Science-guided transfer learning for molecular dynamics of confined fluids in shale nanopores

The phase behavior and properties of confined fluids play a critical role in subsurface energy and environmental operations. Predicting these behaviors in porous media typically relies on Molecular Dynamics (MD) simulations, which, while accurate, are prohibitively expensive for large-scale applications. Deep learning (DL) has recently emerged as a promising alternative for developing surrogate models of such processes. However, conventional DL architectures require large volumes of training data—an impractical requirement given the high cost of generating MD datasets. To address this challenge, transfer learning can be employed: models are first trained on related, lower-cost tasks and subsequently adapted to the target task with limited data. This strategy has been highly effective in domains such as natural language processing and computer vision, but its application to confined fluid modeling remains underexplored. In this work, we present NanoSG, a science-guided deep learning framework for emulating MD simulations of fluid mixtures in confinement. NanoSG integrates domain knowledge with pre-trained representations to enhance learning efficiency and physical consistency. Through extensive experimentation, we show that NanoSG achieves robust generalization, with a minimum performance improvement of 16.26% over baseline models, while maintaining consistency with established scientific principles despite being trained on limited MD data. Our results highlight the potential of science-guided transfer learning to accelerate predictive modeling of confined fluids under data-scarce conditions, opening new avenues for scalable simulation in energy and subsurface applications.