@article{BindiKothaWeatherilletal.2018,
  author    = {Bindi, Dino and Kotha, Sreeram Reddy and Weatherill, Graeme and Lanzano, Giovanni and Luzi, Lucia and Cotton, Fabrice},
  title     = {The pan-European engineering strong motion (ESM) flatfile},
  series = {Bulletin of earthquake engineering : official publication of the European Association for Earthquake Engineering},
  volume    = {17},
  journal   = {Bulletin of earthquake engineering : official publication of the European Association for Earthquake Engineering},
  number    = {2},
  publisher = {Springer},
  address   = {Dordrecht},
  issn      = {1570-761X},
  doi       = {10.1007/s10518-018-0466-x},
  pages     = {583 -- 602},
  year      = {2018},
  abstract  = {We present the results of a consistency check performed over the flatfile extracted from the engineering strong motion (ESM) database. The flatfile includes 23,014 recordings from 2179 earthquakes in the magnitude range from 3.5 to 7.8 that occurred since the 1970s in Europe and Middle East, as presented in the companion article by Lanzano et al. (Bull Earthq Eng, 2018a). The consistency check is developed by analyzing different residual distributions obtained from ad-hoc ground motion prediction equations for the absolute spectral acceleration (SA), displacement and Fourier amplitude spectra (FAS). Only recordings from earthquakes shallower than 40 km are considered in the analysis. The between-event, between-station and event-and-station corrected residuals are computed by applying a mixed-effect regression. We identified those earthquakes, stations, and recordings showing the largest deviations from the GMPE median predictions, and also evaluated the statistical uncertainty on the median model to get insights on the applicable magnitude-distance ranges and the usable period (or frequency) range. We observed that robust median predictions are obtained up to 8 s for SA and up to 20 Hz for FAS, although median predictions for Mw ≥ 7 show significantly larger uncertainties with 'bumps' starting above 5 s for SA and below 0.3 Hz for FAS. The between-station variance dominates over the other residual variances, and the dependence of the between-station residuals on logarithm of Vs30 is well-described by a piece-wise linear function with period-dependent slopes and hinge velocity around 580 m/s. Finally, we compared the between-event residuals obtained by considering two different sources of moment magnitude. The results show that, at long periods, the between-event terms from the two regressions have a weak correlation and the overall between-event variability is dissimilar, highlighting the importance of magnitude source in the regression results.},
  language  = {en}
}
@phdthesis{Kotha2018,
  author    = {Kotha, Sreeram Reddy},
  title     = {Quantification of uncertainties in seismic ground-motion prediction},
  url       = {http://nbn-resolving.de/urn:nbn:de:kobv:517-opus4-415743},
  school      = {Universit{\"a}t Potsdam},
  pages     = {xii, 101},
  year      = {2018},
  abstract  = {The purpose of Probabilistic Seismic Hazard Assessment (PSHA) at a construction site is to provide the engineers with a probabilistic estimate of ground-motion level that could be equaled or exceeded at least once in the structure's design lifetime. A certainty on the predicted ground-motion allows the engineers to confidently optimize structural design and mitigate the risk of extensive damage, or in worst case, a collapse. It is therefore in interest of engineering, insurance, disaster mitigation, and security of society at large, to reduce uncertainties in prediction of design ground-motion levels. In this study, I am concerned with quantifying and reducing the prediction uncertainty of regression-based Ground-Motion Prediction Equations (GMPEs). Essentially, GMPEs are regressed best-fit formulae relating event, path, and site parameters (predictor variables) to observed ground-motion values at the site (prediction variable). GMPEs are characterized by a parametric median (μ) and a non-parametric variance (σ) of prediction. μ captures the known ground-motion physics i.e., scaling with earthquake rupture properties (event), attenuation with distance from source (region/path), and amplification due to local soil conditions (site); while σ quantifies the natural variability of data that eludes μ. In a broad sense, the GMPE prediction uncertainty is cumulative of 1) uncertainty on estimated regression coefficients (uncertainty on μ,σ_μ), and 2) the inherent natural randomness of data (σ). The extent of μ parametrization, the quantity, and quality of ground-motion data used in a regression, govern the size of its prediction uncertainty: σ_μ and σ. In the first step, I present the impact of μ parametrization on the size of σ_μ and σ. Over-parametrization appears to increase the σ_μ, because of the large number of regression coefficients (in μ) to be estimated with insufficient data. Under-parametrization mitigates σ_μ, but the reduced explanatory strength of μ is reflected in inflated σ. For an optimally parametrized GMPE, a ~10\% reduction in σ is attained by discarding the low-quality data from pan-European events with incorrect parametric values (of predictor variables). In case of regions with scarce ground-motion recordings, without under-parametrization, the only way to mitigate σ_μ is to substitute long-term earthquake data at a location with short-term samples of data across several locations - the Ergodic Assumption. However, the price of ergodic assumption is