Dry-joint stone construction, in which precisely shaped blocks are stacked without mortar, is found across the globe, from Inca Peru to ancient Greek fortifications and medieval retaining walls. The technique produces remarkably durable structures, but it also leaves them with a fundamental weakness: with nothing gluing the stones together, the walls depend entirely on gravity and friction for their stability. When an earthquake shakes a wall perpendicular to its face, a loading direction engineers call out-of-plane, the blocks can slide, rock and ultimately overturn in a cascade of falling stones. Understanding the threshold at which that overturning begins is one of the central problems in the seismic assessment of historic masonry.

The research team, led by Leonel Lipa of the Pontificia Universidad Católica del Perú and Universidad San Ignacio de Loyola, together with Nicola Tarque of the Universidad Politécnica de Madrid and Luca Pelà of the Universitat Politècnica de Catalunya, chose one of the most iconic examples of the technique as their template. The Coricancha temple, whose polished stone walls later formed the foundations of the convent of Santo Domingo, is celebrated for its trapezoidal openings, gently inclined wall faces and joints so tight that a knife blade cannot slip between the blocks. Reproducing those characteristics at reduced scale was the first challenge of the project.

Rather than carving scarce and protected andesite, the researchers cast concrete blocks whose geometry and arrangement mirrored a section of the Coricancha’s dry-jointed stonework. Three walls were built at 1:2 scale, each following the block layout, aspect ratios and construction details of the original heritage fabric as closely as the material substitution allowed. The concrete mix was characterised through a battery of standardised tests covering compressive strength, density and aggregate properties, so that the mechanical inputs to the numerical models rested on measured values rather than assumptions. Building the replicas block by block also meant that every joint in the test walls corresponded to a joint in the real structure, preserving the discontinuous nature that governs how these assemblies move.

The test rig was a tilting table, a device that slowly rotates a platform on which the wall stands until gravity’s effective direction tilts away from vertical. This pseudo-static approach is a well-established way to probe out-of-plane stability: as the table inclines, the wall experiences an increasing horizontal component of its own weight, mimicking the lateral forces an earthquake would impose. The researchers tilted each of the three walls until collapse, recording the angle at which the structure first showed visible movement, the sequence of block displacements that led to failure, and the final angle at which overturning became irreversible. These pre-collapse drifts, collapse mechanisms and collapse angles form the experimental benchmark against which any numerical model must be judged.

What the tests revealed was a consistent pattern of behaviour. The walls did not crumble randomly; instead they developed well-defined failure mechanisms in which portions of the masonry separated along joint lines and rotated as coherent groups of blocks. This is the hallmark of dry-joint masonry: because the blocks carry almost no tensile strength across their joints, deformation concentrates in the interfaces, and the wall behaves less like a continuous solid and more like an assembly of rigid bodies pushing against one another. Capturing that behaviour numerically is difficult for conventional finite element codes, which are built around continuous materials and can struggle with the opening, closing, sliding and impacting of many separate contact surfaces.

The team’s answer was a hybrid approach that embeds rigid-body mechanics within a finite element framework. In this formulation, each stone is idealised as a rigid block, an idealisation justified by the fact that the stresses within individual stones remain far below the material’s strength, so the blocks themselves barely deform. All of the action happens at the joints, where the model accounts for contact, friction and the small local deformations that occur where two stone faces meet. The numerical simulations were set up with the geometry of the scaled walls and the measured material and interface properties, then subjected to the same progressive tilting as the physical specimens.

The agreement between experiment and simulation was the study’s decisive result. The numerical models accurately predicted the collapse angles at which the walls overturned, reproduced the experimentally observed failure patterns, and matched the average drift measured at various inclination angles throughout the loading history. In other words, the models did not merely get the final failure threshold right; they tracked the wall’s response along the way, capturing how the structure leaned, cracked open along its joints and redistributed its weight as the tilt increased. That level of fidelity matters, because a model that only matches the collapse angle could do so for the wrong reasons, while one that reproduces the full deformation path gives engineers genuine confidence in its predictions.

For heritage custodians, the implications are immediate. Structures like the Coricancha cannot be instrumented, loaded or dismantled to assess their safety, so any evaluation must rely on numerical analysis. A validated modelling strategy means engineers can now build a digital twin of a real Inca wall, using its surveyed geometry and representative material properties, and estimate with quantified confidence how close it sits to overturning under the seismic demands of a region as earthquake-prone as the Peruvian Andes. The approach also opens the door to testing hypothetical interventions digitally, allowing conservators to compare strengthening options before committing to any physical alteration of protected fabric.

