Data-Driven Many-Body Potentials for Realistic Molecular Simulations
CIC nanoGUNE Seminars
- Speaker
-
Francesco Paesani
University of California - When
-
2024/06/18
15:00 - Place
- CIC nanoGUNE Seminar room, Tolosa Hiribidea 76, Donostia-San Sebastian
- Host
- Pablo Piaggi
- Add to calendar
-
iCal
![Data-Driven Many-Body Potentials for Realistic Molecular Simulations Data-Driven Many-Body Potentials for Realistic Molecular Simulations](https://dipc.ehu.eus/en/scientific-activities/joint-seminar-agenda/cic-nanogune/data-driven-many-body-potentials-for-realistic-molecular-simulations/@@images/e7bfdf61-6a13-44c1-8234-00a612fc4fd2.png)
Molecular simulations are instrumental in gaining detailed insights into the properties of complex systems, from biomolecules to materials. At the heart of these simulations is the potential energy function (PEF), which maps out the multi-dimensional energy landscape of the system in question. The accuracy of the PEF is crucial, as it dictates the realism and, consequently, the predictive power of any simulation. In principle, correlated wave function (WF) methods, such as coupled cluster with single, double, and perturbative triple excitations, i.e., CCSD(T), can provide an accurate description of molecular interactions from isolated molecules to condensed-phase systems. However, their computational cost limits their applications to systems with only a few tens of atoms. Conversely, conventional force fields (FFs) grounded in empirical parameterizations exhibit limited accuracy and lack predictive power, approximating molecular interactions with simplistic models based on harmonic potentials and classical Coulomb interactions. In this presentation, I will describe our data-driven many-body energy (MB-nrg) formalism that exploits the locality of quantum mechanics and allow for bridging the gap between WF methods and FFs, paving the way for realistic computer simulations from the gas to the condensed phase. MB-nrg seamlessly integrates data-driven machine-learned representations of individual many-body interactions with physics-based many-body models, all rigorously derived from CCSD(T) data. I will first illustrate the accuracy, transferability, and predictive power of our MB-nrg potentials for various aqueous systems across different thermodynamic states and in different environments. I will then discuss the extension of our MB-nrg formalism to density functional theory within the density-corrected framework (DC-DFT) and introduce MB-DFT(DC), a general data-driven formalismthat combines the ansatz of the many-body expansion with DC-DFT. I will show that MB-DFT(DC), accurately describes molecular interactions in aqueous systems from the dimer to the condensed phase, effectively elevating DFT to CCSD(T) accuracy for molecular interactions.