ML Algorithms for Forecasting Seismic Demand in RC Buildings

By Muhammad OsamaReviewed by Susha Cheriyedath, M.Sc.Jul 21 2024

In a recent article published in the Journal of Building Engineering, researchers introduced a new method to guide the development of machine learning (ML) models for estimating seismic demand in existing reinforced concrete (RC) buildings. They aimed to obtain interpretable and accurate predictions of the maximum inter-story drift (MID), a key indicator of structural damage and collapse potential, in different RC frames under pulse-like earthquakes.

Background

RC buildings are common structures worldwide, especially in developing countries. However, many are vulnerable to earthquakes due to inadequate design codes, poor construction quality, and aging. Therefore, evaluating their seismic performance and identifying critical ones for retrofitting or replacement is essential.

MID is a key engineering demand parameter (EDP) that reflects the seismic performance of a building. It indicates deformation and damage in structural and non-structural components. However, estimating them for many buildings requires complex and time-consuming numerical simulations.

ML is a branch of artificial intelligence that can learn from data and make pattern predictions. It has been applied to various civil engineering problems, including seismic demand estimation. However, most existing ML models for seismic demand estimation are black-box models, meaning that they are not transparent or interpretable, and do not provide any insight into the underlying physical mechanisms or the influence of the input variables.

About the Research

In this paper, the authors aimed to overcome the limitations of black-box ML models by proposing a novel procedure to develop interpretable ML models for estimating MID in existing RC buildings. The proposed procedure was organized across two scales: large-scale and reduced-scale.

The large-scale ML model uses Gaussian Process Regression (GPR), a nonparametric technique that handles uncertainty and nonlinear relationships. GPR uses all candidate building attributes (such as geometry, material properties, and infill characteristics) and intensity measures (IMs) (such as peak ground acceleration, peak ground velocity, and spectral acceleration) as input variables, and MID as the output variable. This model explores the data and identifies the most relevant and influential input variables.

To interpret the large-scale ML model, SHapley Additive exPlanations (SHAP) values were used. SHAP values explain the contribution of each input variable to the output variable, revealing the most significant IMs and the patterns of their influence depending on the occupancy rate of the infills. Based on the SHAP values, the study selects a small subset of IMs that capture the main features of the data.

Furthermore, the reduced-scale ML model was obtained using genetic programming (GP), a symbolic ML technique that generates explicit mathematical expressions. GP uses the selected subset of IMs as input variables, and MID as the output variable. The researchers used this model to provide accurate and interpretable predictions of the MID for different types of RC frames.

To develop the ML models at both scales, the authors used simplified models of archetype buildings, which are representative of typical RC buildings in Italy. They adopted these models to reduce the simulation time to prepare large datasets. Additionally, they considered types of RC frames: pilotis frames, frames with infills, and bare frames. They also considered pulse-like seismic ground motions, characterized by a single predominant pulse of displacement or velocity, which can cause severe damage to buildings.

Furthermore, to assess the performance and validity of the ML models, the researchers utilized refined models of actual buildings, which are more realistic and detailed than the simplified models. They tested the ML models on unseen data and compared the results with those obtained from nonlinear dynamic analyses.

Research Findings

The outcomes showed that the proposed procedure successfully guided the development of interpretable ML models for estimating MID in existing RC buildings. The large-scale ML model captured the nonlinear and complex relationships between input and output variables, and SHAP values provided useful insights into the data and the model. The reduced-scale ML model generated simple and explicit expressions that accurately predicted MID for different RC frames under pulse-like earthquakes.

The authors demonstrated that the proposed procedure managed a non-homogeneous building stock with diverse structural systems, such as buildings designed against gravity loads only or those complying with outdated seismic codes. They showed that the procedure could provide satisfactory predictions of MID with minimal computational effort, comparable to or better than those obtained from other ML techniques or simplified methods.

Conclusion

In summary, the novel technique proved effective in guiding the development of ML models for estimating MID in existing RC buildings under pulse-like earthquakes. It successfully handled a non-homogeneous building stock with diverse structural systems and can be applied to identify critical buildings, evaluate retrofitting strategies, optimize IM selection, and enhance seismic understanding.

Future work should extend the procedure to other building types, such as steel or masonry structures, and those with irregularities or torsional effects. The method could also be applied to different seismic ground motions, including near-fault or long-period motions, and other EDPs, like floor accelerations or base shear. Incorporating uncertainty and variability in ML models, and integrating ML with different methods, such as empirical or analytical approaches, neural networks, or support vector machines, could further enhance the procedure.