Peak Ground Velocity Predictor

Instead of expensive physics simulations, we're using machine learning to instantly generate simulated maps of earthquake impact.

Comment

Instantaneous Physics-Based Ground Motion Maps

Inspiration

Earthquake early warning systems have made enormous strides — but one gap remains: spatial ground motion maps. When a quake strikes, seismometers can estimate source parameters within seconds. Yet translating those parameters into a full map of where the shaking is worst still requires expensive physics-based simulations that can take hours on high-performance computing clusters. In the critical first minutes of a disaster, that gap costs lives.

We were inspired by the work of Rekoske, Gabriel, & May (2023), who showed that reduced-order modeling could bridge exactly this gap. We wanted to bring that research to life interactively — making the science tangible and deployable.

What We Built

We built an interactive Peak Ground Velocity (PGV) predictor powered by a reduced-order model. Given four earthquake source parameters — hypocentral depth, strike, dip, and rake — our model produces a full 60×60 spatial PGV map.

The frontend lets you drag sliders and watch the ground motion map update in real time, visualizing how fault geometry shapes the spatial distribution of shaking.

How We Built It

The model uses simulated PGV maps from the Ground Velocity from Earthquake Simulations dataset as training data. First, it compresses each 60 x 60 shaking map into a small set of dominant spatial modes using SVD/POD. Then it learns a smooth mapping from earthquake parameters to the weights of those modes using RBFinterpolation. At prediction time, it reconstructs the full shaking map from those predicted weights.

Challenges

  • Dimensionality of the RBF: With 1,800 training points and $r$ retained modes, the RBF system matrix is $1800 \times 1800$. Solving it is feasible but slow — we had to be deliberate about not retraining unnecessarily.

  • Scaling the UI to real physics: Bridging a Python-based scientific model to a responsive web interface required careful API design so that inference latency stayed imperceptible to the user.

  • Interpreting the physics: Understanding why the PGV maps look the way they do — rupture directivity, hanging-wall effects, radiation patterns — helped us build a more meaningful demo rather than just moving sliders.

What We Learned

  • Reduced-order modeling is a powerful but underutilized tool in scientific computing. SVD-based compression can reduce a problem from thousands of degrees of freedom to a handful of modes without sacrificing meaningful accuracy.
  • The gap between a research paper and an interactive, explainable demo is where a lot of real-world impact gets created.
  • Physics and machine learning aren't opposites — the best models respect the structure of the underlying science.
Share this project:

Updates