Peak ground velocity （PGV） has emerged as a critical ground-motion parameter for seismic design and risk assessment, as it better represents cumulative energy and intermediate-frequency content of strong ground motion compared with peak ground acceleration （PGA）. Despite its importance, current Chinese seismic codes do not provide systematic PGV adjustment coefficients for different site classes, which are essential for reliable site-specific seismic hazard analysis. This study addresses this gap by developing a set of PGV site-amplification coefficients based on an unprecedentedly large numerical simulation campaign, integrating four distinct constitutive models, a nationwide borehole database, and a carefully selected set of bedrock strong-motion records.

The research utilizes 19263 engineering site boreholes collected from major seismic safety evaluation projects across China. Each borehole provides detailed soil stratification, measured shear-wave velocity profiles （using down-hole or cross-hole methods）, physical and mechanical properties （density, water content, void ratio）, and dynamic parameters （shear modulus reduction and damping ratio curves） calibrated by regional empirical relationships. All soil layers are classified into six engineering geological categories: soft silt/mud, fill, silty clay, sandy soil, gravelly soil, and bedrock. The dynamic nonlinear parameters of each soil type are taken from well-established Chinese references and are further validated against measured data.

The site response analyses are performed using a one-dimensional wave-propagation code （SiteDyn1D, part of the SEC R2019 platform） developed by the authors and rigorously validated against both Chinese strong-motion records and KiK-net downhole array data. To thoroughly account for methodological uncertainty, four constitutive models are employed: ① frequency-independent equivalent linear, ② frequency-dependent equivalent linear （Yoshida model）, ③ time-domain extended Masing criterion, and ④ time-domain dynamic skeleton curve. The frequency-dependent Yoshida model uses a strain-dependent, frequency-dependent equivalent strain function （with cut-off frequency 15 Hz and exponent 2.0） to better capture broadband wave propagation. The two time-domain methods integrate the equations of motion directly; the extended Masing rule is suitable for regular cyclic loading, while the dynamic skeleton curve allows progressive stiffness degradation and strength loss under strong, irregular loading, making it more appropriate for large-strain conditions in soft soils.

The input ground motions are 110 bedrock records selected from the NGA （Next Generation Attenuation） database. The selection criteria are: ① v_S30≥500 m/s, which corresponds to Chinese site class I₁ （reference bedrock）, and ② a wide range of earthquake magnitudes （M5−M8）, epicentral distances （4−18 km, mostly near-fault but with sufficient intensity and spectral diversity）, PGA （0−1500  cm/s²） and PGV （0−150 cm/s）. Each record’s significant duration （D_5-95, based on Arias intensity） is also computed. These records ensure that the input ensemble covers a broad spectrum of intensities and frequency contents, which is essential for deriving statistically robust site coefficients.

Combining the 19263 boreholes, 110 input motions, and four constitutive models yields a total of 8475720 one-dimensional nonlinear site response simulations — the largest domestic database of its kind for PGV site effects. For each simulation, the PGV amplification coefficient is defined as the ratio of PGV at the ground surface to PGV at the reference bedrock interface （v_S＝500 m/s, class I₁）. All calculations assume vertically incident shear waves and horizontally layered soil profiles, which is physically justified given the low horizontal variability relative to vertical variability.

The results are analyzed in terms of site class （ Ⅱ , Ⅲ , and Ⅳ according to the Chinese seismic code）, input PGA bins （ten intensity intervals from ≤0.05g to 1.0g, with 10−14 records per bin）, and the four constitutive models. For each site class and PGA bin, the mean and median PGV amplification coefficients are computed; because the distributions are approximately normal, the mean is chosen as the central tendency measure. A key finding is that site conditions strongly affect the amplification: as the site becomes softer （from class Ⅱ to Ⅳ）, the mean amplification coefficient increases （class Ⅱ : 1.1−2.0; class Ⅲ : 0.8−2.1; class Ⅳ: 0.6−2.5） and the dispersion also grows. Moreover, with increasing input intensity, the amplification coefficients decrease nonlinearly, reflecting the progressive nonlinearity of soils. This attenuation is mild for stiff sites but very pronounced for soft sites （class Ⅳ）, confirming that soft soils exhibit strong nonlinear behavior even at moderate to high input motions.

Comparing the four constitutive models reveals significant methodological differences at high input intensities. At low intensities, all models give similar results. At high intensities （PGA＞0.2g）, the time-domain nonlinear methods yield consistently lower amplification than the frequency-domain equivalent-linear methods, indicating that the latter may overestimate the response of soft soils under strong shaking. For class Ⅱ and Ⅲ sites, the two equivalent-linear methods agree well among themselves, as do the two time-domain methods, but the differences between the two families become pronounced above 1000  cm/s². For class Ⅳ soft sites, the two frequency-domain methods diverge significantly over the entire intensity range, while the time-domain methods converge at the highest intensities. This behavior demonstrates that the choice of constitutive model introduces substantial uncertainty for soft soils under strong earthquakes, and using multiple methods is strongly recommended for engineering applications.

For engineering application, the results are grouped by standard PGA bins （i.e., the bins adopted in the Chinese Seismic Ground Motion Parameters Zonation Map）. Compared with quadratic polynomial fitting and LOWESS smoothing, the Hanning-window method achieves the best balance between fidelity （RMSE 0.047−0.076, R²＝0.607−0.908 across site classes） and curve smoothness, without introducing artificial oscillations. Accordingly, the final recommended PGV adjustment coefficients are derived from the Hanning-window smoothed results.

The paper also reports the standard deviation of the four-model means for each bin, which quantifies the inter-model uncertainty. For class Ⅱ , the standard deviation is small （0.015−0.13）, indicating good consistency among the models. For class Ⅲ , it increases with input intensity （from 0.006−0.03 to 0.10−0.22, max 0.220）. For class Ⅳ , the increase is dramatic （from 0.013−0.08 to 0.19−0.308, max 0.308）, fully supporting the conclusion that soft soils under strong earthquakes are associated with high methodological uncertainty.

The proposed PGV site coefficients are presented as a function of input PGA （≤0.05g, 0.1g, 0.15g, 0.2g, 0.3g, 0.4g, 0.6g, 0.8g, and 1.0g） and site classes Ⅱ , Ⅲ , and Ⅳ , all referenced to class I₁ （v_S＝500 m/s）. For example, for class Ⅳ at PGA＝0.1g, the coefficient is 1.714; at PGA＝1.0g, it decreases to 1.135.

Finally, the study acknowledges several limitations that point to future research directions: ① the simulations are one-dimensional, ignoring three-dimensional site heterogeneity, bedrock topography, and non-vertical wave incidence; ② the constitutive models are total-stress based and do not account for pore-pressure buildup and liquefaction, which may be critical for soft sites in high-seismicity regions; ③ the input motions are taken exclusively from the NGA database, and although they cover a wide parameter space, local Chinese strong-motion records should be incorporated in future work to validate and refine the coefficients; ④ direct validation against instrumental recordings of PGV site amplification has not been performed in this study — a systematic comparison with strong-motion arrays （e.g., KiK-net, Chinese national strong-motion network） is a high-priority next step. Despite these limitations, the proposed PGV site coefficients represent the first comprehensive model derived from massive Chinese borehole data and multiple constitutive models, and they provide a solid foundation for incorporating PGV-based site effects into the next generation of China’s seismic zonation map.