The work also fits into a broader research effort by the same groups, who have previously applied rigid-body dynamics to the colossal terraced walls of Sacsayhuamán above Cusco and developed a taxonomy of pre-Hispanic stone wall types in southern Peru. By anchoring the new models to physical experiments on faithful scaled replicas, the team has closed the loop between field observation, laboratory evidence and computational prediction. The dataset from the study has been released in the Zenodo repository so that other researchers can scrutinise and extend the results. As seismic risk grows across the mountainous regions where dry-stone heritage concentrates, the image of three concrete replicas of an Inca temple wall toppling on cue in a Lima laboratory may prove to be a turning point: the moment the ancient builders’ empirical mastery of friction and gravity finally became something modern engineering could simulate, verify and defend.

Subject of Research: Out-of-plane seismic behaviour and rigid-body numerical modelling of historical dry-joint stone walls based on scaled Coricancha temple replicas

Article Title: Experimental and numerical analysis of out-of-plane behaviour of historical dry-joint stone walls: insights from a scaled laboratory wall model of the Coricancha temple, Peru

Article References: Lipa, L., Tarque, N., Goicolea, J. M., Santa-Cruz, S., Quispe, L., Pillaca, E., & Pelà, L. (2026). Experimental and numerical analysis of out-of-plane behaviour of historical dry-joint stone walls: insights from a scaled laboratory wall model of the Coricancha temple, Peru. Bulletin of Earthquake Engineering. https://doi.org/10.1007/s10518-026-02710-7

Image Credits: AI Generated

DOI: 10.1007/s10518-026-02710-7

Keywords: dry-joint masonry, out-of-plane behaviour, Coricancha temple, Inca architecture, rigid body dynamics, finite element method, tilting table test, collapse mechanism, heritage structures, seismic assessment, scaled wall model, earthquake engineering

The method, called IOPER, short for Inverse-Optimized Pushover with Energy Redistribution, was developed by Jose Mendoza, Alfredo Canelas, and Berardi Sensale. Its central idea is deceptively simple: instead of guessing how seismic demand should be distributed over the height of a building, let the structure itself reveal the answer. In the first stage, the analysts run a conventional displacement-controlled pushover using a modal force pattern, usually based on the dominant mode shape, which in their benchmark buildings accounted for roughly 78 percent of the effective modal mass. As the structure is pushed, the software records the hysteretic energy dissipated at every node, obtained by numerically integrating the enclosed area of each force-deformation loop. That energy is then aggregated story by story, producing a vertical profile of where the building actually absorbed its inelastic demand.

What happens next is the methodological heart of IOPER. The dissipated energy at each level is multiplied by a residual rigidity coefficient, computed through a logarithmic scheme that penalizes stories with high cumulative hysteretic demand. The logarithmic form is deliberately sublinear: it compresses peak energy values while preserving the contrast between stories, mirroring the saturation behavior seen in dynamic analyses, where stiffness loss slows even as energy demand keeps climbing. Damaged stories therefore retain residual capacity and continue participating in the global response rather than being artificially erased. The resulting weighted vector is normalized and exported as the lateral force pattern for a second pushover run on the original, undegraded model. A third iteration repeats the process using the energy traces of the second run, tracing progressively advanced redistribution scenarios. The output is not a single curve but a triplet of capacity curves, each representing a distinct interpretive layer: modal response, energy-induced redistribution at an initial degraded state, and redistribution at a final degraded state.

This architecture is deliberately external to the simulation engine. Where fully adaptive pushover methods, such as the Force-Based Adaptive Pushover procedure of Antoniou and Pinho, recalibrate the load pattern at every analysis step inside the solver using updated mode shapes and spectral accelerations, IOPER performs its adaptivity outside, through MATLAB scripting that reads SeismoStruct output files, integrates hysteretic loops, computes degradation coefficients, and writes new force vectors back into the model. The authors argue that this decoupling is a feature rather than a compromise. Many commercial platforms restrict access to internal state data, lack restart capabilities, or bury their adaptive logic in undocumented code. By making every redistribution decision explicit, auditable, and reproducible through plain text files, IOPER offers a quasi-adaptive alternative that practitioners can inspect, modify, and trust, without iterative eigenvalue calculations at every step.