an increased σ, due to the region-to-region and site-to-site differences in ground-motion physics. σ of an ergodic GMPE developed from generic ergodic dataset is much larger than that of non-ergodic GMPEs developed from region- and site-specific non-ergodic subsets - which were too sparse to produce their specific GMPEs. Fortunately, with the dramatic increase in recorded ground-motion data at several sites across Europe and Middle-East, I could quantify the region- and site-specific differences in ground-motion scaling and upgrade the GMPEs with 1) substantially more accurate region- and site-specific μ for sites in Italy and Turkey, and 2) significantly smaller prediction variance σ. The benefit of such enhancements to GMPEs is quite evident in my comparison of PSHA estimates from ergodic versus region- and site-specific GMPEs; where the differences in predicted design ground-motion levels, at several sites in Europe and Middle-Eastern regions, are as large as ~50\%. Resolving the ergodic assumption with mixed-effects regressions is feasible when the quantified region- and site-specific effects are physically meaningful, and the non-ergodic subsets (regions and sites) are defined a priori through expert knowledge. In absence of expert definitions, I demonstrate the potential of machine learning techniques in identifying efficient clusters of site-specific non-ergodic subsets, based on latent similarities in their ground-motion data. Clustered site-specific GMPEs bridge the gap between site-specific and fully ergodic GMPEs, with their partially non-ergodic μ and, σ ~15\% smaller than the ergodic variance. The methodological refinements to GMPE development produced in this study are applicable to new ground-motion datasets, to further enhance certainty of ground-motion prediction and thereby, seismic hazard assessment. Advanced statistical tools show great potential in improving the predictive capabilities of GMPEs, but the fundamental requirement remains: large quantity of high-quality ground-motion data from several sites for an extended time-period.},
  language  = {en}
}
@article{KothaBazzurroPagani2018,
  author    = {Kotha, Sreeram Reddy and Bazzurro, Paolo and Pagani, Marco},
  title     = {Effects of epistemic uncertainty in seismic hazard estimates on building portfolio losses},
  series = {Earthquake spectra : the professional journal of the Earthquake Engineering Research Institute},
  volume    = {34},
  journal   = {Earthquake spectra : the professional journal of the Earthquake Engineering Research Institute},
  number    = {1},
  publisher = {Earthquake Engineering Research Institute},
  address   = {Oakland},
  issn      = {8755-2930},
  doi       = {10.1193/020515EQS020M},
  pages     = {217 -- 236},
  year      = {2018},
  abstract  = {In catastrophe risk modeling, a defensible estimation of impact severity and its likelihood of occurrence to a portfolio of assets can only be made through a rigorous treatment of uncertainty and the consideration of multiple alternative models. This approach, however, requires repeating lengthy calculations multiple times. To limit the demand on computational time and resources, a frequent practice in the industry is to estimate the distribution of earthquake-induced portfolio losses using a simulated catalog of events from a single representative mean ground motion hazard model for the region. This simplified approach is faster but may provide biased estimates of the likelihood of occurrence of the large and infrequent losses that drive many risk mitigation decisions. Investigation through case studies of different portfolios of assets located in the San Francisco Bay Region shows the potential for both a bias in the mean loss estimates and an underestimation of their central 70\% interpercentile. We propose a simplified and computationally practical approach that reduces the bias in the mean portfolio loss estimates. This approach does not improve the estimate of the interpercentile range, however, a quantity of no direct practical use.},
  language  = {en}
}
@article{KothaCottonBindi2018,
  author    = {Kotha, Sreeram Reddy and Cotton, Fabrice and Bindi, Dino},
  title     = {A new approach to site classification},
  series = {Soil Dynamics and Earthquake Engineering},
  volume    = {110},
  journal   = {Soil Dynamics and Earthquake Engineering},
  publisher = {Elsevier},
  address   = {Oxford},
  issn      = {0267-7261},
  doi       = {10.1016/j.soildyn.2018.01.051},
  pages     = {318 -- 329},
  year      = {2018},
  abstract  = {With increasing amount of strong motion data, Ground Motion Prediction Equation (GMPE) developers are able to quantify empirical site amplification functions (delta S2S(s)) from GMPE residuals, for use in site-specific Probabilistic Seismic Hazard Assessment. In this study, we first derive a GMPE for 5\% damped Pseudo Spectral Acceleration (g) of Active Shallow Crustal earthquakes in Japan with 3.4 <= M-w <= 7.3 and 0 <= R-JB <= 600km. Using k-mean spectral clustering technique, we then classify our estimated delta S2S(s)(T = 0.01 - 2s) of 588 wellcharacterized sites, into 8 site clusters with distinct mean site amplification functions, and within-cluster site-tosite variability similar to 50\% smaller than the overall dataset variability (phi(S2S)). Following an evaluation of existing schemes, we propose a revised data-driven site classification characterized by kernel density distributions of V-s30, V-s10, H-800, and predominant period (T-G) of the site clusters.},
  language  = {en}
}