The core insight is that every building acts as a mechanical filter sitting between the earth and its occupants. When seismic waves arrive at a foundation, the structure amplifies, attenuates, and delays them according to its own dynamic properties, so the acceleration recorded on an upper floor is not the ground motion itself but a transformed version of it. For decades, engineers have treated this transformation as a nuisance to be corrected for. The new work inverts that logic: if the filtering behavior of a building can be modeled accurately enough, then the building’s own response becomes a rich source of information about the shaking it experienced at its base. In effect, the city’s building stock becomes a distributed, already-installed seismic sensing array.

Technically, the challenge is an inverse problem: given the acceleration time histories measured by structural response sensors, recover the peak ground acceleration, peak ground velocity, and peak ground displacement at the building’s location. The researchers begin by processing the raw structural acceleration records to extract multiple response parameters, which are then fused together to estimate the three ground-motion intensity measures. Fusing several parameters rather than relying on a single feature gives the learning algorithm a more robust signature of the underlying ground motion, because different parameters carry complementary information about amplitude, frequency content, and duration of the shaking.

To make the framework work at the scale of an entire city, the study area is discretized into grid cells, and within each cell a representative building is modeled as a nonlinear multi-degree-of-freedom shear structure. This is a standard idealization in earthquake engineering in which each story of the building is represented by a lumped mass connected to its neighbors by nonlinear springs, capturing how the structure yields and deforms under strong shaking. By simulating earthquakes acting on these representative buildings, the team generates the training data needed to teach a model how structural responses map back to ground motions for each building type, without waiting decades to accumulate real paired observations of ground truth and building response.

The machine learning architecture at the heart of the method, called ECA-Deep Net, is a hybrid framework that combines convolutional neural networks with long short-term memory networks, and it is trained separately for each representative building. The convolutional layers excel at extracting local features from the sensor waveforms, while the LSTM layers capture the temporal dependencies that matter in a seismic record, where the ordering and persistence of pulses carry physical meaning about the wavefield. A distinctive ingredient is the efficient channel attention mechanism, borrowed from the computer vision architecture ECA-Net. Channel attention allows the network to learn which feature channels are most informative for the task and to weight them accordingly, using a lightweight mechanism that adds almost no parameter overhead. In this application, attention helps the model decide which aspects of the multi-sensor structural response deserve emphasis when inferring ground-motion intensity.

The city-scale validation used a simulated earthquake, comparing the ground-motion maps generated by the framework against reference ground-motion fields computed by conventional means. The results showed that the predicted maps effectively reproduce the spatial variability of the reference fields, which is a critical property: ground shaking in real earthquakes varies sharply over short distances because of soil conditions, basin geometry, and source directivity, and a mapping method that smooths away that variability would mislead emergency responders about which neighborhoods were hit hardest. The framework also proved stable under varying levels of sensor noise at city scale, an essential robustness test for a method intended to run on low-cost hardware in real buildings rather than laboratory-grade instruments.

Perhaps the most compelling evidence comes from the field. The team validated the approach using recorded responses from an instrumented six-story real building, testing the models on earthquake records the network had never seen before, without any architectural modification or hyper-parameter retuning. The models achieved correlation coefficients of up to 0.97 between predicted and actual ground-motion parameters, with error values ranging from 7.33 percent to 12.05 percent. That level of transfer performance, on unseen events and with no recalibration, suggests the learned mapping is capturing genuine physics of the building-ground interaction rather than overfitting to the training set.

The practical implications reach well beyond seismology. Rapid ground-motion maps feed directly into shake maps, damage estimation, and emergency routing decisions, and the cost barrier of conventional strong-motion networks is one of the main reasons many earthquake-prone cities in developing regions lack adequate coverage. Structural health monitoring sensors are increasingly being installed in buildings anyway, for condition assessment and safety management, so a method that doubles them as ground-motion sensors offers a scalable, data-driven, and rapidly updatable alternative. Because each model is tied to a specific representative building, the network can grow incrementally: as more instrumented buildings come online, more grid cells gain coverage, and the city map fills in organically.

The work also fits into a broader movement in earthquake science toward machine learning-based ground motion modeling, from neural network prediction of peak ground acceleration to deep learning approaches using single-station waveforms and volunteer-hosted MEMS accelerometer networks. What distinguishes this study is the explicit use of the building as the sensing element, with the inversion handled by a building-specific model rather than a one-size-fits-all network. The authors acknowledge that the approach relies on representative structural models and simulated training scenarios, and the datasets and code files will be made available on request, which should allow other research groups to test the framework on their own instrumented buildings and urban grids.

If the approach matures from simulation and single-building validation into operational deployment, the aftermath of the next major urban earthquake could look very different. Within minutes of the shaking stopping, emergency managers could have a continuously updated, block-by-block picture of ground-motion intensity assembled from the very structures the earthquake was trying to destroy, informing search-and-rescue priorities, inspection queues, and utility shutdown decisions. The buildings that shelter a city, the study suggests, can also become the instruments that watch over it, provided the deep learning models translating their swaying into seismology are built with the care that this new framework demonstrates.

Subject of Research: City-scale seismic ground motion mapping using building-embedded structural response sensors and hybrid deep learning

Article Title: City-scale ground motion mapping from building-embedded structural response sensors using building-specific hybrid deep learning models

Article References: Zar, A., Li, S., Li, C., Li, C., & Zhao, J. (2026). City-scale ground motion mapping from building-embedded structural response sensors using building-specific hybrid deep learning models. Bulletin of Earthquake Engineering. https://doi.org/10.1007/s10518-026-02659-7

Image Credits: AI Generated

DOI: 10.1007/s10518-026-02659-7

Keywords: ground motion mapping, structural response sensors, deep learning, earthquake engineering, channel attention, LSTM networks, convolutional neural networks, peak ground acceleration, seismic monitoring, structural health monitoring, shear building model, sensor data fusion