The validation campaign is substantial. The team tested the framework on two benchmark steel moment-resisting frames originally developed for the ASCE structural control problem: a nine-story building with a square footprint 45.73 meters per side and a first-mode period of 2.28 seconds, and a twenty-story tower rising 80.77 meters with a fundamental period of 3.89 seconds. Both frames qualify as vertically irregular. The nine-story structure shows a stiffness drop at the second level to less than 25 percent of the base value, a classic soft-story condition, while the twenty-story building combines a stiffness reduction below 21 percent at the third floor with a mass jump exceeding 150 percent of the floor below. These irregularities matter because they are precisely the conditions under which fixed load patterns fail, and where elevation-sensitive redistribution strategies earn their keep.

The dynamic benchmark comprised seven real ground motion records, including El Centro, Kobe, Northridge, Landers, Hollister, Loma Prieta, and Imperial Valley, spectrally matched to the Eurocode 8 design spectrum for Ground Type B and scaled to four intensity levels ranging from 0.45 g to 0.90 g. Nonlinear time history analyses used a fiber-based distributed plasticity model with 150 fibers per section, force-based elements, a co-rotational formulation for geometric nonlinearity, and the implicit Hilber-Hughes-Taylor integration scheme. Against this demanding reference, IOPER held its own. In the elastic range, all static methods converged on similar stiffness, but beyond it the differences emerged. The ASCE 41 uniform-load procedure rose monotonically to 7182 kilonewtons on the nine-story frame, blind to degradation, while FAP and IOPER saturated near 6479 and 6518 kilonewtons respectively, closer to the physically plausible post-elastic behavior of the dynamic envelope, which peaked at 7509 kilonewtons. Notably, both adaptive-family methods exhibited a capacity recovery phase beyond 0.6 meters of roof displacement, with IOPER closely following the shape of the incremental dynamic analysis mean curve.

Quantitative metrics reinforced the visual comparisons. At mid- and upper-story nodes, the researchers measured the mean absolute deviation of the IOPER curve from the pointwise median of the NLTHA ensemble and the proportion of drift points at which IOPER fell inside the absolute dynamic envelope. Deviations stayed below 0.05 in normalized base shear across all examined nodes, and the ensemble inclusion ratio ranged from about 71 to 91 percent, peaking at 90.8 percent for the mid-story node of the twenty-story frame. For total drift, root-mean-square errors against NLTHA dropped below 11 percent at the highest intensities, with FAP reaching 1.62 percent on the nine-story frame at 0.90 g. The study also found a meaningful correlation between the energy content of the scaled accelerograms and IOPER’s accuracy: records with high displacement and velocity demand, such as Loma Prieta and Imperial Valley, gave the redistribution logic richer energy traces to work with, improving predictions at severe intensities.

The framework is not without honest limitations, which the authors state plainly. Because the structural model is reset at each iteration, plasticity is not accumulated inside the solver; degradation lives entirely in the external energy metrics and rigidity coefficients. IOPER therefore cannot capture transient amplification, record-to-record variability, or instantaneous modal reshaping, and its accuracy degrades for interstory drift, particularly in upper stories of tall frames, where global errors exceeded 40 percent at 0.90 g on the nine-story building and 60 percent on the twenty-story tower. A refined metric restricted to stories meeting a 15 percent agreement threshold revealed that static methods perform well at lower elevations, with errors as low as 2.8 to 6.4 percent, but diverge near the roof, a persistent dynamic amplification phenomenon that the authors flag as a priority for future work through height-dependent corrections and transient-sensitive formulations.

Those caveats aside, the significance of the work lies in what it demonstrates about the philosophy of seismic assessment. IOPER treats a building’s response not as a geometric abstraction but as an accumulated energy narrative, in which the history of hysteretic dissipation shapes the forces applied next. It recontextualizes modal information as a baseline rather than a straitjacket, and shows that externally scripted control can emulate adaptive behavior without internal feedback loops, challenging the assumption that realism requires embedding adaptivity deep inside the solver. The authors position the framework for preliminary fragility assessment, redistribution-sensitive detailing, and comparative evaluation of modal activation, and note its compatibility with the spirit of ASCE 41 and Eurocode 8, both of which acknowledge stiffness degradation and force redistribution even though neither prescribes energy-based load patterns. Future extensions outlined in the paper include element-wise degradation models, calibration against experimental data for reinforced concrete and composite systems, integration with site-specific hazard spectra, extension to three-dimensional frames with torsional coupling, and evaluation under near-fault pulse-type motions. For a discipline caught between the fidelity of dynamic simulation and the tractability of static methods, IOPER offers a transparent middle path, one in which every force applied to the structure carries the memory of the energy it has already dissipated.

