Probabilistic Data-Driven Modeling & Structural Monitoring Lab

Dr. Audrey Olivier

USC Viterbi Sonny Astani Department of Civil and Environmental Engineering

Our research in a few words...

Our research team at USC explores topics at the intersection of physics-based modeling, data analytics and uncertainty quantification for applications in monitoring and design of civil structures. We are excited about the potential that data, in conjunction with mechanistic models, holds to help build the resilient and smart urban environments of tomorrow.

The applicability of machine learning algorithms to scientific and in particular civil engineering problematics requires careful considerations of data scarcity and imbalance, inherent randomness and stochasticity of the driving systems inputs, and compliance with well-established physics principles. Our research team works on the development of advanced data analytics tools that embed physical knowledge and quantify uncertainties, allowing for reliable decision-making.

... and some applications

In Structural Health Monitoring (SHM), sensing data is leveraged to monitor a structure's behavior and detect potential damage at its onset. Rapidly evolving trends such as increased urbanization or intensification of climate-related hazards generate considerable loads on our aging infrastructure and motivates the need for reliable SHM for better infrastructure management and rapid response to high-intensity events such as earthquakes. In this context, leveraging data in conjunction with mechanistic models is of paramount importance to not only detect but also understand damage and its effect on a structure's integrity. In our group we develop algorithms for probabilistic identification of structural systems using time-varying data, using e.g. enhanced Bayesian filtering algorithms or probabilistic machine learning.

The potential of probabilistic modeling that leverages both physics and data spans various scales, from systems to structures to materials. Scientific machine learning methods can be leveraged to accelerate high-fidelity physics-based computational models. In this context we are working on the development of efficient Bayesian Neural Networks (BNNs) that can complement expensive simulations for applications as varied as materials modeling and power grid contingency analysis. Careful quantification of uncertainties via the Bayesian framework is key to providing confidence in predictions of data-driven models, especially for out-of-distribution predictions, and further allows to future experiments via active learning techniques.

In a smart-city context, probabilistic data analytics tools can also be used to optimize processes and infrastructure systems. As an example, we have worked on building probabilistic ambulance travel time predictors to help optimize emergency medical services.