Machine Learning Rewrites the Rules for Judging How Buildings Survive Earthquakes

For decades, earthquake engineers have relied on a single, deceptively simple number to decide whether a reinforced concrete building will protect its occupants in a major quake: the peak inter-story drift ratio, a measure of how far one floor shifts sideways relative to the floor below it. Building codes around the world, including the linear seismic design procedures prescribed in modern standards, treat this number as a universal yardstick. If the drift stays below a prescribed limit, the structure is deemed to meet its performance target. But a growing body of experimental evidence has exposed a fundamental flaw in that assumption. Two beam-column connections with identical peak drift ratios can exhibit dramatically different levels of damage, strength loss, and residual capacity, depending on their geometric proportions, reinforcement details, material strengths, and loading histories. A new study published in the Bulletin of Earthquake Engineering argues that this one-size-fits-all approach is no longer defensible, and it proposes a data-driven alternative that could reshape how engineers verify seismic performance.

The research, conducted by Hanieh Feyzi, Pouya Khosravi-Hajivand, and Mohammadjavad Hamidia of Shahid Beheshti University in Tehran, introduces a machine learning-aided methodology for identifying the seismic performance level of reinforced concrete moment-resisting frames. Rather than relying on fixed drift thresholds, the framework learns from what laboratory tests have actually shown about how connections behave. At its core is an experimental database containing force-deformation curves from 236 reinforced concrete beam-column connection subassemblies subjected to cyclic loading, meaning the specimens were pushed back and forth repeatedly in the way earthquake shaking alternately loads and unloads a structure. This database spans decades of testing campaigns from laboratories around the world, encompassing joints with widely varying detailing practices, from older non-ductile configurations to modern seismically detailed assemblies.

The team aligned their analysis with the performance framework of ASCE 41-23, the American Society of Civil Engineers standard for seismic evaluation and retrofit of existing buildings. That standard defines discrete performance levels, including Immediate Occupancy, at which a structure remains essentially functional with only minor damage; Damage Control, a intermediate state limiting degradation; Life Safety, at which significant damage occurs but the structure retains margin against partial collapse; and Collapse Prevention, the threshold beyond which the building is on the verge of losing vertical load-carrying capacity. The researchers extracted the peak drift ratios at which each of these performance levels was reached for every specimen in the database. What they found was striking: for specimens with diverse structural characteristics, the range of drift ratios associated with any single performance level was remarkably broad. A drift ratio that signals life-threatening damage in one connection might correspond to barely perceptible cracking in another.

This variability is precisely why the authors contend that the peak drift ratio alone cannot fully characterize the seismic performance of a structure. The finding has profound implications for engineering practice. Under current code-based linear design procedures, an engineer designs a frame, computes the projected peak drift ratio, and checks it against a single prescribed limit. If the number passes, the design is accepted. Yet the experimental record shows that the same projected drift can land a building in radically different damage states depending on the attributes of its beam-column connections, the critical regions where beams frame into columns and where seismic energy is dissipated. In effect, the code’s universal yardstick measures something real, but it measures it with a tolerance so wide that the actual performance outcome remains largely unknown until an earthquake reveals it.

To close that gap, the researchers turned to supervised classification. They trained and compared nine shallow machine learning models alongside one deep learning model, tasking each with mapping structural and geometric attributes of a connection, together with the projected peak drift ratio from code-based linear analysis, to the correct ASCE 41-23 performance level. The standout performer was the Extra Trees algorithm, an ensemble method that builds many randomized decision trees and aggregates their votes. On the held-out testing dataset, the Extra Trees classifier achieved 88 percent accuracy, a level of reliability that the authors argue is sufficient for practical design support. The deep learning model, despite its greater architectural complexity, did not outperform the shallow ensemble, a result consistent with a broader pattern in engineering applications where tabular data of moderate size favors tree-based learners over neural networks.

