The appeal of friction dampers lies in their physics. Conventional steel frames absorb earthquake energy by yielding, which means the structure itself accumulates damage with every loading cycle, a phenomenon engineers call cyclic deterioration or pinching. An RFD behaves differently. Governed by Coulomb’s law of dry friction, it produces stable, rectangular hysteretic loops whose enclosed area, and therefore the energy dissipated per cycle, remains constant no matter how many times the ground shakes. The device consists of a central rotating plate sandwiched between side plates with friction discs, all clamped by a pre-tensioned bolt. When inter-story drift forces the plates to rotate against one another, friction resists the motion with a constant moment, protecting the primary frame. Experimental work dating back to Mualla and Belev in 2002 showed reductions of up to 89 percent in lateral displacement and 41 percent in base shear under optimized configurations.

Simulating such devices accurately, however, is notoriously difficult. The abrupt transition between sticking and slipping is mathematically discontinuous, and the two common workarounds both introduce error. Regularization smooths the discontinuity with an artificial curve, adding fake compliance exactly where the response should be sharp, while equivalent-damping approaches replace the device with a viscous dashpot that cannot reproduce the rate-independent, rectangular hysteretic loop. Both require case-specific calibration, which would be hopeless at the scale of hundreds of thousands of simulations. The authors instead built on the Non-Smooth Dynamics method, reformulating Coulomb friction as a Linear Complementarity Problem solved at each timestep with a Projected Successive Over-Relaxation algorithm. This formulation reproduces the exact friction law with no tuning parameters, and because each step typically converges in fewer than five iterations, sweeping thousands of structural configurations becomes computationally feasible. Crucially, it also means the training labels fed to the machine learning models contain pure physics rather than artifacts of a smoothing choice.

The dataset behind the study is staggering in scope. The team generated 1,000 shear-type frames ranging from 3 to 20 stories, with randomized Latin Hypercube sampling covering fundamental periods from 0.3 to 3.0 seconds, floor masses from 150 to 350 tonnes, and four different mass and stiffness distributions. Each frame was fitted with roughly 50 different damper configurations, varying the number of devices, their vertical placement, and friction moments between 25 and 500 kilonewton-meters. Eight real seismic records, including El Centro 1940, Kobe 1995, and Northridge 1994, were applied at their original historical intensities. In total, 433,984 nonlinear simulations were run, each quantified with a composite index called CR2 that combines peak roof displacement, peak roof acceleration, and peak base shear, each normalized by the uncontrolled response, so that lower values mean better performance.

Six regression models were then trained to predict seismic performance from the design variables, and the comparison itself delivered a scientific finding. Linear models achieved a cross-validated coefficient of determination of only 0.64 for the typical scenario, while XGBoost reached 0.907 and Random Forest 0.904, cutting prediction error nearly in half. The gap proves that the relationship between damper design and seismic response is fundamentally nonlinear, with threshold effects and interactions that simple equations cannot capture. The nonlinearity deepens under severe ground motion: for the conservative 90th-percentile scenario, linear models managed only 0.31, while Random Forest reached 0.786. To prevent data leakage, the team used structure-grouped cross-validation, ensuring models were always evaluated on buildings they had never seen, and verified accuracy on 200 entirely held-out structures with calibrated 90 percent prediction intervals.

The most striking practical insight concerns coverage density, the fraction of floors equipped with dampers. Feature importance analysis identified this single variable as the dominant predictor of performance, and the response surface revealed clear diminishing returns. By formulating the threshold as an explicit optimization over the trained surrogate, the researchers found that the median coverage at which 80 percent of the maximum achievable gain is reached sits at about 0.83, or 0.86 when read from the empirical envelope across the whole population. Resolved by building height, the picture inverts common intuition: low-rise frames of 3 to 5 stories require full coverage, mid-rise frames of 6 to 12 stories need at least 86 percent, and high-rise frames of 13 to 20 stories need only about 73 percent, with none requiring full instrumentation. The mechanical explanation is elegant, since tall buildings concentrate drift demand in a subset of stories that can be covered selectively, whereas the flat drift profile of a low-rise frame leaves no inexpensive stories to omit.

