Speaker
Description
Delicate interplay of non-covalent interactions (NCI) governs the properties and phase behavior of molecular materials, being unprecedentedly rich already for the simplest molecules, such as water. Existing first-principles methods enable to model properties and polymorphism of molecular solids at least qualitatively [1]. However, those methods still frequently fail to yield at least semi-quantitative predictions [2], and are completely inapplicable for modelling properties of liquids or the solid-liquid equilibrium. Immense computational complexity inherently related to large-scale ab initio simulations of liquids and glasses has impeded this field of research for long [3]. Current work presents a newly developed fragment-based ab initio Monte Carlo approach for highly accurate predictions of structural and thermodynamic properties of amorphous molecular materials that require minimum empiric inputs [4]. A sophisticate nested Monte Carlo sampling that combines multiple potentials is a key feature enabling to reach a high computational accuracy at a reasonable computational cost. Since the best description of the underlying non-covalent interactions can be obtained with the coupled-clusters theory, aim of this work is to incorporate the coupled-clusters accuracy into simulations of properties such as thermal expansivity or compressibility of liquids. A suitable composite combination of force fields, semi-empiric quantum-chemical methods [5] and fragment-based ab initio many-body expansion [6] for the individual nested Monte Carlo layers represents a unique development in this field. Adoption of the many-body expansion of the cohesive energy and modern low-scaling formulations the coupled-clusters theory prove to be viable, indeed imparting the coupled-clusters accuracy to the amorphous realm of molecular liquids and glasses. A crucial feature of the computational protocol, greatly improving its efficiency, robustness and the applicability range is a ∆-machine learning scheme that is used to correct the outputs of the cheap and fast quantum-chemical models up to the standards of the coupled-clusters theory. The architecture of equivariant message-passing neural networks [7] is used to exploit the vast libraries of molecular interactions generated during the initial full quantum chemical sampling of the bulk materials. Most of the expensive quantum-chemical evaluations of molecular interactions energies can be then bypassed in latter bulk simulation stages using thence trained potentials. Validation of the computational methodology is presented for three small model systems, but future development for material design of pharmaceuticals or organic semiconductors is expected.
References
1. Červinka, C.; Štejfa, V.; Pokorný, V.; Touš, P.; Růžička, K. Cryst. Growth Des. 2023, 23, 9011.
2. Touš, P.; Červinka, C. Cryst. Growth Des. 2023, 23, 4082.
3. Červinka, C. J. Mol. Liq. 2025, 417, 126600.
4. Červinka, C. J. Chem. Theory Comput. 2025, 21, 9868.
5. Ludík, J.; Kostková, V.; Kocian, Š.; Touš, P.; Štejfa, V.; Červinka, C. J. Chem. Theory Comput. 2024, 20, 2858.
6. Červinka, C.; Beran, G. Chem. Science 2018, 9, 4622.
7. Schütt, K.; Unke, O.; Gastegger, M. Proc. Mach. Learn. Res. 2021, 139, 9377.
Financial support from the Czech Science Foundation (23-05476M) is acknowledged. Computational resources were supplied under the project e-INFRA CZ (ID:90254).