Physics-Informed Neural Network Prediction of Methane Hydrate Flash Crystallization and Dissociation in Microfluidic...

Methane hydratesice-like crystalline solids that trap methane within a water latticerepresent both a potential energy resource and a critical factor in climate change due to their sensitivity to environmental disturbances. Modeling their formation and dissociation involves complex thermodynamic and kinetic interactions that are difficult to capture with traditional numerical methods, particularly when experimental data are limited. The challenge is further amplified when crystallization is governed by mixed transport- and growth-limited processes occurring on time scales of seconds, as is often the case for methane hydrates. This work reexamines empirical data sets for methane hydrate crystallization and dissociation through Physics-Informed Neural Networks (PINN). The PINN framework operates directly on the governing heat and mass transfer differential equations through a neural-network field representation, meaning errors introduced by the simplifications required to yield closed-form solutions or grid discretization are not admitted by our approach. The presented models efficiently solve the inverse problem. For the growth data set, the PINN lowers the in-sample velocity RMSE by 23.5% relative to the closed-form serial-resistance model of the original study (3.26 versus 4.26 μm/s) while replacing the eighteen per-pressure constants of that model with seven global physical parameters. For the dissociation data set, a single global set of kinetic parameters in place of the per-curve fits reproduces all conditions to within the experimental noise floor and recovers an apparatus-level intrinsic rate constant in agreement with the original analysis. Moreover, leave-one-out cross-validation results indicate generalizability to unseen conditions, a capability not offered by the per-condition fittings. This work establishes PINNs as a scalable and computationally efficient approach for hydrate modeling, bridging the gap between data-driven and physics-based methods. The framework offers broad potential for applications in energy production, carbon sequestration, and climate modeling.