Fueled by the advancement in sensor technology and increased seismicity, repeated observations of earthquakes are accumulated at an unprecedented number of localities around the globe. Thus, the systematic effects about source, propagation path and near-surface geology can be disentangled from ground motions recorded by a dense network of sensors. 

Insights into these constituent components can bring about a better understanding and prediction of the ground-motion phenomena, thus improving our forecast of the social and economic impacts of future catastrophic events.

With the accumulation of data from longstanding and new sources, the rapid growth in computational resources and data storage capacity and the recent development of advanced algorithms, artificial intelligence (AI, including machine learning and deep learning) is already broadly applied in many disciplines, e.g., computer vision. AI has been proven to be a formidable tool in addressing many traditionally changeling tasks, for instance, in automation, modelling and scientific discovery.

The modelling of the site effects is an active subfield of earthquake engineering and is an essential element of seismic risk analyses. However, site effects are, in most cases, too complex to be accurately described by a set of differential equations. Thus, AI is gaining traction in the earthquake engineering community. Though promising results have been achieved, AI is certainly not a panacea and has its limitations. In some cases, we may have a small sample size, which poses challenges to developing robust data-driven models. Sometimes we have a sufficient number of instances that have a skewed distribution. In some other applications, we need to deal with the paucity of gold-standard ground truth (for supervised learning), or noisy, incomplete, inhomogeneous uncertain data. Also, the multivariate, nonlinear and non-stationary nature of site effects further complicates the issue. In addition, the “black box” nature of some AI algorithms is among the obstacles prohibiting a more widespread adoption of AI in the community. 

To overcome these limitations and harness the opportunities brought by AI, we devote efforts to the development of large and open benchmark datasets (OpenSite and OpenAmp), and then use novel approaches to address the data challenges, physical consistency, interpretability, generality and uncertainty of AI models. 

The intensity of earthquake-induced ground motions depends on complex interactions between the earthquake source, the path between the source and the site, and the surface and subsurface conditions surrounding the observation site. The impacts of source, path, and site on ground motions are often referred to as “source effects”, “path effects”, and “site effects”, respectively. Site effects at specific sites are highly repeatable during different earthquakes and thus are more predictable than the other two factors. In forward modelings of seismic hazard/risk at a specific location (e.g., new power plants or waste disposal sites), site effects are the only one among the above three factors “within our reach”.

However, current approaches in mapping site data to site amplification can be improved, meanwhile, more accurate new approaches are also being sought. I proposed a novel approach to estimate site effects, c-HVSR, an empirical correction to the horizontal-to-vertical spectral ratio (HVSR) of earthquake recordings. c-HVSR can significantly outperform the current approach, 1D ground response analyses. I, with my co-workers, also systematically assessed and compared the performance of many existing amplification estimation techniques and quantified their epistemic uncertainties based on comparisons to observations. Besides epistemic uncertainty, I investigated the aleatory variability of site response using HVSR. 

An earthquake recording site was simply delineated as either rock or soil sites in some early ground-motion models (GMMs) before being explicitly characterized according to a piecewise site classification scheme or, more recently, a continuous site proxy. Undoubtedly, the time-averaged shear-wave velocity in the first 30 m, VS30 is the most widely used site or site class delineator. However, many studies have shown that VS30 alone is not adequate to distinguish the site effects at one site from those at another. 

We investigated the best-performing site characterization proxy alternative and complementary to VS30, as well as the optimal combination of proxies in characterizing site response. Then we establish OepnSite: an open-source site database for a total number of 1742 earthquake recording sites in the K-NET (Kyoshin network) and KiK-net (Kiban Kyoshin network) networks in Japan. To our knowledge, OpenSite is the most detailed site database in the world. This large and homogeneous site database can benefit the application of sophisticated machine learning techniques in site-effects studies. 

For a long time now, researchers have recognized basin effects as the changes in strong ground motions induced by the lateral finiteness of surficial and sub-surficial soil layers. Basin effects have received much attention for two reasons: firstly, basin effects may amplify ground motion by a large factor; secondly, many urban areas subjected to large seismic risk are situated atop sedimentary basins (e.g., Los Angeles, San Fransisco, Tokyo, Mexico City, Beijing, etc).

Although basin effects have long been recognized, current engineering practice to account for site effects is mainly based on the 1D (one-dimensional) approaches. To contribute to integrating the basin effects into the site-specific hybrid (probabilistic+deterministic) seismic hazard analysis, I developed a practical and robust approach to consider the aggravation of ground motions due to basin structures, which could help mitigate seismic risks of safety-critical buildings and infrastructure such as tall buildings, long-span bridges, and nuclear power plants.