Subject of Research: A quasi-adaptive pushover framework using hysteretic energy redistribution for seismic assessment of steel moment-resisting frames

Article Title: IOPER: Inverse-optimized pushover with energy redistribution: A framework for enhanced control and degraded stiffness-aware seismic analysis

Article References: Mendoza, J., Canelas, A., & Sensale, B. (2026). IOPER: Inverse-optimized pushover with energy redistribution: A framework for enhanced control and degraded stiffness-aware seismic analysis. Results in Engineering, 32, Article 113209. https://doi.org/10.1016/j.rineng.2026.113209

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.113209

Keywords: pushover analysis, seismic assessment, hysteretic energy, steel moment frames, nonlinear static analysis, stiffness degradation, NLTHA, ASCE 41, Eurocode 8, adaptive load patterns, structural engineering, earthquake engineering

The research, led by Xiaohui Yu, Zenghui Li, Zihan He, Kuangyu Dai and Feng Fu, with affiliations spanning Guilin University of Technology, Zhengzhou University and City, University of London, focuses on the hysteretic behavior of reinforced concrete columns. Hysteresis describes the looping force–displacement curves a column traces as it is pushed back and forth past its elastic limit. These loops encode everything engineers need to know about a column’s energy dissipation, stiffness degradation and strength deterioration, but measuring them for real damaged structures is slow, expensive and often impossible after a disaster.

To build a foundation for their predictive models, the team first assembled a comprehensive database of 1,053 reinforced concrete column specimens tested under cyclic loading. For each column, the hysteretic response was simulated using the modified Ibarra–Medina–Krawinkler (I-M-K) model, a phenomenological framework widely used in seismic collapse assessment. The I-M-K model represents a column’s backbone curve — the envelope of its strength and stiffness — together with rules for cyclic deterioration, capturing how repeated inelastic cycles progressively erode yield strength, unloading stiffness, and post-capping behavior.

The central innovation lies in a two-stage simulation strategy designed to mimic what an earthquake actually does to a structure. In the first stage, a partial loading history drawn from the experimental data is applied to each intact column, producing a wide range of initial damage states, from hairline cracking to severe strength loss. In the second stage, the complete loading history is applied to these now-damaged columns, generating the full hysteretic response of a structure that has already been through a mainshock. This approach effectively turns one experimental record into many, allowing the researchers to explore the continuum between pristine and heavily damaged states without physically breaking hundreds of additional specimens.

From the simulated responses of the damaged columns, the researchers identified the backbone and cyclic deterioration parameters of the modified I-M-K model. By comparing the backbone parameters of damaged columns with those of their intact counterparts, they quantified damage-induced reduction factors, or DRFs, for key quantities including yield strength, elastic stiffness, strain hardening ratio, capping strength, and the chord rotation of the softening stage. These factors express, in precise numerical terms, how much of a column’s original capacity survives a given level of earthquake damage.

With the database of DRFs and deterioration parameters in hand, the team trained six different machine learning models to predict these values directly from a column’s initial damage state and its design parameters — inputs such as axial load ratio, longitudinal and transverse reinforcement ratios, concrete compressive strength, stirrup spacing, shear span-to-depth ratio, and cross-sectional dimensions. Rather than relying on a single arbitrary train-test split, the models were trained and tested across 1,000 random splits, a rigorous protocol that guards against lucky partitions and gives a robust picture of each algorithm’s true performance.

The verdict was clear: XGBoost, the gradient-boosted decision tree algorithm that has become a favorite in engineering applications, exhibited the best performance in predicting the hysteretic behavior of damaged columns. XGBoost builds an ensemble of decision trees sequentially, with each new tree correcting the errors of the previous ones, making it particularly adept at capturing the nonlinear interactions between design variables and damage states that govern structural response. To test that the model was not simply memorizing its training data, the researchers also evaluated it on three reinforced concrete column specimens outside the original database, examining its general applicability to unseen cases.

Crucially, the study does not treat its model as a black box. Using SHapley Additive exPlanations, or SHAP, a technique borrowed from cooperative game theory that attributes each prediction to the contributions of individual input features, the researchers interrogated what their model had learned. The analysis indicated that initial damage is the most influential feature governing the reduced hysteretic parameters — an intuitive but now quantitatively confirmed result, since the severity of pre-existing damage directly controls how much strength and stiffness remain for subsequent loading.

