Operator-informed GP solves complex Helmholtz wavefields with uncertainty

Researchers from Johns Hopkins University (Boyuan Deng, Kshitiz Upadhyay, Michael Shields) have developed a probabilistic solver for the complex Helmholtz equation, which governs wave propagation in dissipative media like biological tissue. Their key innovation is an operator-informed Gaussian process (GP) that realifies the complex operator into an equivalent coupled real block, enabling standard real-valued GP conditioning. The method supports multiple priors—diagonal, coregionalized, and multiscale—and conditions on both PDE residuals and boundary traces.

On synthetic benchmarks from 1D to 3D, the solver is competitive with deterministic finite-difference and neural-network methods while requiring far fewer interior constraint points. Critically, it quantifies uncertainty via a posterior distribution over the wavefield. In a real-world test on in vivo brain magnetic resonance elastography, a proper multiscale prior achieved a correlation of 0.77 against measured shear curl fields, exceeding the 0.75 target. The improvement came from the multiscale kernel, not from real–imaginary coupling. However, the authors note a low-frequency accuracy ceiling from model mismatch and that the posterior uncertainty is not yet calibrated—marking calibrated uncertainty as the next frontier for probabilistic wavefield inference.