Berlin 2024 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

DY: Fachverband Dynamik und Statistische Physik

DY 5: Machine Learning in Dynamics and Statistical Physics I

DY 5.2: Talk

Monday, March 18, 2024, 09:45–10:00, BH-N 243

Machine-learned Potentials for Vibrational Properties of Acene-based Molecular Crystals — •Shubham Sharma and Mariana Rossi — Max Planck Institute for the Structure and Dynamics of Matter

Machine-learning potentials (MLPs) have allowed the efficient modelling of complex atomistic systems with ab-initio accuracy. Normally, the construction of sufficiently large and diverse reference datasets, using first-principles calculations, is a bottleneck for training. Therefore, several active-learning strategies have been proposed, which aim to make the training more efficient, especially when used together with molecular-dynamics techniques [1]. In this work, we explore building protocols for training sets of high-dimensional neural network potentials (HDNNPs), targeting specifically an accurate description of the vibrational properties of weakly-bound condensed-phase systems. For that, we show how we can use and augment the committee-model framework within the i-PI code [2]. We show results for acene-based molecular crystals and discuss the advantages and limitations of different learning strategies to treat different crystal polymorphs, at various thermodynamic conditions. [1] C. Schran et. at., J. Chem. Phys. 153, 104105 (2020). [2] V. Kapil et. al., Comput. Phys. Commun. 236, 214 (2019).

Keywords: Machine Learning Potentials; Active Learning; Molecular Dynamics; Weakly-bonded Materials