UC San Diego

We conducted ambient noise tomography (ANT) using data from 342 seismographs distributed over a $50 \times 50$ km area encompassing the July 2019 Ridgecrest earthquakes (M7.1 and M6.4). By employing the locally sparse tomography (LST) method—an unsupervised machine learning approach—we effectively modeled small-scale geophysical structures using only data from the study region. The Rayleigh group speed derived from LST showed superior accuracy in predicting travel times compared to conventional regularized least-squares inversion techniques. From surface wave dispersion maps, we constructed a 3D shear velocity model, which revealed a heterogeneous low-velocity zone (LVZ) surrounding the causative faults, with a significant 40% reduction in shear wave velocity primarily concentrated in the upper 2–3 km. Additionally, imaged LVZs associated with inactive portions of the Little Lake Fault System suggest the presence of enduring damage zones.

Building on this work, we introduced a method to fuse multiresolution seismic tomography models with physics-informed probabilistic graphical models (PIPGMs), which incorporate physical constraints like ray-path density. Using synthetic checkerboard models and fault zone structures from the Ridgecrest earthquake sequence, the PIPGM fusion method demonstrated marked improvements in travel-time residuals, image quality, and peak signal-to-noise ratios compared to conventional approaches. This novel fusion framework enables the integration of gridded velocity models of varying resolutions, offering a powerful computational tool for enhancing the quality and interpretability of seismic imaging data. We extended the Probability Graphical Model (PGM) to 3D velocity models combining tasks, including synthetic and real fault zone structures, and demonstrated significant improvements—reducing computed travel-time residuals by 44% for 3D models, respectively, over conventional methods. Unlike traditional techniques, the PGM’s adaptive weighting preserves complex features from high-resolution data and effectively transfers these enhancements to the broader low-resolution background, making it a valuable tool for advancing seismic tomography and ground motion simulation.

We applied the newly proposed Mixture of Experts (MoE) to combine the features learned from multiple machine learning models and enhance the capabilities of medical Internet of Things (IoT) systems for blood pressure monitoring by integrating advanced technologies such as flexible sensors, wireless data communication, and intelligent algorithms. This integration facilitates efficient data collection, real-time analytical processing, and seamless online sharing of results. This model employs a dynamic MoE gating network, which intelligently directs incoming data to specialized expert models that are each optimized for distinct measurement scenarios. This model fusion through the MoE framework enables the system to adaptively respond to diverse environmental and individual physiological characteristics, thereby significantly enhancing the accuracy of blood pressure predictions by 15–18% over conventional methods.