Language Priors Enhance Darcy-Flow Inversion in Geological Modeling

In complex inverse problems, the accuracy of a solution heavily relies on the prior information provided, often as much as on the data itself. Much of the valuable engineering knowledge, especially in fields like geology, exists in qualitative descriptions rather than formal mathematical models. This study investigates whether sentence embeddings can bridge this gap by serving as an inference-time interface to integrate geological descriptions into a learned Darcy-flow inverse solver. The researchers tested this approach across six synthetic geological classes and a benchmark reservoir model. By varying only the conditioning representation, they found that text conditioning reduced reconstruction error by 81% compared to a no-text baseline. This improvement largely stemmed from categorical, class-level constraints, particularly in areas where hydraulic head data left the conductivity field underdetermined. While within-class geometric detail was secondary, sentence embeddings also improved training stability and enabled advanced analyses like paraphrase-based sensitivity and open-vocabulary inputs, showcasing their potential as an engineering-informatics tool for injecting expert knowledge.