Woodframe construction is commonly used for single and multifamily residential buildings in the United States. In many parts of California, multifamily woodframe residential buildings are constructed with open first stories, which have much less strength and stiffness compared to the ones above. In older single-family residences, the “crawl space” is constructed with unbraced and unbolted cripple walls. Both these conditions lead to a soft-story response during seismic loading, resulting significant damage, economic losses and even collapse. This type of vulnerability is often addressed through seismic retrofits, which can be mandated by local jurisdictions (e.g., the Los Angeles Soft-Story Ordinance) or incentivized by state or local entities (e.g., the California Earthquake Authority Brace and Bolt Program). A key challenge in implementing these retrofit programs (mandated or incentivized) is quantifying the improvements in performance at the individual and portfolio scale and creating design procedures that maximize the overall benefit. This research integrates nonlinear structural modeling, performance-based assessments and advanced statistical and machine learning techniques to quantify the benefit of soft-story woodframe building retrofit and develop optimal design solutions that maximize regional performance. The considered construction types include single-family houses with unbraced cripple walls developed as part of the recently completed Pacific Earthquake Engineering Research Institute (PEER) and California Earthquake Authority (CEA) project and multi-family residences with soft, weak and open front wall lines (SWOF). An end-to-end computational platform is developed to automate the construction and analysis of archetype numerical models in OpenSees and conduct seismic evaluations based on the PEER performance-based earthquake engineering framework. The performance of existing and retrofitted buildings is assessed in terms of collapse safety and direct (due to earthquake damage) economic losses.
The effect of retrofit and various structural characteristics is illuminated for the single-family cripple wall houses. 2^k full factorial experiment design combined with hypothesis testing is used to identify the most influential structural properties. Two story buildings performed worse than their one-story counterparts and pre-1945 buildings performed better than pre-1955 construction. Building performance is found to be positively correlated with cripple wall heights and cripple wall retrofits provided significant overall improvements. Surrogate models are developed as a compact statistical link between key structural characteristics and seismic performance. Several machine learning algorithms are investigated for predicting the building median collapse intensity and expected annual loss using the cripple wall height, seismic weight, damping ratio and material properties as features. The XGBoost algorithm provides the most accurate prediction and on average, limits the prediction error to less than 10%. Using the well-developed machine learning models, additional sensitivity analyses are conducted and the effect of model uncertainty on collapse safety and expected annual losses is quantified using Monte Carlo simulation.
For the SWOF buildings, a multi-scale cost-benefit analysis of the Los Angeles Soft-Story Ordinance Retrofit is performed. Individual buildings take an average of four to five years for the reduced earthquake losses to exceed the one-time retrofit cost. At the portfolio-scale, the average cost-benefit ratio is found to be 0.32 for the hypothetical M 7.1 Puente Hills scenario earthquake. A stochastic event-set cost-benefit assessment is also performed, where all events (approximately 8,000) that are significant to the region are considered. From this assessment, it is determined that the probability of achieving a desirable cost-benefit ratio (value between 0.0 and 1.0) within a 50-year period is approximately 0.9.
Lastly, a retrofit design optimization framework is proposed with the goal of maximizing performance-based benefits at the regional scale. The methodology relies a machine learning-based surrogate model to predict seismic performances of retrofitted buildings given the design parameters. Then, a stochastic optimization algorithm is implemented to find the retrofit designs that maximize the improvement in seismic performance for the entire portfolio under a set of pre-defined constraints. The algorithmic retrofit leads to collapse losses that are comparable to the Los Angeles Ordinance guidelines while using only 60% of the resources. The performance-oriented framework is shown to address the inefficiency of conventional strength-based retrofit policies.