The core problem the researchers set out to solve is one of granularity. Existing damage state definitions for reinforced concrete structural walls, found in standards such as ASCE/SEI 41, FEMA 356, Eurocode 8 and various performance-based design guidelines, are based largely on empirical limits derived from past laboratory tests and post-earthquake observations. These limits typically specify a single drift ratio, the lateral deflection of the wall expressed as a percentage of its height, at which a wall is deemed to have reached a given damage state, from hairline cracking through concrete spalling to reinforcement buckling and outright collapse. The difficulty is that no two walls are identical. A tall, slender wall with well-confined boundary elements and heavy longitudinal reinforcement can sustain far larger deformations than a stockier, lightly reinforced counterpart, yet the traditional limits largely ignore these differences. Sarkar and Dasgupta argue that this blind spot leads to predictions that can be either dangerously optimistic or needlessly conservative, both of which carry costs in seismic risk assessment, loss estimation and retrofit decision-making.

Their solution draws on a dataset of 8,125 analytically studied specimens of reinforced concrete structural walls, an unusually large numerical sample for this class of component. The virtual specimens were generated using nonlinear finite element modeling in the OpenSees platform, with material constitutive behavior represented by well-established models, including the concrete model of Scott, Park and Priestley for confined and unconfined concrete and the Menegotto-Pinto steel model for reinforcing bars, along with provisions for post-yield buckling of reinforcement based on the work of Dhakal and Maekawa and on buckling strain formulations by Papia and Russo. By systematically varying the key geometric and reinforcement parameters across thousands of virtual walls and subjecting each to simulated cyclic loading histories, the researchers built a comprehensive numerical laboratory capturing the deformation capacity of walls spanning a wide range of practical design configurations.

The intelligence at the heart of the study is genetic programming, an evolutionary computation technique pioneered by John Koza in the 1990s. Unlike conventional regression, which forces data into a predefined functional form, or black-box machine learning methods such as neural networks, which produce predictions without a transparent underlying formula, genetic programming evolves mathematical expressions directly from data. The algorithm begins with a randomly generated population of candidate equations built from mathematical operators and the input variables of interest. Each candidate is tested against the dataset and scored for accuracy. The fittest expressions are then selected, and their mathematical “genes” are recombined and randomly mutated to produce a new generation, mimicking natural selection. Over many generations, the population of equations converges toward compact, accurate formulas that capture the underlying physics without any human-imposed structure. The critical advantage for engineering practice is that the output remains a closed-form expression that any designer can evaluate with a hand calculator or spreadsheet, preserving full transparency about which parameters matter and how they interact.

In this application, the input variables chosen by the researchers reflect the design characteristics that structural engineers most plausibly control at the drawing board. The developed drift prediction expressions account for the wall aspect ratio, which governs whether a wall responds primarily in flexure or shear; the axial load ratio, which influences concrete confinement demand and deformation capacity; the boundary element longitudinal reinforcement ratio, which controls the flexural strength and ductility of the carefully detailed edges of the wall where plastic hinges typically form; the web longitudinal reinforcement ratio, which affects distributed cracking behavior across the wall panel; and the ratio of boundary element length to wall length, which determines the extent of confined concrete available at the wall ends. Each of these parameters is readily available from standard design documents, which means the resulting equations can be applied directly to both new designs and existing buildings without specialized analysis.

The statistical rigor applied to the resulting models reflects a discipline that has sometimes been lacking in machine learning studies in structural engineering. The researchers evaluated their expressions using multiple goodness-of-fit measures, including the coefficient of determination and correlation statistics, and they conducted external validation on data not used in model development, following the cautionary principles articulated by Golbraikh and Tropsha regarding the necessity of external validation for predictive models. The predictions were also benchmarked against established deformation limits from design standards and against drift capacity expressions reported in the peer-reviewed literature, and then tested against experimental results of reinforced concrete wall components from laboratory test programs. On every measure, the genetically evolved expressions outperformed the existing empirical damage state definitions, delivering significantly higher accuracy and more reliable predictions across the full spectrum of wall configurations in the dataset.

The implications for practice are considerable. Quantitative damage state thresholds are the backbone of performance-based seismic design, the philosophy that buildings should be engineered not merely to avoid collapse but to achieve specified performance targets, from immediate occupancy after a frequent moderate earthquake to life safety in a rare major event and collapse prevention in the rarest, strongest shaking. Fragility analyses, which estimate the probability that a structure will exceed a given damage state under a given level of shaking, and seismic loss estimation frameworks used by insurers and governments both depend on credible component-level damage thresholds. If the thresholds for reinforced concrete walls can be made sensitive to actual design characteristics rather than generic constants, then the entire chain of risk assessment, from component behavior to building performance to portfolio-level loss, becomes more accurate. This is particularly consequential in high seismic zones across Asia, the Americas and the Mediterranean, where frame-wall systems are the dominant structural choice for residential and commercial high-rise construction.

The study also situates itself within a rapidly growing body of work applying artificial intelligence to earthquake engineering. Recent contributions in the same journal and elsewhere have used neural networks to identify pulse-like ground motions, genetic algorithms for structural vibration control, machine learning for seismic fragility analysis and failure mode identification of shear walls, and hybrid evolutionary algorithms for predicting ground vibration. Sarkar and Dasgupta’s contribution to this wave is distinctive in its emphasis on symbolic, interpretable expressions rather than opaque predictive models, and in its focus on damage state drift thresholds rather than strength or failure mode prediction alone. Their work builds directly on earlier efforts such as the mechanics-guided genetic programming formulation for shear strength of squat walls by Gondia and colleagues, extending the approach to deformation-based damage limit states for slender rectangular walls, which behave quite differently from their squat counterparts.

The authors are careful to frame the scope of their achievement. The expressions were calibrated on analytically generated data covering slender rectangular walls, and while the underlying finite element models were grounded in experimentally validated constitutive laws, the ultimate test of any numerical prediction framework is correspondence with physical tests. The researchers addressed this by comparing their damage state drift predictions with experimental results reported in the literature, including cyclic and quasi-static tests of isolated walls conducted over several decades at laboratories in the United States, Canada, Europe and Asia. The favorable comparison with these tests provides an important bridge between the computational dataset and real structural behavior, and it suggests that the framework could be extended in future work to walls with different cross-sectional shapes, opening configurations, higher-order reinforcement detailing or different construction materials.

For the engineering community, the immediate payoff is a set of practical, parameterized drift expressions that can be incorporated into next-generation seismic evaluation guidelines and displacement-based design procedures. For the wider public, the significance is quieter but no less real: better predictions of when the walls holding up a building will crack, spall or fail translate directly into smarter building codes, better-targeted retrofits and more trustworthy estimates of how many people a future earthquake might harm. As machine learning and evolutionary computation continue to penetrate structural engineering, this study offers a compelling demonstration that the goal need not be mysterious artificial intelligence making inscrutable judgments. Sometimes the most valuable form of machine intelligence is the kind that hands the engineer a short, elegant equation, one that has been rigorously evolved from thousands of simulated earthquakes’ worth of structural response, and that tells her, with greater fidelity than decades of empirical rules ever could, exactly how far a wall can bend before the damage becomes real.