Machine-learning-based many-body energy analysis of argon clusters: Fit for size?

Machine-learned potentials (MLPs) have gained attention to fill the gap between empirical and quantum-mechanical methods. Their speed can be comparable to empirical force fields, whilst preserving ab initio accuracy. Their prediction of systems' expected physical/chemical characteristics presents the many-body energy as a key challenge: although empirical potentials cannot compute this owing to their pairwise-additive nature, ab initio methods can do so, albeit not for large systems. Here, we examine if a finely-trained MLP can predict the many-body energy for argon clusters caS0301-0104(21)00258-5lculated by DFT methods and HDNNs for MLPtraining. The final model predicts the many-body energy with error of a few meV.

Automated summary

Machine-learned potentials (MLPs) are used to bridge the gap between empirical and quantum-mechanical methods, and face the challenge of predicting many-body energy when determining system characteristics. A finely-trained MLP model successfully predicts the many-body energy of argon clusters with an error within a few meV.