Machine Learning Interatomic Potentials and Long-Range Physics

Dylan M. Anstine, Olexandr IsayevThe Journal of Physical Chemistry A 2023 · DOI

TL;DR

Machine learning interatomic potentials (MLIPs) offer a way to simulate molecular systems with near-quantum-mechanical accuracy at a fraction of the computational cost. However, many existing MLIPs struggle to accurately capture long-range interactions like dispersion and electrostatics, limiting their applicability. This paper reviews current strategies and models for developing MLIPs that effectively incorporate these crucial long-range physics, paving the way for more accurate simulations of complex systems.

What Problem Does This Paper Address?

Traditional quantum mechanics (QM) methods, while highly accurate, are computationally expensive, limiting their use for large systems or extensive simulations. Machine learning interatomic potentials (MLIPs) have emerged as a powerful alternative, learning the relationship between atomic configurations and energies/forces from QM data. The challenge, however, is that most MLIPs are inherently "short-sighted," relying on local atomic environments. This limitation becomes critical when simulating systems where long-range interactions—such as van der Waals dispersion forces and electrostatic interactions—play a significant role in determining properties like phase behavior, molecular conformation, and material stability. Accurately modeling these phenomena requires MLIPs that can go beyond local atomic descriptors.

Methods

The paper categorizes MLIPs based on how they handle long-range interactions:

  • Augmenting with Dispersion Corrections: This involves adding explicit terms for dispersion interactions to MLIPs, often using established functional forms like the D3 correction. The key distinction lies in whether these corrections are applied "on-the-fly" during inference or learned implicitly by the MLIP during training.
  • Electrostatics via Charge Prediction: This approach focuses on accurately predicting atomic charges, which are then used to calculate electrostatic interactions. Methods include:
  • Atomic Environment Descriptors: Training ML models to predict charges based on local atomic environments.
  • Self-Consistency and Message Passing: Iterative refinement schemes where charges (or related properties) are updated until a consistent state is reached. Message passing allows information to propagate across the system, enabling the capture of more nonlocal effects.
  • Charge Equilibration (Qeq) Schemes: Solving a system of linear equations to determine a global charge distribution that minimizes electrostatic energy, often incorporating atomic electronegativity and hardness.
  • Models Incorporating Polarization: This advanced strategy aims to capture the dynamic response of electron densities to electric fields, which is crucial for polarizable systems. This often involves more complex representations of atomic interactions beyond simple point charges.

Datasets

The paper does not explicitly list specific datasets used for training or evaluation. However, it frequently refers to the use of density functional theory (DFT) calculations as the source of reference data for training MLIPs. It mentions that generalized gradient approximation (GGA) functionals are often used, and that dispersion corrections (like DFT-D3) are frequently applied to these DFT calculations to generate training data. The paper also notes the importance of experimental ground truths for benchmarking, but acknowledges their limited availability for noncovalent interactions.

Key Insights and Ideas

  • The "Nearsightedness" Problem: The core challenge is that MLIPs trained on local atomic environments inherently miss long-range physics, leading to inaccurate predictions for many real-world systems.
  • Dispersion is Not Always Negligible: While short-range interactions dominate many systems, the collective effect of dispersion forces can be substantial in larger systems, impacting properties like cohesive energy and material stability.
  • Charge Prediction is Tricky: Mapping complex electron densities to point charges is an ambiguous task with no single correct solution, leading to various schemes (e.g., Hirshfeld, Löwdin) that can affect MLIP transferability.
  • Iterative Refinement is Powerful: Self-consistency and message-passing schemes offer a promising avenue for propagating nonlocal information, allowing MLIPs to learn more flexible and generalizable chemical representations.
  • Global vs. Local Approaches: Charge equilibration methods offer a global perspective on charge distribution, which can be advantageous for capturing nonlocal charge transfer, contrasting with purely local ML approaches.
  • The Need for Physics-Informed Design: The choice of MLIP strategy should be guided by the specific physical phenomena governing the system of interest, rather than a one-size-fits-all approach.

Key Results

The paper is a perspective, focusing on reviewing existing methodologies and highlighting their potential rather than presenting new quantitative results. However, it points to several successful applications:

  • Accurate Short-Range Models: MLIPs can achieve near ab initio accuracy for short-range interactions at significantly reduced computational cost.
  • Successful Dispersion Corrections: Models augmented with dispersion corrections have shown promise in accurately describing systems like condensed-phase phosphorus and multilayer graphene.
  • Effective Charge Prediction: Models like PhysNet, HIP-NN, and Behler-Parrinello variants have demonstrated the ability to predict conformation-dependent charges.
  • Message-Passing Success: Models like AIMNet have shown applicability to neutral organic molecules and can predict properties like atomic partial charges with DFT accuracy. AIMNet-NSE has further demonstrated the ability to model spin states.
  • Qeq Applications: Charge equilibration methods have been successfully integrated into MLIPs for systems like ionic crystals and organic molecules.
  • Polarization Challenges: While progress is being made, accurately capturing polarization effects in MLIPs remains an active area of research.

Limitations and Future Work

  • Data Bottleneck: The accumulation of sufficient and diverse reference data for training remains a significant bottleneck for increasing MLIP capabilities. Active learning strategies are suggested to address this.
  • Generalizability: While many models show promise, their generalizability to a wide range of chemical systems and phenomena still requires further validation.
  • Integration with Simulation Software: Efficiently integrating advanced MLIPs, especially those with complex nonlocal features, into existing molecular simulation software is an ongoing challenge.
  • Experimental Benchmarks: There is a need for more experimental ground truths to rigorously benchmark MLIP performance, particularly for noncovalent interactions.
  • Applied MLIP Modeling: The field of applied MLIP modeling, where these potentials are used to solve real-world chemical and materials science challenges, lags behind model development.
  • Future Directions: The authors encourage further development in MLIPs that can accurately capture polarization, many-body effects, and complex chemical reactions, and envision their use in simulating large-scale systems with intricate short- and long-range behaviors.

Summary updated

Abstract

Advances in machine learned interatomic potentials (MLIPs), such as those using neural networks, have resulted in short-range models that can infer interaction energies with near ab initio accuracy and orders of magnitude reduced computational cost. For many atom systems, including macromolecules, biomolecules, and condensed matter, model accuracy can become reliant on the description of short- and long-range physical interactions. The latter terms can be difficult to incorporate into an MLIP framework. Recent research has produced numerous models with considerations for nonlocal electrostatic and dispersion interactions, leading to a large range of applications that can be addressed using MLIPs. In light of this, we present a Perspective focused on key methodologies and models being used where the presence of nonlocal physics and chemistry are crucial for describing system properties. The strategies covered include MLIPs augmented with dispersion corrections, electrostatics calculated with charges predicted from atomic environment descriptors, the use of self-consistency and message passing iterations to propagated nonlocal system information, and charges obtained via equilibration schemes. We aim to provide a pointed discussion to support the development of machine learning-based interatomic potentials for systems where contributions from only nearsighted terms are deficient.
BibTeX
@article{anstine2023machine,
  author = {Dylan M. Anstine and Olexandr Isayev},
  title = {{Machine Learning Interatomic Potentials and Long-Range Physics}},
  journal = {The Journal of Physical Chemistry A},
  year = {2023},
  doi = {10.1021/acs.jpca.2c06778},
}
Figure