Assessing machine learning for seismic performance: A preliminary look at building safety

Engineers have identified Gradient Boosting and Convolutional Neural Networks as the most accurate computational models for predicting how concrete slab-column joints fail during earthquakes. Achieving this level of predictive accuracy has historically been difficult because the physical forces involved in sudden structural collapses are highly complex and non-linear. The application of machine learning for seismic performance offers a new method to anticipate these brittle failures before buildings are actually constructed.

The context of machine learning for seismic performance

Flat plate structures rely on connections where horizontal slabs meet vertical columns. During seismic activity, these joints are highly vulnerable to brittle failure, a scenario where the concrete suddenly punches through the column.

Traditionally, civil engineers have relied on established empirical formulas to estimate connection safety. By replacing conventional calculations with dynamic algorithms, researchers hope to model the breaking point of building materials more accurately. This approach evaluates complex data to find predictive patterns human engineers might miss.

Evaluating the predictive algorithms

In this early-stage, non-peer-reviewed preprint research, investigators tested twelve different algorithms to predict two specific metrics: punching moment and drift ratio. The study compared traditional machine learning against deep learning alternatives.

The algorithms analysed included:

• Traditional linear models like Ridge and Lasso Regression.
• Ensemble methods such as Random Forest and Gradient Boosting.
• Deep learning architectures including Convolutional and Recurrent Neural Networks.

For punching moment, Gradient Boosting outperformed all other models, achieving a coefficient of determination (R²) of 0.87. Among the deep learning options, Convolutional Neural Networks (CNN) performed best with an R² of 0.82. The R² value measures how well the model explains the variance in the data; a score of 0.87 indicates a strong, though imperfect, correlation.

Drift ratio proved much harder to predict. Random Forest models achieved the highest accuracy here, though only reaching an R² of 0.56. This lower score highlights a significant gap in the predictive capability of current algorithms.

Current limitations in the modelling

Despite these promising metrics, the research does not yet solve the problem of real-world, dynamic application. The markedly lower accuracy in predicting drift ratio suggests the current algorithms struggle to capture the full complexity of building movement, at least based on the specific parameters evaluated in this study.

Implications for structural engineering

If validated through peer review, these findings suggest that structural design could become heavily reliant on algorithmic pre-testing. Engineers might soon optimise joint dimensions and reinforcement strategies entirely in software before pouring any concrete.

While physical testing will remain necessary, this early-stage research indicates a practical shift in how safety margins are calculated. By identifying the specific algorithms best suited for different stress metrics, the study provides a focused direction for future computational engineering.