Skip to content

[Feature Request] Last layer feature (latent feature) output in deepmd-kit function or CLI #5396

@QuantumMisaka

Description

@QuantumMisaka

Summary

In DeePMD-kit we have eval-desc CLI based on eval_descriptor function to generate structural features from all the DeepPot models for structural distinction. However, for single-model uncentainty quantification, last layer feature is needed for UQ models to map model latent feature to uncentainty quantification. Practical examples are Gaussian Mixture Model [1] and Shallow Ensemble methods [2]. For integrating modern single-model UQs in deepmd-kit, especially DPA LAM, the generation of last layer feature is needed.

Ref:
[1] Zhu, A.; Batzner, S.; Musaelian, A.; Kozinsky, B. Fast Uncertainty Estimates in Deep Learning Interatomic Potentials. J. Chem. Phys. 2023, 158, 164111.
[2] Kellner, M.; Ceriotti, M. Uncertainty Quantification by Direct Propagation of Shallow Ensembles. Mach. Learn.: Sci. Technol. 2024, 5 (3), 035006. https://doi.org/10.1088/2632-2153/ad594a.

Detailed Description

  1. Python functions for calculate last layer feature for all types of deepmd-kit models (especially DPA3)
  2. CLI interface like dp eval-latent to dump last layer feature like dp eval-desc

Further Information, Files, and Links

No response

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions