Implementing a neural network interatomic model with performance portability for emerging exascale architectures

“…For example, HDNNP (Behler and Parrinello 2007;Behler 2017), which is a three-layered fully-connected NN is introduced in RuNNer (Behler 2011). Other packages such as BIM-NN (Yao et al 2017), Simple-NN (Lee et al 2019), CabanaMD-NNP (Desai, Reeve, and Belak 2022), SPONGE (Huang et al 2022), DeepMDkit (Wang et al 2018) are also implemented via fullyconnected NN. Many physical phenomena such as a diagram of water and Cu 2 S (Singraber et al 2019), chemical molecule (Yao et al 2017) Recent progress of NNMD packages implemented with graph NN (GNN) is gaining momentum.…”

Section: Related Workmentioning

confidence: 99%

RLEKF: An Optimizer for Deep Potential with Ab Initio Accuracy

Zhang

Qiuchen

et al. 2023

AAAI

View full text Add to dashboard Cite

It is imperative to accelerate the training of neural network force field such as Deep Potential, which usually requires thousands of images based on first-principles calculation and a couple of days to generate an accurate potential energy surface. To this end, we propose a novel optimizer named reorganized layer extended Kalman filtering (RLEKF), an optimized version of global extended Kalman filtering (GEKF) with a strategy of splitting big and gathering small layers to overcome the O(N^2) computational cost of GEKF. This strategy provides an approximation of the dense weights error covariance matrix with a sparse diagonal block matrix for GEKF. We implement both RLEKF and the baseline Adam in our alphaDynamics package and numerical experiments are performed on 13 unbiased datasets. Overall, RLEKF converges faster with slightly better accuracy. For example, a test on a typical system, bulk copper, shows that RLEKF converges faster by both the number of training epochs (x11.67) and wall-clock time (x1.19). Besides, we theoretically prove that the updates of weights converge and thus are against the gradient exploding problem. Experimental results verify that RLEKF is not sensitive to the initialization of weights. The RLEKF sheds light on other AI-for-science applications where training a large neural network (with tons of thousands parameters) is a bottleneck.

show abstract

“…NNPs are also amenable to parallel and GPU computing, so simulations with thousands or millions of atoms are possible. 50…”

Section: Performancementioning

confidence: 99%

A Neural Network Potential with Rigorous Treatment of Long-Range Dispersion

Phuc

Rezajooei

Johnson

et al. 2022

Preprint

View full text Add to dashboard Cite

Neural Network Potentials (NNPs) have quickly emerged as powerful computational methods for modeling large chemical systems with the accuracy of quantum mechanical methods but at a much smaller computational cost. To make the training and evaluation of the underlying neural networks practical, these methods commonly cut off interatomic interactions at a modest range (e.g., 5.2 Å), so longer-range interactions like London dispersion are neglected. This limits the accuracy of these models for intermolecular interactions. In this work, we develop a new NNP designed for modeling chemical systems where dispersion is an essential component. This new NNP is extended to treat dispersion interactions rigorously by calculating atomic dispersion coefficients through a second set of NNs, which is trained to reproduce the coefficients from the quantum-mechanically derived exchange-hole dipole moment (XDM) model. The NNP with this dispersion correction predicts intermolecular interactions in very good agreement with the QM data, with a mean absolute error (MAE) of 0.67 kcal/mol and a coefficient of determination (R2) of 0.97. The dispersion components of these intermolecular interactions are predicted in excellent agreement with the QM data, with a mean absolute error (MAE) of 0.01 kcal/mol and an R2 of 1.00. This combined dispersion-corrected NNP, called ANIPBE0-MLXDM, predicts intermolecular interaction energies for complexes from the DES370K test set with an MAE of 0.69 kcal/mol and an R2 of 0.97 relative to high-level ab initio results (CCSD(T)), but with a computational cost that is billions of times smaller. The ANIPBE0-MLXDM method is effective for simulating large-scale dispersion-driven systems, such as molecular liquids and gas adsorption in porous materials, on a single computer workstation.

show abstract

Implementing a neural network interatomic model with performance portability for emerging exascale architectures

Cited by 13 publications

References 64 publications

A neural network potential with rigorous treatment of long-range dispersion

A neural network potential with rigorous treatment of long-range dispersion

RLEKF: An Optimizer for Deep Potential with Ab Initio Accuracy

A Neural Network Potential with Rigorous Treatment of Long-Range Dispersion

Contact Info

Product

Resources

About