The practical implications extend well beyond the laboratory. After a damaging earthquake, engineers must rapidly decide whether a building is safe for occupancy, whether it needs retrofitting, or whether it should be demolished. Current post-earthquake assessment practice often relies on simplified damage indices, such as the Park–Ang damage index, and conservative assumptions that can lead to unnecessary closures or, worse, underestimated risk. A trained surrogate model like the one developed here could take measurable column properties and an estimate of sustained damage and return, in seconds, a full prediction of the residual hysteretic envelope — feeding directly into aftershock fragility analysis and resilience planning for entire building stocks.

The work also fits into a broader movement in earthquake engineering toward machine-learning-assisted seismic assessment, with recent studies applying similar methods to fragility estimation, drift capacity prediction, and parameter identification for nonlinear structural models. By releasing their column database as supplementary material and making the source code openly available on GitHub, the authors have lowered the barrier for other researchers to build on the approach. As cities in seismic regions face aging infrastructure and increasing exposure, tools that translate raw data into rapid, explainable predictions of residual capacity may become as essential to post-earthquake response as the seismometers that record the shaking itself.

Subject of Research: Machine learning prediction of the residual hysteretic behavior of post-earthquake damaged reinforced concrete columns

Article Title: Hysteretic behavior prediction of post-earthquake damaged reinforced concrete columns using machine learning

Article References: Yu, X., Li, Z., He, Z., Dai, K., & Fu, F. (2026). Hysteretic behavior prediction of post-earthquake damaged reinforced concrete columns using machine learning. Bulletin of Earthquake Engineering. https://doi.org/10.1007/s10518-026-02704-5

Image Credits: AI Generated

DOI: 10.1007/s10518-026-02704-5

Keywords: machine learning, XGBoost, reinforced concrete columns, hysteretic behavior, post-earthquake damage, seismic assessment, Ibarra-Medina-Krawinkler model, damage-induced reduction factors, SHAP, aftershock resilience, structural engineering, Bulletin of Earthquake Engineering

The core problem the researchers tackled is the nonlinear static analysis known in the engineering community as pushover analysis. In a pushover analysis, a finite element model of a building is subjected to a progressively increasing lateral load that mimics the inertial forces of an earthquake, and the response is tracked until key structural members reach their deformation limits. The procedure produces a capacity curve, which is typically simplified into a bilinear acceleration-displacement relationship characterized by an equivalent vibration period, a yield acceleration and an available ductility. These three quantities, evaluated for both the Damage Limitation and Severe Damage limit states and for different lateral load distributions, determine whether a building can survive a design-level earthquake. The difficulty is that each pushover analysis demands an hour or more of computation on a nonlinear finite element model, which makes portfolio-level assessment of millions of buildings an intractable task.

The team’s answer is a hybrid deep neural network that fuses two architectural paradigms: Long Short-Term Memory networks, which excel at modeling sequential dependencies, and Convolutional Neural Networks, which specialize in spatial pattern recognition. The central insight behind the design is that the seismic behavior of a frame building depends simultaneously on how forces propagate vertically through the storeys and on how stiffness and strength are distributed spatially across the floor plan. Neither an LSTM alone nor a CNN alone can capture both aspects, so the researchers constructed an architecture in which fifty parallel LSTM branches process the columns of the building column by column, storey by storey, while a downstream CNN analyzes the spatial arrangement of their outputs.

The input representation is one of the most innovative features of the study. Each building is described by three complementary sources of information: nine global parameters characterizing the structure as a whole, such as the number of storeys and material properties; eight local parameters for each individual reinforced concrete column, including its lateral stiffness, plastic shear capacity and rotational capacities in both plan directions; and a mask that encodes the building’s layout on a 50 by 50 grid spanning a 50 meter square, where each one meter cell records whether a column is present. Because the columns are aligned across storeys, a single mask suffices for the whole building. Every column along a given direction of the plan is represented as a 400-dimensional feature vector, and the stack of these vectors across the storeys forms a sequence whose length equals the number of storeys, allowing the LSTM branches to learn how structural effects accumulate from the foundations upward.

The fifty LSTM outputs are then arranged as columns of a new two-dimensional matrix and combined with the layout mask to form a two-channel image, which is processed by a convolutional network with two convolutional layers and max-pooling stages. The CNN output is concatenated with the nine global building parameters and passed through fully connected layers that produce twenty regression targets: the equivalent period, yield acceleration and kinematic ductility for both limit states, across four pushover directions. The authors emphasize that the mask acts as a spatial filter, telling the network where real columns exist and filtering out the empty background of the grid, which turns out to be critical for performance. A masking layer inside each LSTM branch similarly ignores padded entries, which is what allows the same architecture to process buildings with any number of storeys without modification.

