Understanding the phase behavior of water under extreme conditions is crucial in fields such as geoscience and planetary science. Under high temperatures and pressures (e.g., 1000s of K and 10s to 100s of GPa), water exhibits a rich phase diagram that includes complex “superionic” phases characterized by highly diffusive hydrogen atoms moving among crystalline oxygen, in addition to molecular and ionic solid and liquid phases. However, many of its properties remain a subject of ongoing debate. Simulations using Machine Learned (ML) interatomic potentials (IAPs) provide a powerful framework for exploring these behaviors by enabling quantum-accurate simulation on far greater spatiotemporal scales. However, developing models for highly dense reactive systems, such as water under extreme conditions, remains computationally intensive, in part due to the lack of quantum mechanical (QM) training data in high-energy, non-equilibrium configurations. To address this challenge, this study integrates hierarchical transfer learning within the ChIMES ML-IAP framework [1–3] , which allows the fitting problem to be decomposed into smaller independent and reusable parameter blocks – e.g., fitting a model for H/O systems can be achieved by fitting separate models for H and O systems, then learning parameters describing O/H cross interactions, and then linearly combining models at simulation time. Beyond simplifying the fitting problem, this strategy makes resulting models robust for pure element systems as well. The resulting model is presented, along with a discussion of strategies used for hyperparameter optimization, data weighting, and active learning.
References
[1] Rebecca Lindsey. ChIMES Generator. https://github.com/rk-lindsey/chimes_lsq.
[2] Rebecca Lindsey. ChIMES Calculator. https://github.com/rk-lindsey/chimes_calculator.
[3] Rebecca Lindsey. ChIMES Active Learning Driver. https://github.com/rk-lindsey/al_driver.
