Molecular force fields with gradient-domain machine learning (GDML): Comparison and synergies with classical force fields

Huziel E. Sauceda, Michael Gastegger, Stefan Chmiela, Klaus‐Robert Müller, Alexandre TkatchenkoThe Journal of Chemical Physics 2020 · DOI

TL;DR

This paper bridges the gap between highly accurate but computationally expensive quantum mechanics-based methods and fast but less accurate classical force fields. By comparing gradient-domain machine learning (GDML) force fields with traditional molecular mechanics (MM) force fields, the authors identify shortcomings in MM force fields and propose strategies to improve them, paving the way for more efficient and accurate molecular simulations.

What Problem Does This Paper Address?

Simulating molecular systems is crucial across many scientific disciplines. Traditional electronic structure methods (like DFT) offer high accuracy but are computationally prohibitive for large systems or long timescales. Classical molecular mechanics force fields (MM-FFs) are much faster and can handle larger systems, but their fixed functional forms limit their accuracy and ability to capture complex quantum mechanical phenomena like bond breaking/formation or many-body effects. Modern machine learning force fields (ML-FFs) promise a middle ground, achieving near-quantum accuracy at significantly reduced computational cost. However, a direct comparison between ML-FFs and MM-FFs, and understanding how they can complement each other, has been lacking. This paper aims to fill that gap by analyzing the strengths and weaknesses of both approaches and exploring synergistic integration strategies.

Methods

The core of the study involves comparing the performance of a specific ML-FF, the symmetric Gradient Domain Machine Learning (sGDML) model, with several classical MM-FFs. The sGDML model is trained on forces and reconstructs potential energy surfaces (PESs) and energies. The MM-FFs investigated include the Generalized Amber Force Field (GAFF), Merck Molecular Force Field (MMFF94), Ghemical, and Universal Force Field (UFF). The comparison focuses on the accuracy of reconstructed PESs, the ability to capture specific chemical phenomena (like lone pair interactions and non-classical electronic effects), and the resulting molecular dynamics and vibrational properties.

The paper also explores two main avenues for synergy:

  1. Generating reference data: Using MM-FFs to generate configurations for training ML-FFs, leveraging the speed of MM-FFs for sampling.
  2. Augmenting MM-FFs: Using ML to improve MM-FFs by either reparametrizing existing functional forms or introducing new, more expressive terms to capture quantum effects.

Datasets

The study utilizes several molecular systems for training and evaluation, including:

  • Ethanol: Used to study lone pair interactions and rotor coupling.
  • Keto-malondialdehyde (keto-MDA): Used to investigate non-classical electronic effects and symmetric degrees of freedom.
  • Acetylsalicylic acid (Aspirin): Used to examine the ordering of equilibrium configurations and the impact of specific quantum interactions (n → π*).

Reference data for training the sGDML models was generated using high-level ab initio methods like CCSD(T) and DFT (PBE with van der Waals corrections). For MM-FF comparisons, PES scans were performed. The paper also discusses the use of the MD17 dataset, which consists of DFT-generated molecular dynamics trajectories, as a source for reference data.

Key Insights and Ideas

  • MM-FFs miss crucial quantum effects: Classical MM-FFs, even those with more complex functional forms, often fail to capture subtle but important quantum mechanical interactions, such as the coupling between functional groups due to lone pair electrons or specific attractive interactions like n → π*. This leads to inaccurate PESs and potentially incorrect predictions of molecular behavior.
  • ML-FFs offer a more faithful representation: sGDML, as a representative ML-FF, can accurately reconstruct PESs from high-level quantum chemistry data, capturing these subtle quantum effects and providing a benchmark against which MM-FFs can be evaluated.
  • Synergy is key: The limitations of MM-FFs and the computational cost of ML-FFs highlight the need for integration. MM-FFs can be used for efficient sampling, and ML can be used to improve MM-FFs by reparametrization or by adding ML-derived terms that capture missing physics.
  • Data generation is critical: The quality and representativeness of the training data are paramount for ML-FFs. While MM-FFs can generate data quickly, their inherent limitations mean that ML models trained solely on MM-FF data may not generalize well to higher-level quantum mechanics.
  • Augmenting MM-FFs can be powerful: Even simple ML-based extensions to MM-FFs can significantly improve their accuracy and ability to capture specific phenomena, demonstrating the potential for hybrid approaches.

