DPA-2: Multi-Task Atomic Model Pre-Training and Fine-Tuning Datasets

DPA-2 is a multi-task pre-trained model for atomic modeling, trained across 18 upstream datasets for 1 million steps. The associated data archive includes both upstream pre-training and downstream fine-tuning datasets in DeePMD format, used for tasks like learning curve analysis and model distillation. The work originates from Peking University and is hosted on the paperswithcode platform.

Use Cases

• Pre-training machine learning interatomic potentials based on the 18 upstream datasets mentioned.
• Fine-tuning models for specific chemical systems based on the 15 downstream datasets referenced.
• Analyzing learning curves for model transferability as described for the downstream evaluation workflow.
• Distilling fine-tuned models for systems like H2O-PBE0TS-MD, SSE-PBE-D, and FerroEle-D as illustrated in the manuscript.

Strengths

• Includes a multi-task pre-trained model (DPA-2) trained across 18 distinct upstream datasets.
• Data is structured for both upstream pre-training and downstream fine-tuning, with dedicated train/test splits.
• Accompanied by scripts for training, fine-tuning, and distillation workflows as detailed in the description.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Last update date is unknown; freshness unverified.

Provenance

Source

paperswithcode, with datasets described in the associated research paper's Supplementary Materials and accessible on AIS Square.

Collection Method

Datasets compiled for research on the DPA-2 atomic model, likely from computational chemistry simulations.