ScoreField boosts EM imaging by 1.8 dB with generative priors

Inverse scattering is a notoriously ill-posed problem: you need to reconstruct an object's electromagnetic properties from scattered wave measurements, while respecting nonlinear full-wave physics. Classical solvers struggle with strong multiple scattering, where wave interactions become deeply entangled. A new approach called ScoreField, from researchers Wenhan Guo, Yuan Gao, and Yu Sun, attacks this by combining two implicit neural representations (INRs)—one for the unknown permittivity contrast, one for the induced current fields—and jointly optimizing them under the Lippmann-Schwinger equations, the exact integral equations governing scattering. Crucially, they layer in a pretrained score-based generative prior that provides a learned gradient on the contrast, propagated through the contrast INR via the chain rule. This hybrid gives the model both a physics constraint and a powerful data-driven prior, avoiding the need for explicit regularization tuning.

On benchmarks that include weak- and strong-scattering simulations, the canonical Austria phantom, and experimental Fresnel measurements, ScoreField consistently outperforms both classical full-wave methods and deep learning baselines. The paper reports an average PSNR improvement of 1.8 dB over the best competing method on real Fresnel data, with visibly reduced artifacts and better edge preservation. Several extensions have been suggested, including testing on more complex 3D scattering geometries and exploring the ability of the score-based prior to transfer across a range of frequencies. ScoreField's workflow is also modular—its regularization comes from generative priors, which are reusable across many inverse problems, not just scattering. The authors note that because the method is self-supervised through physics, it requires no labeled ground-truth images for training. This is an exciting step forward for computational imaging, as it demonstrates a practical fusion of physics-based models and generative AI for solving tough inverse problems.