Modeling Porous Liquids with Machine-Learning-Assisted Molecular Dynamics
<p dir="ltr">Diverse applications benefit from porous materials, such as molecular separations and catalysis, enabling the efficient capture of greenhouse gases (CO<sub>2</sub> and CH<sub>4</sub>) and valuable noble gases (Xe, Ar, and Kr). Xenon, widely used i...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Published: |
2024
|
| Subjects: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | <p dir="ltr">Diverse applications benefit from porous materials, such as molecular separations and catalysis, enabling the efficient capture of greenhouse gases (CO<sub>2</sub> and CH<sub>4</sub>) and valuable noble gases (Xe, Ar, and Kr). Xenon, widely used in optics, medicine, and nuclear processes, poses extraction challenges due to its low atmospheric abundance and inert nature, driving high commercial costs. Efficient xenon isolation requires materials with precise selectivity and high adsorption capacity. Porous liquids (PLs) incorporating cavities formed by porous organic cages (POCs) show promise in addressing these challenges. Understanding the binding mechanisms, occupancies, dynamics, and equilibrium between host (PL/POC) and guest (Xe) is pivotal for the design of novel POCs tailored to specific functionalities. Molecular dynamics (MD) simulations have proven essential for understanding and exploring the physicochemical processes governing these systems. In MD simulations, atom movements are described by the potential energy surface (PES) of the system. Typically, the PES is accurately obtained by calculating the electronic structure using methods such as density functional theory (DFT). However, the combination of MD with DFT, although provides precise interatomic forces, is constrained by computational scalability, limiting simulations to tens of picoseconds and a few hundred atoms, thereby falling short of capturing realistic timescales and size of these porous systems. In recent years, machine learning, particularly neural networks (NNs), has emerged as a promising avenue to address these limitations, by learning accurate interatomic potentials from a set of high-fidelity reference calculations while maintaining computational efficiency.</p><p dir="ltr">In this work, we present accurate and data-efficient machine learning interatomic potential (MLIP) models built using Allegro, a local equivariant deep NN architecture. These models were trained, validated, and tested on DFT level data, covering energies, forces, and virials in structures with 600 to 1170 atoms (H, C, N, O, F, Cl, Xe). The structures include varying numbers of xenon atoms in different PLs and POCs, totaling 1.7 million atoms with around 12 million data points. The MLIP models enable simulating large-scale porous liquids at realistic physicochemical conditions and applied to provide microscopic interpretation of experimental <sup>129</sup>Xe NMR — a local probe critical for understanding the condition-dependent dynamic processes present in these systems — data both at the static and dynamic levels.</p> |
|---|