specifically, based on an extensive literature review, a building property database comprising 909 real-world samples is constructed. A multi-layer stack ensembling strategy combined with repeated five-fold cross-validation is employed for model training. Predictive performance of the optimal model on the test set is then compared to that of conventional empirical models. Results show that the proposed model significantly outperforms empirical approaches, offering more comprehensive and accurate predictions of building mass. In addition, Shapley additive explanations technique is employed to interpret the model, providing both global and local insights into the importance of various building features and their contributions to the prediction output. Finally, the trained model is deployed on an online platform, enabling both prediction and automatic updates, thereby improving accessibility and ensuring continuous enhancement through incoming data.