[Feature Request] Release a universal pretrained model with frame-parameter support for electronic temperature #6117

Summary

Could you consider training and releasing a universal pretrained model whose energy fitting network supports a one-dimensional frame parameter (numb_fparam=1)? This would provide a starting point for fine-tuning models for applications involving electronic temperature or other frame-level parameters.

Detailed Description

Some applications require frame parameters, particularly electronic temperature. Training a model from scratch for these applications can be computationally expensive. However, the pretrained models I have used do not include fparam inputs in their fitting networks, which makes direct fine-tuning for these applications difficult.

I would therefore like to request a universal pretrained model with numb_fparam=1 included in its fitting network from the pretraining stage.

For existing datasets such as OMat24 or MatPES, one possible approach would be to assign an appropriate baseline frame-parameter value based on the original DFT settings. If this parameter represents electronic temperature, its value should be determined from the smearing method and temperature used in the reference calculations.

Once the pretrained model includes this input, users could fine-tune it using their own datasets containing different electronic temperatures, rather than training a universal model from scratch. I understand that pretraining at a single temperature would not by itself establish temperature dependence; that dependence would need to be learned from temperature-dependent fine-tuning data.

Would it be feasible to provide such a pretrained model in a future release?

I have also tried training a DPA4C-plus foundation model from scratch on the MatPES dataset, with fparam set to a constant value of 300 K for every frame. However, the results were not as good as I had hoped. For example, the dimer energy curves were less smooth than those produced by the pretrained models available on AI Square. This experience further motivated my request for an officially pretrained model with built-in frame-parameter support.

Further Information, Files, and Links

No response