(637f) Developing Phase-Transferable Machine Learning Force Fields of Molten Salts

A fundamental understanding of the behavior of molten salts is important for their applications in molten salts reactors, which are considered a leading candidate for the next generation of nuclear reactors. In order to obtain accurate physical properties in molecular dynamics (MD) simulations, accurate and computationally efficient force fields are required. In recent years, machine learning (ML) techniques have been widely applied in force field development. ML force fields are able to capture many-body interactions with accuracy close to ab initio calculations, and they are not much more computationally expensive than classical force fields. ML force fields are particularly appealing for simulations of molten salts, where polarizability (i.e., many-body interactions) are often important.

ML force fields are commonly trained on ab initio MD (AIMD) simulations, and evaluated by their ability to reproduce ab initio energies and forces. However, this training procedure is necessarily performed with very small simulation boxes and short time scales, due to the large computational cost of AIMD simulations. The ML force field is then generally used to simulate much larger systems over longer timescales and calculate properties that would be difficult-to-impossible to evaluate with AIMD. Therefore, an important question to consider is whether a model trained on small simulation boxes can predict properties that can only be observed in larger systems. To rigorously answer this question, it is necessary to know the properties of the model used to generate the training data, which is not possible with AIMD.

In order to investigate the effect of small training system sizes on ML force fields, we trained an ML force field for MgCl₂ to reproduce an induced dipole model, named the polarizable ion model (PIM). MgCl₂ was selected because the liquid structure has long range ordering that PIM simulations capture and which is not observed in small box sizes (e.g., AIMD) due to system size effects. The PIM also captures complicated many-body interactions with induced dipoles, and is therefore more sophisticated than simple fixed charge models. Moreover, most physical properties of the PIM for MgCl₂ can be, or have already been, accurately calculated. This makes PIM MgCl₂ a great candidate to examine how the accuracy of a ML force field depends upon the time- and length-scales of the simulations used to train the model.

We develop a training scheme to make the ML force field phase-transferrable. Our training scheme efficiently generates a large amount of uncorrelated training data from both the crystalline and liquid phases by a two-step procedure. We evaluated a wide range of physical properties, including density, self-diffusivity, crystal and liquid structure, and melting point. We find excellent agreement between the ML force field and the PIM used for training. The ML force field of MgCl₂ is phase-transferrable, as it successfully reproduces accurate physical properties in both liquid and crystalline phases, compared to the original PIM. The melting point of the ML force field is also within 10 K of the PIM. Of particular note, we find that when performing the simulations in larger boxes, the ML force field reproduces the long range ordering of MgCl₂ liquid, even though the force field is entirely trained from PIM simulations in small simulation boxes where such structures cannot form. Our work provides strong evidence that ML force fields of molten salts can correctly capture properties that appear on time and length scales beyond those observed in the training simulations. By training an ML force field on a computationally tractable model, we were able to rigorously test this hypothesis and validate our training procedure for developing ML phase-transferable force fields of molten salts.