AI and the Future of Seismic Performance Prediction in Urban Centres

Predicting exactly when and how concrete slabs will fail during an earthquake has long relied on slow, computationally heavy engineering models. This historical limitation means city planners often lack rapid tools for accurate seismic performance prediction. Now, an early-stage preprint study suggests artificial intelligence could bypass this bottleneck entirely.

The research focuses on slab-column joints within flat plate structures. These specific connections are highly sensitive to seismic forces and are prone to sudden, brittle failure. When the earth moves, these structural nodes take the brunt of the kinetic stress.

Advancing Seismic Performance Prediction

To address this, the researchers tested an extensive array of Machine Learning (ML) and Deep Learning (DL) algorithms. Their goal was to forecast two specific metrics: punching moment and drift ratio. As an early-stage preprint, the findings currently apply to these specific structural components rather than whole-building dynamics, but the trajectory is clear.

The team measured the accuracy of various algorithms, including Gradient Boosting, Random Forest, and Convolutional Neural Networks (CNN). For forecasting the punching moment, Gradient Boosting achieved the highest accuracy among the ML models, measuring a coefficient of determination of 0.87. When evaluating the drift ratio, Random Forest performed best within the ML category.

Interestingly, across the deep learning options, CNNs consistently led the pack for both metrics. While CNNs are typically associated with image recognition, this study suggests they may also excel at identifying spatial patterns in structural stress data.

The Next Decade of Urban Resilience

What does this mean for the next five to ten years of civil engineering? If these early-stage algorithms are further validated and expanded, they could alter how we design and retrofit urban centres. Engineers might soon feed architectural plans into an AI system to instantly flag vulnerable joints before a single concrete pour occurs.

The downstream applications could be highly practical for structural management:

• Architects and engineers could rapidly test thousands of slab-column configurations to find the most resilient designs in minutes rather than months.
• Construction firms might optimise material use, reinforcing only the specific connections that algorithms identify as high-risk.
• Retrofitting projects could become more targeted, focusing resources on the exact structural nodes most likely to experience brittle failure.

Moving forward, this approach to structural analysis could transition from academic theory into standard engineering software. Algorithms will not stop tectonic plates from shifting. However, integrating predictive models into our building codes suggests that minimising structural failure may soon become a faster, more accurate science.