Zhe Jia

My research stands at the intersection of earthquake physics, computational geophysics, and scientific AI. While earthquakes remain one of the deadliest hazards on our planet, their physics are still obscured by the complexity of the Earth’s environment and dynamics. I am dedicated to developing physics-informed computational frameworks that synergize multi-geophysical data with advanced modeling to decipher such high-dimensional dynamical systems. My philosophy is that to fundamentally understand the enigmatic physics of rupture, including how earthquakes initiate, propagate, and arrest, we need to move beyond traditional paradigms, to build a framework that integrates global observations, first-principle simulations, and physics-informed artificial intelligence.

Through systematic analysis of global seismic events, my work has challenged conventional views of how earthquakes rupture. I discovered that large earthquakes are rarely simple slips on a single plane, and instead they often operate as complex cascades that jump across fault networks and/or evolve through different mechanisms. For example, my research on intermediate and deep earthquakes revealed that ruptures sometimes initiate in cold regions with one mechanism yet penetrate into warm zones by transitioning into another mechanism, which allowing their magnitude to grow unexpectedly. To capture these complexities, I utilize a multi-scale observational framework on multi-modal data, to build comprehensive dataset from large global events to understand their rupture complexity.

To resolve complex non-linear processes where standard analytical methods struggle, an important component of my work is the development of next-generation analytical tools. I developed the Subevent MCMC Inversion Tool (SubMIT), a robust Bayesian inference framework that overcomes the limitations of simplified planar fault assumptions. By modeling large earthquakes as sequences of dynamic subevents, this probabilistic framework allows for rigorous characterization of multi-fault multi-stage rupture processes with high computational efficiency. On the deep learning front, I developed scalable algorithms to process big data from dense sensor arrays. Specifically, I applied U-Net deep learning architectures to automate dispersion feature extraction across millions of spectrograms derived from time-series traces. Furthermore, I architect a multi-modal CNN-Transformer algorithm for predictive modeling of source parameters. Validated on large-scale datasets, this end-to-end model synergizes raw time-series data with scalar physical constraints, which shows the power of attention mechanisms in extracting interpretable insights from noisy, high-dimensional data.

My ongoing research aims to operationalize these insights by integrating physics-informed deep learning with simulation and inversion. I am actively architecting hybrid AI models, by leveraging the advantages of Transformers, Graph Neural Networks (GNNs), and Fourier Neural Operators, to process massive spatiotemporal datasets. By training these models on both real-world data and physics-based simulations, I aim to make use of subtle signal patterns that human analysts miss. The hope is to build intelligent systems capable of real-time inference, thereby informing seismic hazard assessment and responses.

See more on my personal website: https://jiazhe868.github.io/