Long range Deepmd

[chtchelkatchev]

A significant limitation of current DeepMD machine learning potentials (MLPs) is their inherent locality and inability to naturally predict atomic charges. This presents a major obstacle for modeling processes and systems where long-range electrostatic interactions are critical, such as chemical reactions in solution or at interfaces. Furthermore, the lack of explicit charges prevents the direct calculation of material properties like dielectric permittivity.

While training DeepMD using Wannier centers offers a partial solution, this approach requires considerable care and significant computational resources. A more elegant and fundamental solution is needed: the ability to train atomic charges by directly minimizing the loss function. It is unclear why this capability, which has been a standard feature in many other MLP frameworks for some time, has not yet been implemented in DeepMD.

Another desirable feature is the support for hybrid potentials. This would allow a user-defined analytical potential (with tunable parameters) to be combined with the neural network's local potential. During the training procedure, both the NN weights and the analytical parameters would be optimized simultaneously. These parameters could potentially depend on both the local atomic environment and the global chemical composition of the system, offering a powerful and flexible framework for complex simulations.

[wanghan-iapcm]

The partial-charge approach is not preferred because it does not guarantee convergence when modeling long-range interactions. In contrast, the DPLR method does provide such convergence.

A possible workaround is to use atomic energies—that is, to supply the partial charges as atomic energy labels. For further discussion, see #5031

[chtchelkatchev]

Thank you for your prompt response. However, I must respectfully disagree. Several years ago, the well-known review by Behler presented numerous successful examples (including his own machine-learning potential) where the challenge of charge convergence was effectively solved. Behler introduced the term "Fourth-Generation Machine Learning Potential" for such models, a classification that has since become standard.

New fourth-generation potentials continue to be published annually; for example, here is a recent one: https://arxiv.org/html/2502.07907v1. There is also a fresh review that cites many fourth-generation ML potentials: https://doi.org/10.1016/j.cossms.2025.101214. A fourth-generation MTP was also just released (I included a photo of the poster above).

The defining characteristic of these fourth-generation potentials is the automatic charge assignment; no Wannier functions (or atomic energies) are needed. Therefore, in this regard, the DEEPMD-kit has fallen significantly behind.

[wanghan-iapcm]

The fourth-generation method may offer stronger generalizability than other long-range treatment approaches (although I do not have any direct comparisons). However, it fundamentally requires an SCF procedure for assigning partial charges, which leads to cubic computational scaling—recently reduced to quadratic scaling in [arXiv:2502.07907]. In contrast, DPLR scales as (N \log N) (PPPM scaling).

I do not find your style of discussion—simply asserting that one method “goes beyond” others without a detailed analysis of their respective advantages and limitations—to be constructive. Such an approach does not provide meaningful value to the community.

I have therefore decided to close this discussion thread.