Research area 1
Combining physical wave equations with deep learning operators to build subsurface models that stay interpretable, not just accurate on a benchmark.
Core focus & scope
As the foundational methodology area of the Institute, this work bridges wave mechanics and numerical analysis with modern AI. Physical laws are built directly into deep learning loss functions, which keeps predictions geophysically consistent instead of leaving them as a black box.
In practice: physics-informed neural networks (PINNs), multi-dimensional acoustic and elastic forward modeling, HPC acceleration, and automated processing of high-density seismic datasets.
Embedding acoustic and elastic wave equations directly into neural operator loss functions, to solve forward and inverse problems without needing massive labeled training sets.
GPU-accelerated finite-difference and spectral-element algorithms for real-time 3D wavefield simulation and large-scale elastic parameter inversion.
Algorithms for processing high-throughput Distributed Acoustic Sensing (DAS) streams, for microseismic event detection and continuous monitoring.
Bayesian neural networks and variational inference, to attach honest confidence intervals to inverted velocity and impedance models rather than a single point estimate.