(310l) Transfer Learning of Symmetry-Adapted Learned Collective Variables from Classical to Ab Initio Simulations: Enhanced Sampling of Spin Defect Dynamics in Silicon Carbide

Solid-state spin defects in silicon carbide (SiC) crystal lattices present promising qubit candidates with applications in quantum information science, sensing, and metrology [1]. The mechanisms and conditions for the generation, transport, and annihilation of these defects remains poorly understood. Improved understanding of these mechanisms can inform the realization of long-lived, stable qubits in an industrially relevant substrate. Accurate treatment of defect dynamics requires ab initio calculations accounting for the electronic structure, charge, and spin effects, but the high cost of these calculations typically precludes the direct simulation of these activated rare events, and direct application of enhanced sampling techniques to accelerate these processes is frustrated by a lack of knowledge of good collective variables (CVs) along with to bias sampling.

In this work, we report a strategy to learn CVs in inexpensive classical mechanical calculations and employ transfer learning to deploy them within ab initio enhanced sampling calculations. The underlying hypothesis is that the classical CVs can be easily learned and can provide a sufficiently good approximation for the many-body collective motions underpinning the defect dynamics to accelerate these processes within the ab initio simulations. Importantly, the indistinguishable nature of the C and Si atoms surrounding the defect means that the CVs capturing the defect transitions must respect the underlying translational, rotational, and permutational invariance of the underlying crystal. To this end, we employ our recently developed Permutationally Invariant Networks for Enhanced Sampling (PINES) approach for CV discovery that combines the Permutation Invariant Vector (PIV) featurization with regularized autoencoders to discover CVs that are, by construction, invariant to these symmetries [2].

We demonstrate our approach in applications to divacancy formation, reorientation, and antisite formation in SiC. We conduct classical CV discovery using PINES coupled to LAMMPS and then deploy the ab initio enhanced sampling calculations using the SSAGES enhanced sampling libraries coupled to the Qbox and CP2K simulation packages. As an additional outcome, we use the data harvested from our enhanced sampling calculations to parameterize new equivariant deep learning potentials for these processes in SiC.

References

1. E.M.Y. Lee, A. Yu, J.J. de Pablo, and G. Galli, Nat. Commun. 12 6325 (2021)
2. N.S.M. Herringer, S. Dasetty, D. Gandhi, J. Lee, and A.L. Ferguson, J. Chem. Theory Comput. 20 1 178-198 (2024)