Strong motion records have shown that spectral accelerations commonly exhibit narrow peaks characterized by large amplitudes and small widths, which fail to adequately capture in current ground motion attenuation relationships （i.e., strong motion prediction equations）. Leveraging a comprehensive dataset of 190 215 horizontal strong motion records from Japan’s KiK-net arrays, this paper systematically parameterized these deviations, identified their governing factors, and developed an empirical model to enhance the accuracy of spectral acceleration predictions for seismic hazard analyses and structural engineering applications.

The dataset included earthquakes with M≥3 and peak ground acceleration PGA≥5 cm/s², encompassing various seismic scenarios featuring different magnitude, epicentral distances, and ground motion intensities. Acceleration response spectra （5% damping ratio） were computed for all records and normalized by PGA to isolate spectral shape characteristics. Residuals between observed spectra and predictions from a bilinear regression model （dependent on magnitude and epicentral distances） revealed distinct single-peaked deviations in 71.6% of records, with 24.9% showing multi-peaked deviations and only 3.5% lacking significant peaks. These deviations, characterized by their narrow bandwidth and elevated amplitudes, were parameterized using Gaussian curve fitting into three key features: peak height, peak width, and central period. The analyses demonstrated that such deviations occurred across diverse site conditions including bedrock stations, indicating contributions from both the spectral properties of bedrock motions and site-specific amplification mechanisms.

Statistical analyses were conducted to evaluate correlations between the three key features of single-peaked regular deviations and strong motion characteristics as well as site condition parameters. The results indicated that the peak heights of single-peaked regular deviations exhibited a strong positive correlation with the concentration of spectral energy within the deviation bandwidth, where higher energy ratios amplified peak amplitudes. Central period and peak width of single-peaked regular deviations showed significant positive relationships with the spectral peak period of borehole motions, earthquake magnitude, and epicentral distance. These relationships reflected the increasing dominance of long-period energy in large-magnitude distant-earthquakes. Site conditions further modulated deviations: softer soils （lower 30-meter average shear wave velocity v_S30）, thicker overburden （greater depth to bedrock with shear wave velocity v_S≥1 km/s）, and longer site predominant periods shifted deviations toward longer central periods and broader bandwidths. Weak-motion spectral ratios between surface and borehole motions also demonstrated moderate correlations with peak height, supporting the hypothesis that these deviations were derived from resonance between site layers and borehole motions.

To address multicollinearity among strong motion characteristics or site condition parameters, Lasso regression was employed to identify dominant factors. For peak height of single-peaked regular deviations, the key predictors included concentration of spectral energy within the deviation bandwidth, 30-meter average shear wave velocity v_S30, depth to bedrock with shear wave velocity v_S≥1 km/s, and weak-motion spectral ratios between surface and bedrock motions. For central period of single-peaked regular deviations, significant predictors were the spectral peak period of borehole motions, v_S30, depth to bedrock with v_S≥1 km/s, and longer site predominant periods. For peak width of single-peaked regular deviations, several parameters were retained. These include longer site predominant periods, spectral peak period of bedrock motions, v_S30, and depth to bedrock with v_S≥1 km/s.

A multivariate linear regression model was developed to predict the three deviation parameters. Comparison results showed the model achieved highly consistencies between predicted and observed parameters, and the prediction model considering single-peaked regular deviations of spectral accelerations significantly improved the accuracy of spectral shape predictions. The Jaccard similarity coefficients, which quantified the overlap between predicted and observed spectra, indicated that both the median and lower-bound of spectral similarity improved to a certain degree. The average Jaccard similarity coefficients increased from 0.728 for the model without considering regular spectral deviations to 0.792 for the model incorporating such deviations, highlighting the model’s ability to replicate narrow spectral peaks and improve then overall spectral fidelity.

The study underscored the necessity of integrating single-peaked regular deviations into seismic hazard analysis. While the methodology relied on statistical analysis of strong-motion records and did not delve into the physical mechanisms deriving these deviations, the empirical model demonstrated significant improvements in spectral shape predictions. The proposed framework, however, depended on the selected ground motion attenuation relationship; its applicability with alternative prediction equations requires further validation.

By bridging the gap between statistical averaging and site-specific spectral features, the findings had potential implications for strong motion prediction, seismic zonation, site-specific hazard assessments, and the design of vibration-sensitive structures. The proposed model enabled engineers to account for localized spectral amplifications that conventional attenuation models overlook, thereby enhancing the resilience of structures to earthquake-induced forces. Additionally, the methodology provided a framework for future studies to incorporate high-resolution site and ground motion data into probabilistic seismic hazard analyses, fostering a more nuanced understanding of ground motion variability across diverse geological settings.