Abstract
Well log interpretation faces significant challenges due to data scarcity, distribution shift across geological basins, and a lack of physics-informed representations in current foundation models. While recent masked autoencoder approaches show promise for self-supervised pretraining on subsurface data, they fail to consider fundamental petrophysical constraints that govern measurement relationships. This paper presents a physics-aware masked pretraining framework that incorporates petrophysical consistency losses into self-supervised learning on multivariate well logs. The proposed method combines masked patch reconstruction with differentiable petrophysical forward models that use Archie's law, Wyllie's time-average equation, and density-porosity relationships as auxiliary objectives. A geologically adaptive masking curriculum also gradually increases task difficulty, moving from random point masking to cross-curve masking and stratigraphic interval masking. Extensive experiments on the FORCE 2020 benchmark show that physics-aware pretraining improves lithology classification accuracy by 4.2% and reduces porosity estimation error by 5.8% compared to standard masked autoencoders, while cross-basin transfer experiments show a 6.3% lower optimality gap when transferring datasets from Norwegian to Kansas wells. Ablation studies show that petrophysical consistency losses contribute 2.8 percentage points to classification improvement, while geologically adaptive masking adds another 1.4 percentage points. The framework maintains computational efficiency, with inference times of less than 12 milliseconds per well section, making it suitable for real-time drilling applications in data-scarce, multi-basin settings.