Capturing the Vibrational Dynamics of Acene-Based Molecular Crystals with Machine Learning

Vibrational dynamics in molecular crystals are inherently complex and play an important role in applications ranging from organic electronics to pharmaceuticals. However, due to long-range molecular interactions, modeling these dynamics is computationally expensive. Machine Learning Interatomic Potentials (MLIP) offer a promising approach to address this challenge.

In this work, we construct and test MLIPs based on MACE architecture [1], aiming for accurate modeling of vibrational dynamics in acene-based molecular crystals. We utilize a committee-based active-learning approach [2] for training and benchmark the MLIP against on-the-fly machine learning force field method [3]. By propagating the uncertainty from the committee, we further quantify training-related errors in the vibrational density of states. Furthermore, we explore the ability of MLIP to generalize, both interpolatively and extrapolatively, by systematically applying different active-learning strategies across several acene members. Finally, we test the validity of the generalized MLIP for a host-guest system relevant to quantum technologies [4].

[1] Adv. Neural Inf. Process. 35, 11423-11436 (2022)

[2] J. Chem. Phys. 153, 104105 (2020)

[3] Phys. Rev. B 100, 014105 (2019)

[4] Phys. Rev. Lett. 127, 123603 (2021)