The Right Measure for Physics-Constrained Generation: A Co-Area Correction for Posterior-Consistent PDE Inverse Problems
Researchers have identified a fundamental issue in how generative models, such as diffusion and flow matching, are used to solve physics-constrained inverse problems. The common practice of enforcing physics as a hard constraint leads to sampling from an incorrect distribution, omitting a crucial co-area Jacobian factor. This omission can inflate posterior errors significantly, up to 20 times the sampling-noise floor. To address this, a new measure-aware sampler called CoCoS has been developed, which accurately targets the correct co-area posterior, aligning with gold-standard results. AI
IMPACT Corrects a fundamental flaw in scientific inference using generative models, improving uncertainty quantification.