[Paper Review] Machine learned Force-Fields for an ab-initio Quality Description of Metal-Organic Frameworks

Sandro Wieser, Egbert Zojer|arXiv (Cornell University)|Aug 2, 2023

Machine Learning in Materials ScienceMaterials Science3 citations

TL;DR

This paper presents two highly accurate, computationally efficient machine learning force fields—based on moment-tensor and kernel-based potentials—for metal-organic frameworks (MOFs). Trained via active learning molecular dynamics on ab initio reference data, they achieve near-DFT accuracy in forces, structural parameters, elastic constants, phonon band structures, and thermal conductivity, outperforming classical force fields while maintaining high efficiency.

ABSTRACT

Metal-organic frameworks (MOFs) are an incredibly diverse group of highly porous hybrid materials, which are interesting for a wide range of possible applications. For a reliable description of many of their properties accurate computationally highly efficient methods, like force-field potentials (FFPs), are required. With the advent of machine learning approaches, it is now possible to generate such potentials with relatively little human effort. Here, we present a recipe to parametrize two fundamentally different types of exceptionally accurate and computationally highly efficient machine learned potentials, which belong to the moment-tensor and kernel-based potential families. They are parametrized relying on reference configurations generated in the course of molecular dynamics based, active learning runs and their performance is benchmarked for a representative selection of commonly studied MOFs. For both potentials, comparison to a random set of validation structures reveals close to DFT precision in predicted forces and structural parameters of all MOFs. Essentially the same applies to elastic constants and phonon band structures. Additionally, for MOF-5 the thermal conductivity is obtained with full quantitative agreement to single-crystal experiments. All this is possible while maintaining a high degree of computational efficiency, with the obtained machine learned potentials being only moderately slower than the extremely simple UFF4MOF or Dreiding force fields. The exceptional accuracy of the presented FFPs combined with their computational efficiency has the potential of lifting the computational modelling of MOFs to the next level.

Motivation & Objective

• To develop machine learning force fields that achieve ab initio-level accuracy for metal-organic frameworks (MOFs) while maintaining computational efficiency.
• To reduce human effort in parametrizing force fields by leveraging active learning and automated data generation from ab initio molecular dynamics.
• To benchmark the performance of two distinct machine learning potential families—moment-tensor and kernel-based—on a representative set of MOFs.
• To validate the force fields against diverse properties including forces, structural parameters, elastic constants, phonon band structures, and thermal conductivity.
• To demonstrate that the resulting force fields are only moderately slower than simple classical force fields like UFF4MOF and Dreiding, enabling large-scale simulations.

Proposed method

• Employ active learning molecular dynamics to generate high-quality reference configurations for training machine learning force fields.
• Use moment-tensor and kernel-based machine learning potentials as the underlying models for force and energy prediction.
• Train the models on a diverse set of configurations sampled from ab initio molecular dynamics simulations of representative MOFs.
• Validate the models against a random set of test structures for forces, geometry, elastic constants, and phonon dispersion relations.
• Benchmark thermal conductivity predictions for MOF-5 against single-crystal experimental data.
• Ensure computational efficiency by optimizing model architecture and inference speed, achieving performance close to classical force fields.

Experimental results

Research questions

• RQ1 Can machine learning force fields achieve near-DFT accuracy in predicting forces and structural parameters for metal-organic frameworks?
• RQ2 To what extent do moment-tensor and kernel-based machine learning potentials maintain accuracy across diverse MOF properties, including elastic constants and phonon band structures?
• RQ3 How does the computational cost of these machine learning force fields compare to classical force fields like UFF4MOF and Dreiding?
• RQ4 Can the machine learning force fields quantitatively reproduce experimental thermal conductivity values, as demonstrated for MOF-5?
• RQ5 Can active learning molecular dynamics reliably generate sufficient and representative training data to enable high-accuracy force field parametrization with minimal human intervention?

Key findings

• The machine learning force fields achieve near-DFT precision in predicted forces and structural parameters across all tested MOFs.
• Elastic constants and phonon band structures predicted by the models show excellent agreement with ab initio reference data.
• Thermal conductivity of MOF-5 is quantitatively reproduced, matching single-crystal experimental measurements.
• The force fields are only moderately slower than classical force fields such as UFF4MOF and Dreiding, enabling efficient large-scale simulations.
• The active learning workflow successfully generates representative training data with minimal human input, ensuring high model accuracy and generalization.
• Both moment-tensor and kernel-based potentials demonstrate robust performance across diverse MOF structures, confirming their reliability and transferability.