Automated Construction of Quantum–Classical Hybrid Models

Christoph Brunken, Markus ReiherJournal of Chemical Theory and Computation 2021 · DOI

TL;DR

This paper introduces QM/SFAM, a fully automated method for building quantum mechanics/molecular mechanics (QM/MM) models. It addresses the challenge of defining accurate QM regions and classical force fields by leveraging a system-focused, self-parametrizing approach, enabling more reliable and efficient simulations of complex chemical systems.

What Problem Does This Paper Address?

Simulating complex chemical systems, such as reactions in proteins or on surfaces, requires a hybrid approach that combines accurate quantum mechanical (QM) descriptions of reactive centers with efficient classical molecular mechanics (MM) models for the surrounding environment. However, traditional QM/MM methods demand significant manual effort for model construction, including defining the QM region, parametrizing the MM force field, and ensuring consistency. This manual process is prone to errors, lacks standardization, and hinders high-throughput studies. The authors aim to automate this entire process, making QM/MM modeling more accessible, reliable, and efficient.

Methods

The core of this work is the QM/SFAM approach, which extends the authors' previous Self-Parametrizing System-Focused Atomistic Models (SFAM) for the MM part. Key methodological contributions include:

  • Automated QM Region Definition: Instead of relying on chemical intuition, the QM region is determined automatically based on first principles, focusing on accurately describing atomic forces in the reaction center. This involves evaluating differently sized QM regions and using a descriptor based on atomic forces to select the most reliable models.
  • System-Focused SFAM for MM: The classical part of the model uses SFAM, which is automatically parametrized from quantum mechanical reference data. This eliminates the need for pre-existing force field parameters and ensures consistency between the QM and MM descriptions.
  • Link-Atom Approach for QM/MM Boundary: The authors employ a link-atom approach to handle the boundary between the QM and MM regions, specifically focusing on automated placement and force distribution.
  • Charge Redistribution Schemes (RC and RCD): To mitigate overpolarization effects at the QM/MM boundary in electrostatic embedding, two charge redistribution schemes are implemented.
  • Automated Structure Optimization and Molecular Exploration: The paper outlines strategies for structure optimization and molecular explorations using the QM/SFAM model, including microiteration-based optimization to reduce computational cost.
  • Local Reparametrization: A scheme is introduced to efficiently re-parametrize SFAM locally when new covalent bonds are formed or broken and moved between the QM and MM regions.

Datasets

The paper uses the following for demonstration and validation:

  • Insulin Peptide Hormone (Chain A): Used as a model system for both dry and microsolvated (with water) scenarios. Full QM calculations were performed as a reference.
  • 1-Propanol and Carboxylic Acid Group: These were added to the insulin system to create an esterification reaction for studying chemical transformations.

Key Insights and Ideas

  • Automating QM/MM is Achievable and Beneficial: The work demonstrates that a fully automated QM/MM model construction is feasible, significantly reducing the manual effort and potential for human error.
  • Forces as a Reliable Descriptor for QM Region Selection: The authors found that atomic forces, particularly in the vicinity of the reaction center, are a robust descriptor for evaluating the quality of a QM/MM model and selecting an appropriate QM region size.
  • SFAM's Versatility: The system-focused, self-parametrizing nature of SFAM makes it an ideal classical component for QM/MM, as it adapts to the specific system and QM method, avoiding parameter transferability issues.
  • Flexibility in QM Region Definition: The automated approach allows for convenient redefinition of the QM region during molecular exploration, which is crucial for studying dynamic reactive centers.
  • Controlling Uncertainty: By generating and evaluating multiple QM/MM models with varying QM region sizes, the authors can estimate and control the uncertainty in calculated properties.

Key Results

  • The QM/SFAM approach successfully automates the construction of QM/MM models.
  • The automated QM region selection process, based on atomic forces, effectively filters out unreliable models. For the dry insulin example, only four out of many generated models showed a large force error (>0.75 kcal mol⁻¹ bohr⁻¹), all with fewer than 60 QM atoms.
  • The method demonstrated good transferability of force accuracy across different points on the potential energy surface (reactant vs. product structures).
  • For the solvated insulin system, a reference set of models with large QM regions ( > 390 atoms) achieved a mean deviation of 0.05 kcal mol⁻¹ bohr⁻¹ for atomic forces.
  • The automated selection process, using a tolerance of 0.05 kcal mol⁻¹ bohr⁻¹, identified seven models with high force accuracy.
  • The reaction energy error was found to decrease with increasing QM region size, and models with large force errors generally also exhibited large energy errors.

Limitations and Future Work

  • Computational Cost of Reference Calculations: While the SFAM parametrization is efficient, generating reference data for large QM regions for the QM region selection can still be computationally intensive, although it is performed only once per model candidate.
  • Energy vs. Force Accuracy: While force accuracy is a good indicator, it doesn't perfectly guarantee energy accuracy. The authors note that for well-performing models, reaction energies can still fluctuate by a few kcal mol⁻¹.
  • Heuristic Rules for Final Selection: Currently, the final selection of a QM region involves some heuristic rules (e.g., favoring fewer link atoms, discarding highly asymmetric regions). The authors plan to extend this to a purely first-principles data-driven selection.
  • Extending Descriptors: The authors suggest that incorporating energy differences between structures into the descriptor could further improve the selection process.
  • Generalization to Other QM Methods: While the framework is designed to be general, specific QM methods might require further adaptation.

Summary updated

Abstract

We present a protocol for the fully automated construction of quantum mechanical (QM)-classical hybrid models by extending our previously reported approach on self-parametrizing system-focused atomistic models (SFAMs) [Brunken, C.; Reiher, M. <i>J. Chem. Theory Comput.</i> 2020, 16, 3, 1646-1665]. In this QM/SFAM approach, the size and composition of the QM region are evaluated in an automated manner based on first principles so that the hybrid model describes the atomic forces in the center of the QM region accurately. This entails the automated construction and evaluation of differently sized QM regions with a bearable computational overhead that needs to be paid for automated validation procedures. Applying SFAM for the classical part of the model eliminates any dependence on pre-existing parameters due to its system-focused quantum mechanically derived parametrization. Hence, QM/SFAM is capable of delivering high fidelity and complete automation. Furthermore, since SFAM parameters are generated for the whole system, our ansatz allows for a convenient redefinition of the QM region during a molecular exploration. For this purpose, a local reparametrization scheme is introduced, which efficiently generates additional classical parameters on the fly when new covalent bonds are formed (or broken) and moved to the classical region.
BibTeX
@article{brunken2021automated,
  author = {Christoph Brunken and Markus Reiher},
  title = {{Automated Construction of Quantum–Classical Hybrid Models}},
  journal = {Journal of Chemical Theory and Computation},
  year = {2021},
  doi = {10.1021/acs.jctc.1c00178},
}
Figure