Symmetry-adapted graph neural networks for constructing molecular dynamics force fields

Abstract

Molecular dynamics is a powerful simulation tool to explore material properties. Most realistic material systems are too large to be simulated using first-principles molecular dynamics. Classical molecular dynamics has a lower computational cost but requires accurate force fields to achieve chemical accuracy. In this work, we develop a symmetry-adapted graph neural network framework called the molecular dynamics graph neural network (MDGNN) to construct force fields automatically for molecular dynamics simulations for both molecules and crystals. This architecture consistently preserves translation, rotation, and permutation invariance in the simulations. We also propose a new feature engineering method that includes high-order terms of interatomic distances and demonstrate that the MDGNN accurately reproduces the results of both classical and first-principles molecular dynamics. In addition, we demonstrate that force fields constructed by the proposed model have good transferability. The MDGNN is thus an efficient and promising option for performing molecular dynamics simulations of large-scale systems with high accuracy.

Graph neural networks for materials science and chemistry

“…Analogous to atomistic dynamics simulations of molecular systems, GNNs can also be used to simulate the dynamic behavior of crystals, i.e., to predict potential energy, forces, and partial charges, in order to drive molecular dynamics simulations. To predict forces for molecular dynamics of crystals, most approaches use conventional machine learning methods 252 , but there are first examples that use GNNs 190,253 . Raza et al use a GNN with node-level readouts to predict partial charges of MOFs for molecular simulations of gas adsorption 254 .…”

Applications of Artificial Intelligence and Machine Learning Algorithms to Crystallization

“…A promising application of ML is to learn the crystal lattice energies from crystal structures, partially substituting expensive DFT calculations at 1/5 to 1/10000 of the computational cost often via ANNs with symmetry functions based on the work of Behler and Parrinello. ,− For instance, a deep neural network (DNN) trained on DFT data successfully predicted total energies for organic molecules (including small drug substances), speeding up the calculations while performing significantly better than semiempirical QM methods (such as DFTB and PM6) . Recently, a molecular dynamics graph neural network was developed by Wang et al which includes high-order terms of interatomic distances and reproduces the force fields of molecules and crystals resulting from both classical and first-principles molecular dynamics, also imposing additional constraints to preserve translation and rotation invariance. Yao, Herr, and Parkhill used an ANN to reproduce the accuracy of ab initio many-body expansion calculations without specialized force fields while reducing the computational overhead by a factor of 10 6 .…”

“…Graph theory representing molecules can be powerful in this approach. 255,256 Machine learning for CSP should be demonstrated for drug molecules of higher molecular mass in more intricate chemical systems increasingly present in the current pharmaceutical development pipelines, as the emerging methodologies focus on inorganic materials 247,248 and APIs of modest chemical complexity, similar to aspirin and salicylic acid, 241,243 olanzapine, 53 fentanyl, and lisdexamfetamine. 240 Moreover, although machine learning formation energies can be learned with high accuracy from DFT calculations, the relative stability may not be as accurate because DFT benefits from a systematic error cancellation while machine learning does not.…”

Scalars are universal: Equivariant machine learning, structured like classical physics

“…There is significant work on imposing permutation symmetry in jet assignment for high-energy particle collider experiments with self-attention networks [24,46]. In molecular dynamics, rotationally invariant neural networks have been shown to better learn molecular properties [2,78,71], and Hamiltonian neural networks have been constructed to better preserve molecular conformations [48]. More broadly, Hamiltonian networks have been shown to improve physical characteristics, such as better conservation of energy, and to better generalize [32,68,86], and Lagrangian neural networks can also enforce conservation laws [53,17].…”