Learning Machine-Learned Interatomic Potentials with DeePMD: A Methane Example

Simulating how atoms in a system move and interact with time or Molecular Dynamics (MD) has applications in physics, chemistry and related fields. For a long time, Ab initio molecular dynamics (AIMD) was used, which had the accuracy of Density Functional Theory (DFT), but the computational cost limited its use case to simpler systems and smaller time periods. In the recent years, huge development has been taking place in the field of Machine Learning methodologies to reproduce original data, by training the machine on large data sets of atomic configurations and corresponding potential energies and forces.

One of the software which does the same is DeePMD. Here, the total energy of a system is constructed as a sum of atomic energies. To do this, the structure is first transformed into a set of descriptors which serve as a framework for establishing a mathematical relationship between structure and energy. We take translational, rotational and permutational invariance for granted and try to create the most complete picture of the atom using its surroundings. As we are working with Neural Nets for the “learning”, the descriptors are made smooth. Learning process is impacted by the choice of descriptor but can be improved thanks to DeePMD’s introduction of a flexible family of loss functions.

For this experiment, I trained models on methane data generated using ABACUS (a DFT package) and compared different descriptor choices.

Here are some of the observations:

· Training/validation losses decrease consistently.

· The three-body embedding (DeepPot-SE descriptor) produces predictions much closer to DFT energies.

· The simpler two-atom embedding shows noticeably larger scatter.

Interesting to see how descriptor design directly affects learning quality in atomistic ML. Incorporating richer local geometric information significantly improved predictive quality.

Training and validation RMSE for energies and forces during DeePMD training on methane data. Both energy and force losses decrease steadily, indicating stable learning and reasonable generalization.

Parity plot comparing DFT energies with predictions from the three-body embedding (DeepPot-SE) descriptor. The close alignment with the diagonal line indicates strong agreement with reference quantum calculations.

Parity plot for the simpler two-atom embedding descriptor. Compared to the three-body embedding, a larger scatter around the ideal diagonal line is observed, reflecting reduced predictive performance.