Accurate global machine learning force fields for molecules with hundreds of atoms

Abstract

Global machine learning force fields, with the capacity to capture collective interactions in molecular systems, now scale up to a few dozen atoms due to considerable growth of model complexity with system size. For larger molecules, locality assumptions are introduced, with the consequence that nonlocal interactions are not described. Here, we develop an exact iterative approach to train global symmetric gradient domain machine learning (sGDML) force fields (FFs) for several hundred atoms, without resorting to any potentially uncontrolled approximations. All atomic degrees of freedom remain correlated in the global sGDML FF, allowing the accurate description of complex molecules and materials that present phenomena with far-reaching characteristic correlation lengths. We assess the accuracy and efficiency of sGDML on a newly developed MD22 benchmark dataset containing molecules from 42 to 370 atoms. The robustness of our approach is demonstrated in nanosecond path-integral molecular dynamics simulations for supramolecular complexes in the MD22 dataset.

Denoise Pre-training on Non-equilibrium Molecules for Accurate and Transferable Neural Potentials

“…Unlike the previous datasets which only investigate small organic molecules, MD22 74 and SPICE 73 cover a wider variety of molecule types. In particular, MD22 includes the AIMD trajectories of proteins, carbohydrates, nucleic acids, and supramolecules (i.e., buckyball catcher and nanotube).…”

“…SPICE,73 and MD2274 ), which cover different molecular systems from the pre-training datasets.Figure 2a and 2b compare the performance of GNN models with and without denoise pre-training on ISO17 containing isomers of C 7 O 2 H 10 . It is demonstrated that pre-trained GNN models achieve better prediction accuracy when evaluated by both RMSE and MAE.…”

Reconstructing Kernel-Based Machine Learning Force Fields with Superlinear Convergence

“…To retain the original PSD structure, we require that a symmetric decomposition P = LL T ( L ∈ double-struckR n × n ) exists. ,, Another key requirement is that the construction of the preconditioner P should not dominate the overall computational costs. Generally, there are two principled ways to achieve this: either via sparsification of the kernel matrix (e.g., zero-out entries to obtain a block-diagonal form) or via a compression into a low-rank representation.…”

“…In this setting, both neural networks − and kernel-based approaches − have been successfully applied. Kernel machines are generally considered to be more data efficient in modeling high-quality MLFFs, but yield large linear optimization problems when many training samples and/or large molecule sizes are involved (e.g., in materials ,, or biomolecules). Linear systems are generally solved in closed-form, which has quadratic memory complexity.…”

q-pac: A Python Package for Machine Learned Charge Equilibration Models

“…These for example include the use of global or non-local descriptors. [18][19][20] In many cases, physical baseline models can also provide the correct long-range physics at affordable computational cost (∆-ML). 16,17,[21][22][23] Finally, message-passing neural networks can extend the range of local interatomic potentials by a multiple of the employed cutoff, though without including the full long-range interactions present in a periodic system.…”