Training the network required a substantial synthetic dataset. The researchers generated 22,576 finite element models of reinforced concrete frame buildings representative of Italian construction practice after World War II, when structures were designed for gravity loads only under outdated codes. Using the OpenSEES-Py framework and high performance computing resources with AMD EPYC processors, they carried out parallel pushover analyses on buildings of two to five storeys, a range that reflects the real Italian building stock, in which two-storey and three-storey structures respectively account for 44 and 26 percent of reinforced concrete buildings. The simulated design procedure reproduced the allowable stress method, typical gravity loads and the single concrete class R’bk 250 characteristic of that era, giving the dataset a realistic grounding in a vulnerable and widespread class of structures.

The trained model, which contains roughly 8.5 million parameters and occupies about 100 megabytes, achieved errors that the authors describe as negligible for practical engineering purposes. On the test set, the mean absolute error for the equivalent period averaged about 0.022 seconds, while the yield acceleration errors averaged around 0.013 to 0.014 meters per second squared, or roughly 0.0014 g, and ductility errors averaged about 0.014. Crucially, the model generalized across building heights despite an intentionally unbalanced dataset: the average error for two-storey buildings was nearly identical to that for five-storey buildings, indicating that the network did not simply overfit the majority class. Because the architecture uses padding and masking rather than a fixed sequence length, the authors note that the design could in principle handle single-storey or taller buildings as well, although accuracy outside the validated two-to-five storey range has not yet been empirically demonstrated.

The study went well beyond a single training run. Repeating the entire training procedure with five additional random seeds produced almost identical results with low standard deviations, demonstrating that the reported performance is not the product of a lucky initialization. A stratified five-fold cross-validation confirmed the stability of the model across different data partitions. An ablation study isolated the contribution of each component: removing the masking mechanism increased the average error from 1.56 to 9.97, removing the LSTM branches raised it to 9.68, and removing the CNN caused the largest degradation, to 14.79, confirming that all three elements are essential to the architecture’s synergy. The hybrid model was also benchmarked against five baselines, including a standard deep neural network, a one-dimensional CNN, a standalone LSTM, XGBoost and a Transformer encoder. It achieved the lowest error on nearly all twenty outputs, outperforming the best baseline by approximately nineteen percent, and differential Wilcoxon signed-rank tests confirmed that the improvements over the CNN, LSTM and Transformer baselines were statistically significant across every output.

The researchers are careful to frame the tool’s role. It is not intended to replace finite element analysis of an individual structure, and practitioners will still rely on professional nonlinear analysis software when designing retrofit interventions for a specific building. What the network offers instead is a fast, interpretable second opinion, capable of checking the order of magnitude of pushover results and, more importantly, enabling vulnerability assessments at a scale that traditional methods cannot reach. Because the outputs are the same bilinear capacity parameters used in code-based performance assessment, the predictions feed directly into regional risk frameworks without requiring qualitative guesswork.

The implications extend well beyond Italy. Most of the world’s seismically exposed cities contain large stocks of pre-modern-code reinforced concrete buildings, from Los Angeles to Istanbul to Kathmandu, and the ability to triage thousands of structures rapidly could transform how governments prioritize retrofit funding before the next major earthquake. The Salerno team suggests that extending the dataset to a broader range of structural configurations and validating out-of-distribution generalization are the next steps, but the study makes a persuasive case that hybrid LSTM-CNN architectures, guided by simple geometric masks, can compress an hour of nonlinear mechanics into a fraction of a second, bringing large-scale seismic resilience assessment within practical reach.

Subject of Research: A mask-guided hybrid LSTM-CNN neural network for rapid seismic assessment of existing reinforced concrete frame buildings

Article Title: A mask-guided multi-LSTM-CNN architecture for rapid seismic analysis of existing RC frames

Article References: D’Angelo, G., Nigro, F., & Martinelli, E. (2026). A mask-guided multi-LSTM-CNN architecture for rapid seismic analysis of existing RC frames. Neural Computing and Applications, 38(18), Article 741. https://doi.org/10.1007/s00521-026-12458-5

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12458-5

Keywords: seismic assessment, reinforced concrete frames, pushover analysis, LSTM networks, convolutional neural networks, deep learning, structural engineering, finite element method, earthquake engineering, neural network architecture, seismic retrofitting, machine learning