Can AI make earthquake simulations more reliable? A new hybrid model says yes

Istanbul, Türkiye - When engineers assess whether a building can withstand a major earthquake, they increasingly rely on sophisticated computer simulations. Yet even advanced models may require experienced engineers to manually calibrate key parameters using experimental data before the simulations can adequately reproduce observed structural behavior.

Researchers from Istanbul Technical University (ITU) and the University of California, Los Angeles (UCLA) have developed a new approach that enables artificial intelligence to perform much of the expert calibration automatically, while keeping engineering decisions firmly grounded in the laws of physics. Published in the Journal of Building Engineering, the study presents a hybrid modeling framework that combines machine learning with validated finite element simulations to improve predictions of how reinforced concrete shear walls, the primary structural elements that help buildings resist earthquakes, behave under severe seismic loading.

Rather than asking AI to predict whether a building will fail, the researchers designed it to serve as an intelligent assistant that improves the engineering simulation itself. That distinction matters. Many recent AI applications in structural engineering rely on machine learning as a "black box," producing predictions that can be difficult for engineers to interpret or verify. In contrast, the new framework keeps decades of established structural mechanics at the heart of the analysis. Machine learning is used only to estimate one of the simulation's most challenging calibration parameters, enabling the physics-based model to produce more reliable and consistent results.

"There is growing interest in using artificial intelligence to replace engineering calculations," said Dr. Zeynep Tuna Deger, corresponding author and researcher at Istanbul Technical University. "Our work takes a different approach. We use AI to assist the engineer, not replace engineering. The physics remains transparent, while AI helps remove one of the most challenging and time-consuming parts of the modeling process."

The researchers focused on reinforced-concrete shear walls, which play a critical role in preventing buildings from collapsing during earthquakes. Simulating their behavior is particularly challenging because engineers must account for multiple interacting failure mechanisms, including concrete crushing, reinforcing-bar buckling, low-cycle fatigue, shear deformation, and strain localization as damage accumulates.

Although sophisticated finite element models already exist, their accuracy often depends on manually calibrating material properties using laboratory test data—a process that is impractical for evaluating large numbers of buildings or new structural designs when experimental results are unavailable. The new framework addresses this challenge by using machine learning to estimate the optimal calibration parameter directly from measurable wall characteristics, including geometry, reinforcement configuration, concrete strength, and loading conditions. That parameter is then incorporated into an established OpenSees finite element model that simulates the wall's physical response during an earthquake.

To validate the approach, the team tested the framework on experimental data from 140 reinforced concrete wall specimens, one of the larger datasets assembled for this type of nonlinear structural modeling. Across the dataset, the hybrid approach accurately reproduced wall displacement capacity while preserving the physical transparency expected by structural engineers. Because the final predictions still stem from a validated engineering model rather than a purely data-driven algorithm, engineers can understand how the results are produced and evaluate whether they are physically reasonable, an increasingly important consideration as AI becomes integrated into safety-critical infrastructure design.

"Engineering decisions are fundamentally different from those in many other AI applications because lives may depend on the outcome," said Dr. Zeynep Tuna Deger. "Our objective was to build a system that combines machine learning's efficiency with the transparency and scientific rigor engineers require."

The researchers believe the concept extends beyond reinforced-concrete walls. Similar hybrid approaches could eventually enable simulations of bridges, tunnels, industrial facilities, and other critical infrastructure where accurate, explainable engineering models are essential. As governments and engineers worldwide explore how artificial intelligence can support infrastructure resilience, the study suggests that the future may not lie in replacing engineering expertise with AI, but in combining AI with decades of validated physical knowledge.

For earthquake-prone regions facing aging infrastructure, rapid urbanization, and growing demand for structural assessment, such hybrid models could help engineers evaluate building performance more efficiently and consistently, providing greater confidence before the next major earthquake.