Key Results

  • PES Comparison: sGDML models trained on CCSD(T) data accurately reproduced complex PES features for ethanol and keto-MDA, while GAFF qualitatively reproduced global minima but missed important couplings and local minima. Other MM-FFs showed varying degrees of success, with some misrepresenting large portions of the PES.
  • Vibrational Density of States (VDOS): GAFF provided good predictions in the fingerprint region but struggled with bond stretching vibrations involving hydrogen atoms. It also showed a lack of temperature-dependent shifts in frequencies, indicative of its harmonic approximations. sGDML, on the other hand, captured these anharmonic effects and temperature dependencies.
  • Aspirin Minima Ordering: GAFF incorrectly ordered the lowest energy minima of aspirin, failing to capture the stabilizing n → π* interaction, which was correctly represented by sGDML.
  • MM-FF Sampling: DFT and GAFF molecular dynamics simulations at 500K showed distinct sampling behaviors, particularly for aspirin, where GAFF sampled incorrect minima.
  • MM-FF Augmentation: Refitting GAFF parameters or using ML-based functional forms (GAFF-Morse, GAFF-NN) improved accuracy, with GAFF-NN showing promising results even in low-data regimes. A specific ML-based term for the n → π* interaction in aspirin significantly improved GAFF's sampling behavior.
  • Cross-Comparison: Training sGDML models on GAFF-sampled data and validating on DFT data led to higher errors than training and validating on DFT data, highlighting the importance of the underlying PES for sampling. However, sGDML models trained on GAFF data still showed reasonable performance, suggesting potential for adaptive sampling strategies.

Limitations and Future Work

  • Limitations of MM-FFs: The paper clearly demonstrates that MM-FFs, due to their fixed functional forms and reliance on simplified physical approximations, struggle to accurately represent complex quantum mechanical phenomena. This limits their predictive power for certain chemical systems and properties.
  • Computational Cost: While ML-FFs offer significant speedups over ab initio methods, they are still computationally more expensive than MM-FFs, limiting their application to very large systems.
  • Data Generation: The reliance on high-level quantum chemistry data for training ML-FFs remains a bottleneck. The paper explores using MM-FFs for data generation, but the mismatch in PESs can still lead to generalization issues.
  • Future Work: The authors suggest several promising directions:
  • Developing hybrid MM/ML force fields where ML terms are localized to capture specific quantum effects.
  • Investigating the use of ML for improving the treatment of long-range interactions, which often involve many-body effects.
  • Extending ML-FFs to larger systems like proteins and solutions, where generating reference data is particularly challenging.
  • Further exploring the synergistic potential of MM-FFs and ML, such as using ML to identify missing terms in MM-FFs and then incorporating them.

Summary updated

Abstract

Modern machine learning force fields (ML-FF) are able to yield energy and force predictions at the accuracy of high-level ab initio methods, but at a much lower computational cost. On the other hand, classical molecular mechanics force fields (MM-FF) employ fixed functional forms and tend to be less accurate, but considerably faster and transferable between molecules of the same class. In this work, we investigate how both approaches can complement each other. We contrast the ability of ML-FF for reconstructing dynamic and thermodynamic observables to MM-FFs in order to gain a qualitative understanding of the differences between the two approaches. This analysis enables us to modify the generalized AMBER force field by reparametrizing short-range and bonded interactions with more expressive terms to make them more accurate, without sacrificing the key properties that make MM-FFs so successful.
BibTeX
@article{sauceda2020molecular,
  author = {Huziel E. Sauceda and Michael Gastegger and Stefan Chmiela and Klaus‐Robert Müller and Alexandre Tkatchenko},
  title = {{Molecular force fields with gradient-domain machine learning (GDML): Comparison and synergies with classical force fields}},
  journal = {The Journal of Chemical Physics},
  year = {2020},
  doi = {10.1063/5.0023005},
}
Figure