Hybrid graph surrogate model predicts seismic response of steel building portfolios

When an earthquake strikes a city, the question that keeps engineers and emergency planners awake at night is not whether the ground will shake—it is which buildings will bend, which will crack, and which will collapse. Answering that question across an entire urban portfolio, potentially tens of thousands of structures at once, has long forced researchers into a difficult compromise. The most accurate tools, finite-element analyses that simulate the nonlinear behavior of every beam and column, are far too slow to run at city scale. The fastest tools, simplified empirical correlations, are fast precisely because they ignore too much of what makes each building unique. A new study published in the Bulletin of Earthquake Engineering proposes a way out of this trade-off, using graph neural networks to blend physics-based simulation with data-driven learning—and the results cut prediction errors by more than half compared with conventional approaches.

The work, carried out by Shandy Rianto and Xinzheng Lu of the Department of Civil Engineering at Tsinghua University in Beijing, addresses a stubborn problem in regional seismic damage assessment. Traditional regional models typically describe buildings using a handful of coarse attributes—height, construction era, occupancy class—and then map those attributes to damage states through simplified relationships. The approach is computationally cheap, but it flattens away the very details that govern how a structure actually behaves under strong shaking: the arrangement of its frames, the distribution of member stiffness, the topology of its lateral-force-resisting system.

At the other extreme sit multi-degree-of-freedom, or MDOF, models, which condense a building into a stack of lumped masses connected by nonlinear shear or flexural springs. These models can reproduce hysteretic behavior—strength and stiffness deterioration under cyclic loading—reasonably well, but they must be parameterized from limited structural data. When those parameters are assigned from generic statistical rules rather than from the building’s actual design, the simulated response can drift far from reality. Meanwhile, machine learning surrogates trained to imitate finite-element analysis have grown popular, but most of them ingest buildings as fixed-length feature vectors. A vector, however long, cannot natively capture the fact that a structure is a network of connected members whose behavior depends on adjacency, hierarchy, and load paths.

The Tsinghua team’s insight was to represent each building as a graph. In this representation, nodes correspond to structural components or stories, and edges encode the physical connections between them—beam-to-column joints, story-to-story coupling, lateral load transfer. Node features carry engineering attributes such as member dimensions, material properties, and story masses, while edge features describe connectivity and geometry. Graph neural networks, or GNNs, are uniquely suited to such data because their message-passing architecture propagates information along the actual edges of the structure. During each layer of the network, a node aggregates features from its neighbors, transforms them, and passes the result onward. After several rounds of message passing, each node’s embedding reflects not just its own properties but the structural context in which it sits—exactly the kind of relational reasoning that fixed vectors struggle to supply.

The framework the researchers developed is not a single model but a spectrum of data-physics integration strategies, offering three distinct surrogate modeling approaches. Two of them are hybrid schemes that retain an MDOF model as the physics backbone and use graph-based learning to correct or enrich its predictions. In effect, the MDOF simulation provides a physically consistent first estimate of the seismic response, and the GNN learns the systematic discrepancies between that estimate and the true nonlinear finite-element response, using the graph-structured description of each building to explain why the simplified model falls short for a particular structure. This division of labor matters: the physics component constrains the prediction to plausible dynamic behavior, while the data component supplies building-specific nuance that generic parameterization cannot capture.

The third approach is more ambitious—a physics-aware, end-to-end GNN that estimates seismic responses directly from the graph representation, embedding structural mechanics knowledge into the network’s architecture and training rather than relying on an intermediate simulation step. This end-to-end route eliminates the need to run even a simplified MDOF simulation at prediction time, promising the greatest speed, while the embedded physics keeps the network from learning spurious correlations that would fail under earthquakes unlike those in its training set.

To train and validate these models, the researchers drew on established datasets of seismic designs, nonlinear models, and response simulations for steel moment-resisting frame buildings—a class of structure whose ductile steel connections and frame-based lateral systems make them a natural testbed for graph methods. Steel moment frames were modeled with nonlinear hysteretic behavior consistent with the deterioration phenomena documented in the earthquake engineering literature, and the ground-motion demand was characterized through intensity measures that capture spectral shape, not just peak acceleration. The graph-based structural representation dataset compiled for the study has been deposited in the Science Data Bank, making both the dataset and the GNN algorithms available to other researchers—a transparency move that matters in a field where reproducibility of machine learning benchmarks has often lagged behind headline accuracy claims.

The benchmark comparisons are where the framework makes its case. Against traditional parameterization-based MDOF modeling—the standard practice when only coarse building data are available—the proposed approaches reduced response estimation errors by more than 50 percent. They also outperformed commonly used data-driven surrogates that rely on fixed-length vector features, confirming the central hypothesis: the accuracy bottleneck in regional machine learning for earthquakes is not merely the amount of training data but the adequacy of the structural representation itself. When the model can see the building as a connected system rather than as a list of attributes, it learns correspondences that map far more faithfully onto actual structural behavior.

The timing of this work is no accident. Cities worldwide are generating increasingly rich digital descriptions of their building stock—digital twins, semantic city information models, and detailed urban survey databases that record geometry, materials, and structural systems at a fidelity unimaginable when the first regional loss estimation tools were assembled in the 1980s. The missing ingredient has been a computational method that can exploit this detail at scale. Full finite-element analysis of a million-building portfolio remains impractical even with modern GPU clusters, but a trained GNN performs inference in a fraction of a second per building, once trained. The hybrid framework thus occupies a sweet spot: fidelity approaching refined structural simulation, speed approaching that of the crudest empirical tools.

The practical stakes extend well beyond academic benchmarks. Rapid post-earthquake damage assessment determines where rescue teams are sent first, which hospitals are presumed operational, and how emergency budgets are allocated in the hours and days after a major event. Pre-earthquake risk mapping, in turn, drives retrofit prioritization and insurance pricing for entire metropolitan building stocks. In both scenarios, systematic over- or under-estimation of damage has real consequences: misallocated resources, mispriced risk, and misplaced confidence in buildings that may not perform as the regional model assumes. A surrogate that halves the estimation error without slowing the analysis down changes the reliability of every downstream decision.

There are, of course, caveats inherent to the approach. Surrogate models inherit the distributional boundaries of their training data; a GNN trained on code-compliant steel moment frames cannot be trusted, without retraining, to predict the response of unreinforced masonry or irregular older buildings that behave in fundamentally different ways. The physics-aware design mitigates but does not eliminate extrapolation risk, which is precisely why the authors emphasize the hybrid tiered structure—users can choose the level of data-physics integration appropriate to their data availability and confidence requirements. The authors also note the framework’s applicability hinges on the quality of the urban-scale databases feeding it, though the trend line for such data availability is clearly pointing upward.

What the study ultimately demonstrates is a conceptual shift in how the field thinks about structural representation for machine learning. For a decade, progress in seismic surrogates came largely from better regression architectures stacked on top of the same hand-engineered feature vectors. This work argues that the representation itself—whether the model perceives a building as a graph of connected physical components—is a first-order determinant of accuracy. Given the field’s trajectory, graph-based hybrid surrogates are likely to spread quickly beyond steel moment frames to shear-wall buildings, reinforced concrete frames, and increasingly to million-scale urban clusters, where related AI frameworks are already being developed for city-scale simulation.

The research was supported by the National Natural Science Foundation of China. As urban populations in seismically active regions continue to grow, tools that can predict, quickly and accurately, how an entire city’s buildings will respond to shaking are moving from academic aspiration to operational necessity. This study offers a concrete, open, and demonstrably accurate step in that direction—teaching machines to see buildings the way structural engineers do: as systems, not statistics.