When an earthquake hits, you want the building you’re in to be secure.
Civil & Environmental Engineering Assistant Professor Elnaz Seylabi is developing a method to make structures safer by improving predictions of how they will behave when the ground shakes. She is advancing a new computational framework that integrates machine learning with physics-based knowledge of geostructural systems.
This innovative framework, Seylabi explains, could lead to surrogate models that run significantly faster than today’s high-fidelity numerical simulations while still honoring the underlying mechanics. Such a leap in both speed and reliability could transform earthquake engineering practices concerning soil-structure systems.
“When I joined the 91·çÃùÄñ³ª in 2019, that was a field I wanted to explore,” Seylabi said.
Her research has now been recognized by the National Science Foundation (NSF), which this spring awarded her a five-year, $608,000 grant through its Faculty Early Career Development Program (CAREER) — the NSF’s most prestigious award for early-career faculty.
“Dr. Seylabi is an extremely talented researcher in the area of engineering mechanics and seismology,” said Indira Chatterjee, interim dean of the College of Engineering. “We are very pleased that the NSF has recognized her potential with this prestigious CAREER award.”
The challenge of high-fidelity numerical models
Seylabi’s CAREER-funded project, “A Physics-Informed Machine Learning (PIML) Framework for Surrogate Modeling of Geostructural Systems,” seeks to overcome longstanding barriers in geostructural research. These are challenges she has been engaged with since earning her doctorate degree in civil engineering from the University of California, Los Angeles (UCLA) in 2016.
Today’s state-of-the-art numerical models — essentially, mathematical representations of physical systems — can simulate the behavior of soil and structures with high fidelity. However, these models come at a steep computational cost: a single probabilistic analysis, which incorporates problem uncertainties through statistical modeling, may require a significant computing budget.
Compounding this is the difficulty of calibrating such models with experimental and field data, as well as the persistent challenge of characterizing modeling bias and errors.
“We need to solve this optimization problem at the system level,” Seylabi said.
Toward rapid, reliable predictions
Seylabi’s CAREER project aims to tackle both challenges. By embedding governing physics directly into the machine learning training process, the PIML framework penalizes violations of physical laws—reducing the need for massive datasets and improving model generalizability.
The outcome is a surrogate model that maintains physical interpretability, lowers training data requirements and offers near-real-time seismic response predictions. This advancement paves the way for more responsive sensor integration, enhanced seismic risk assessments and adaptive management strategies for geostructural systems.
Although her CAREER project is primarily computational, Seylabi plans to validate the framework using data collected from prior large-scale experiments.
“In the long run, this work can benefit not only my own research but also other researchers working on complex dynamical systems,” she said. “It can serve as a new framework for efficiently analyzing such systems while accounting for uncertainty at the system level.”
The project also has an important educational component. Seylabi plans to integrate scientific machine learning concepts into her civil engineering courses and develop outreach modules for high school students in the region.
“We’re equipping the next generation of civil engineers to apply machine learning responsibly,” she said.
