Developing data-driven surrogate models for holistic performance-based assessment of mid-rise RC frame buildings at early design

• Published in: Engineering structures 2021-10, Vol.245, p.112971, Article 112971
• Main Authors: Zaker Esteghamati, Mohsen, Flint, Madeleine M.
• Format: Article
• Language: English
• Subjects: Building design; Buildings; Concrete; Concrete construction; Early design; Environmental assessment; Environmental impact; Frame structures; Interpretable Machine learning; Learning algorithms; Machine learning; Office buildings; Parameter identification; Performance-based Engineering; Regression analysis; Seismic hazard; Seismic response; Seismic surveys; Structural engineers; Support vector machines; Surrogate modelling; Topology; Variance-based sensitivity
• ISSN: 0141-0296; 1873-7323
• DOI: 10.1016/j.engstruct.2021.112971

•Apply directional sampling to auto-generate an inventory of 720 RCMRF office building of varying topology, and use performance-based earthquake engineering methods to assess economic and environmental impacts associated with seismic hazard in Charleston, SC.•Leverage inventory to build a machine learning (ML) pipeline and compare five common ML models as surrogate for life cycle resilience and sustainability impacts using only input parameters known/estimable during early design.•Analyze the sensitivity of models to identify the most influential parameters and assess tradeoffs associated with using only crude topological parameters.•Support vector regression found to achieve an acceptable balance of accuracy, interpretability and generalizability, even when using only topological parameters and can serve as an aid to structural engineers in early design. This paper presents a framework to develop generalizable surrogate models to predict seismic vulnerability and environmental impacts of a class of buildings at a particular location. To this end, surrogate models are trained on a performance inventory, here simulation-based seismic and environmental assessments of 720 mid-rise concrete office buildings of variable topology in Charleston, South Carolina. Five surrogate models of multiple regression, random forest, extreme gradient boosting, support vector machine and k-nearest neighbors were trained in a machine-learning pipeline including hyperparameter tuning and cross-validation. Variance-based sensitivity and accumulated local effect analysis were performed on the most accurate model to identify the most influential parameters and interpret the trained surrogate model. Support vector machines achieved the highest accuracy for total annual loss with an average 10-fold adjusted R2 of 0.96, whereas simpler linear regression was adequate to estimate the initial and seismic-induced embodied carbon emission. Floor area, building height, lateral-resisting frame weight, and average beam section sizes were found to be the most influential features. As these features may be approximated by an experienced structural engineer the results indicate that, with suitable performance inventories available, it should be possible to employ surrogate models in early design to narrow the initial design space to highly resilient and sustainable configurations.