The core problem the team set out to solve is one that plagues every data-driven surrogate model in engineering: error accumulation. Traditional physics-based constitutive models, which describe material and structural behavior through differential equations, are accurate but slow, often requiring substantial computational resources to simulate a single earthquake scenario. Data-driven surrogates, typically neural networks trained on the outputs of those expensive simulations, can produce predictions in milliseconds. The catch is that when a surrogate model is used for multi-step-ahead prediction, meaning it forecasts the structural state at the next time step and then feeds that forecast back in as input for the step after that, small inaccuracies compound. Over the duration of a long earthquake record, these errors snowball until the predicted response drifts far from reality, sometimes producing hysteretic loops that violate basic physical principles such as the conservation of energy.

The new framework attacks this problem from two directions simultaneously. The first is architectural: the backbone of the model is a hybrid Transformer-LSTM network. The Long Short-Term Memory component, a form of recurrent neural network, is designed to retain information over long sequences, making it well suited to tracking the evolving state of a structure as ground motion unfolds. The Transformer component, the architecture that underpins modern large language models, brings powerful attention mechanisms that can weigh the relevance of different moments in the shaking history when making a prediction. Together, the two components extract robust temporal features from the seismic input and the structural response history, allowing the network to directly predict the restoring force, the force the structure exerts in resistance to deformation, at each moment of the earthquake.

Architecture alone, however, cannot guarantee that a neural network behaves like a physical system. This is where the second innovation comes in: a consistency constraint based on hysteretic loop energy conservation, embedded directly into the loss function that guides training. During an earthquake, the area enclosed by a hysteretic loop represents the energy dissipated by the structure through yielding, friction, and other inelastic mechanisms. The researchers defined a time window and required that, within that window, the energy balance implied by the model’s predictions remain consistent with the physical energy dissipated in the corresponding hysteretic loop. Any prediction that would create or destroy energy out of nothing is penalized during training. The result is a network that is not merely fitting curves but is constrained to obey one of the most fundamental laws of physics, which in turn sharpens both its predictive accuracy and its physical plausibility.

The second line of attack addresses cumulative error propagation head-on through a cumulative error correction mechanism built on what the authors describe as a hybrid true-pseudo sample training strategy. In conventional training, a network learns to map true historical states to true future states. But at inference time, when the model is rolled forward through a long earthquake record, it must often consume its own predictions as inputs, states the training data never contained. This mismatch between training conditions and deployment conditions is a classic source of drift. The hybrid strategy exposes the network to both genuine samples and pseudo samples, states generated from the model’s own predictions during training, so that it learns to correct its own mistakes before they compound. The mechanism effectively teaches the surrogate to recognize and pull back toward physically meaningful trajectories even when it has wandered slightly off course.

Validation of the framework was comprehensive and systematic. The researchers first tested the model against two canonical mathematical descriptions of hysteretic behavior: the bilinear hysteretic model, a simplified representation of yielding systems, and the Bouc-Wen model, a smooth differential model widely used to capture the gradual stiffness degradation and pinching behavior of real structures. These benchmarks matter because they have known analytical behavior, allowing the team to quantify precisely how well the neural network reproduces the underlying physics. The framework was then implemented in both single-degree-of-freedom systems, which behave like a simple mass on a spring, and multiple-degree-of-freedom structural systems, which more realistically represent buildings with many interacting floors and components, all subjected to seismic ground motions.

Across these tests, the proposed framework demonstrated superior prediction accuracy compared with conventional data-driven approaches, along with enhanced extrapolation capability and robust generalization performance. Extrapolation capability is particularly significant in earthquake engineering, because the ground motions a real building will face in its lifetime may be stronger or of longer duration than anything in the training database. A surrogate that only interpolates within its training data is of limited use for safety assessment; one that can extend reliably beyond that data, while still respecting energy conservation, opens the door to rapid seismic performance evaluation across a much wider range of scenarios. The authors emphasize that these results highlight the framework’s potential as a physics-informed, data-driven approach for efficient seismic response prediction.

The broader context makes clear why this line of research has attracted intense interest. In recent years, deep learning has swept through structural engineering: LSTM networks have been used to predict the hysteresis of steel braces, convolutional and attention-based hybrids have been applied to long-period earthquake response, and physics-guided generative models have been proposed for damage prognosis in reinforced concrete bridge columns. Yet many of these models remain purely data-driven, with no explicit guarantee that their outputs obey conservation laws. The new study positions itself within a growing movement toward physics-informed machine learning, in which the equations and principles that engineers trust are woven into the fabric of the model rather than left for the data to rediscover. The authors also point to prior corrective training strategies for surrogate hysteretic models as an important precursor to their error correction mechanism, suggesting their contribution refines and extends an established idea with the addition of an explicit energy constraint.

The practical implications could extend well beyond the laboratory. Fast and reliable surrogates of nonlinear structural response are the engine behind many modern seismic risk tools, from fragility analysis of reinforced concrete frames to post-earthquake damage prediction for high-speed railway bridge systems, and from the design of seismic isolation devices to the safety assessment of trains running on bridges during shaking. If a surrogate can deliver accurate long-duration response predictions in a fraction of the time of a full nonlinear simulation, engineers could evaluate thousands of earthquake scenarios for a given structure, enabling probabilistic performance assessment at a scale that is currently impractical. The research was supported by the National Natural Science Foundation of China, the National Key Research and Development Program of China, the Japan Society for the Promotion of Science, and several provincial science foundations, reflecting the international and interdisciplinary nature of the effort.

The authors are careful to frame the current work as a foundation rather than a finished product. Their validation relied on idealized hysteretic models, bilinear and Bouc-Wen, which capture essential features of inelastic behavior but do not encompass the full complexity of real materials, including rebar buckling, strength deterioration under cyclic loading, or the asymmetric hysteresis of certain structural systems. The team explicitly identifies future investigations involving more complex hysteretic behaviors and engineering scenarios as the natural next step. Still, the central achievement stands: a neural network that predicts the violent, energy-dissipating dance of a structure during an earthquake, guided at every step by the conservation of energy and disciplined against its own drift. In a field where the difference between an accurate and a drifting prediction can shape decisions about life safety, that combination of speed, physics, and self-correction may prove to be exactly what earthquake engineering has been waiting for.

Subject of Research: Physics-guided deep learning for predicting hysteretic seismic response of structures with energy conservation constraints

Article Title: Physics-guided neural network for seismic hysteretic response prediction with energy conservation constraints and error correction

Article References: Shaohui, L., Lizhong, J., Wangbao, Z., Yuntai, Z., Lingxu, W., Yulin, F., & Tuo, Z. (2026). Physics-guided neural network for seismic hysteretic response prediction with energy conservation constraints and error correction. Bulletin of Earthquake Engineering. https://doi.org/10.1007/s10518-026-02657-9

Image Credits: AI Generated

DOI: 10.1007/s10518-026-02657-9

Keywords: physics-guided neural network, seismic response prediction, hysteretic behavior, energy conservation, Transformer-LSTM, cumulative error correction, Bouc-Wen model, earthquake engineering, surrogate modeling, structural dynamics, machine learning, Bulletin of Earthquake Engineering