Robustness was a central concern of the methodology. Machine learning models in engineering risk a subtle failure mode: they can appear accurate while merely memorizing the quirks of their training data, a phenomenon known as overfitting. To guard against this, the team determined the optimal hyperparameters of each model through Bayesian optimization, implemented via the BayesSearchCV framework, which uses probabilistic surrogate models to search the hyperparameter space efficiently rather than exhaustively. They then validated generalizability using nested five-fold cross-validation, a rigorous scheme in which hyperparameter tuning and performance estimation are carried out on strictly separated data partitions. Nested cross-validation is widely regarded as one of the most trustworthy protocols for estimating how a model will behave on unseen data, and its use here signals a deliberate effort to meet the evidentiary standards of structural engineering rather than those of a machine learning benchmark competition.

Transparency posed another challenge. Black-box predictions carry little weight in a field where lives depend on the reasoning behind a design decision. To open the box, the researchers applied SHAP analysis, or SHapley Additive exPlanations, a technique borrowed from cooperative game theory that quantifies the contribution of each input feature to every individual prediction. The SHAP assessment revealed which structural and geometric attributes most strongly shaped the model’s judgments about performance levels, giving engineers a window into the learned relationships and a means of checking them against physical intuition. This interpretability layer matters not only for professional acceptance but also for scientific scrutiny, since it allows other researchers to verify that the model has learned plausible mechanics rather than statistical artifacts.

The framework’s practical value was demonstrated through a seismic design case study of a reinforced concrete moment frame. Following conventional code-prescribed linear procedures, the study showed how the projected peak drift ratio, fed into the trained classifier together with the relevant structural descriptors, yields a probabilistic identification of the resulting performance level. Instead of a binary pass-fail check against a fixed drift limit, the engineer receives a data-informed assessment of whether the design is likely to achieve Immediate Occupancy, Damage Control, Life Safety, or Collapse Prevention. This capability fits squarely within the broader movement toward performance-based seismic design, which seeks to give owners and engineers explicit control over the damage state a building will experience at given hazard levels, rather than merely ensuring that code minimums are met. Recent comparative studies have shown that performance-based approaches can produce materially different, and often more rational, designs than conventional force-based procedures.

The study also connects to a rapidly expanding research program on data-driven earthquake engineering. In recent years, machine learning has been applied to estimate drift capacity of reinforced concrete walls, classify earthquake damage to buildings from sensor and inspection data, predict seismic demands on steel moment frames, and even infer damage states from the visual features of surface crack patterns using computer vision. The present work extends this lineage in a distinctive direction: rather than assessing damage after an earthquake, it embeds learning models directly into the design workflow, using the outputs of standard linear analysis as inputs to a far more nuanced performance evaluation. The authors’ related research on post-earthquake performance identification using residual drift ratios suggests a complementary pathway, in which the same learning infrastructure could serve both pre-event design verification and post-event damage assessment.

The implications for practice are considerable. Codes such as ASCE 7 and ACI 318, along with Eurocode 8, continue to anchor seismic design in linear analysis with drift checks, and a wholesale replacement of these procedures is neither imminent nor necessarily desirable. But the study’s evidence suggests that supplementing the drift check with a learned performance classifier, calibrated on hundreds of laboratory tests and validated with rigorous statistical protocols, could give engineers a far sharper picture of what their designs will actually do when the ground shakes. The broad scatter of drift limits behind every code threshold means that many designs accepted today carry hidden uncertainty about their true performance level. A framework that reduces that uncertainty, while remaining interpretable and compatible with existing workflows, offers a pragmatic bridge between the codified simplicity of the past and the performance-based precision that the discipline has long aspired to. As experimental databases grow and learning models improve, the era in which a single number could stand in for a building’s seismic fate may finally be drawing to a close.

Subject of Research: Machine learning-aided performance-based seismic design of reinforced concrete moment-resisting frames

Article Title: Performance-based machine learning-aided seismic design methodology for reinforced concrete moment frames

Article References: Feyzi, H., Khosravi-Hajivand, P., & Hamidia, M. (2026). Performance-based machine learning-aided seismic design methodology for reinforced concrete moment frames. Bulletin of Earthquake Engineering. https://doi.org/10.1007/s10518-026-02598-3

Image Credits: AI Generated

DOI: 10.1007/s10518-026-02598-3

Keywords: seismic design, machine learning, reinforced concrete, moment-resisting frames, beam-column connections, ASCE 41-23, performance-based design, peak drift ratio, Extra Trees, Bayesian optimization, SHAP, earthquake engineering