Are the dpa3 omol models only trained on omol?

[JonathanSchmidt1]

We noticed that the dpa3 omol models return values for effectively the whole periodic table while omol only includes a limited number of elements. So I was wondering if the outputs for non-omol elements are basically random or are they trained on additional data.
Thank you so much!

[dosubot[bot]]

Hi @JonathanSchmidt1! I'm Dosu and I’m helping the deepmd-kit team.

The DPA3-Omol-Large model is trained exclusively on the OMol25 dataset — no additional datasets are mixed in [1].

OMol25 covers 83 elements (Z = 1–83, hydrogen through bismuth) [2]. All radioactive elements after bismuth (Po, At, Rn, Fr, Ra, and all actinides/transactinides) are explicitly absent from the training data [2].

So the model does have trained coverage for a large portion of the periodic table — 83 elements is quite broad — but for elements Z ≥ 84, any outputs would be untrained extrapolations with no validated accuracy. The DPA-3 architecture includes an electronic configuration type embedding (econf_embd.py) that defines representations for all 118 elements, which is likely why the model accepts those elements as input rather than erroring out. But accepting input is not the same as producing meaningful predictions — for elements outside OMol25's coverage, the results would essentially be extrapolations based on electronic configuration similarity to trained elements, not genuinely trained predictions.

In short: if you're working with elements 1–83, the model has legitimate training data behind it. For anything beyond that, treat the outputs with heavy skepticism.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}

  

[JonathanSchmidt1]

Thank you so much for the quick reply. That's about what I expected.