The analysis also uncovered a counterintuitive failure mode. A small number of oversized dampers, high friction moments paired with low coverage, can leave a frame worse off than no dampers at all. In this lock-up regime, the devices never reach their slip threshold under the imposed demand, act as rigid links, and locally stiffen the structure, amplifying floor accelerations and base shear. These configurations account for just 1.4 percent of the dataset but nearly 10 percent of the total prediction error, and more than 80 percent of them occur in frames of five stories or fewer. The practical lesson is a lower bound on coverage for any given device capacity: coverage cannot be traded freely against friction moment. The study identifies an optimal friction moment window of roughly 300 to 460 kilonewton-meters for mid- and high-rise frames, with a lower optimum near 250 kilonewton-meters for short buildings.

Robustness across seismic scenarios proved to be another strength of the framework. When the researchers compared the optimal damper count recommended under typical conditions with that recommended under the conservative 90th-percentile scenario, 67.9 percent of the 1,000 structures received identical recommendations, and in the remaining cases the median discrepancy was just one damper. The conservative scenario never suggested fewer devices, only occasionally one or two more, meaning designers can optimize for typical demand and then verify against severe conditions with minimal adjustment. This multi-record approach, summarizing performance with 10th-percentile, median, and 90th-percentile statistics consistent with the FEMA P-58 framework, addresses a long-standing gap in earlier studies that optimized dampers against a single ground motion.

The authors are candid about the limits of their results. The structural models are elastic shear frames, and a substantial fraction of the simulated demand levels actually lie at or beyond the yield limit of real steel frames, so the coverage recommendation should be read as an upper bound for quasi-elastic serviceability design rather than a validated rule for ductile response. The design space is a factorial sweep rather than a representative building stock, and the eight-record suite carries sampling uncertainty of about 0.10 in CR2 units, which exceeds the surrogate’s own prediction error of 0.070 by roughly half. That last point may be the study’s most consequential methodological message: enlarging the ground motion database, not refining the machine learning model, is what would most sharpen future design guidance. Future extensions include soil-structure interaction, three-dimensional frames with torsional coupling, nonlinear material behavior, floor-specific friction moments, and cost-based optimization.

Even with those caveats, the work marks a genuine shift in how earthquake protection devices might be designed. Instead of running expensive simulations case by case, engineers can now query a trained surrogate that exposes the entire response surface across damper placement, capacity, and building geometry in milliseconds, complete with calibrated uncertainty bounds. The finding that taller buildings need proportionally fewer instrumented floors, that a coverage threshold near 86 percent captures most of the achievable benefit for mid-rise construction, and that a handful of oversized devices can actively harm a structure, together give practitioners an evidence-based starting point that previously required months of computation. As machine learning surrogates spread through earthquake engineering, this study offers both a template, exact physics in the labels, grouped validation, honest uncertainty, and a reminder that the biggest gains sometimes come from questioning design intuition.

Subject of Research: Machine learning surrogate modeling of nonlinear seismic response of steel moment frames with rotational friction dampers

Article Title: A machine learning-based surrogate model for analyzing the nonlinear seismic response of steel moment frames with rotational friction dampers

Article References: Ponce-López, E., & Yáñez-Godoy, H. (2026). A machine learning-based surrogate model for analyzing the nonlinear seismic response of steel moment frames with rotational friction dampers. Machine Learning with Applications, 26, Article 101030. https://doi.org/10.1016/j.mlwa.2026.101030

Image Credits: AI Generated

DOI: 10.1016/j.mlwa.2026.101030

Keywords: rotational friction dampers, machine learning, seismic response, steel moment frames, XGBoost, non-smooth dynamics, earthquake engineering, surrogate model, coverage density, structural optimization, energy dissipation, time-history analysis