How to choose fidelity(?) of DPA3-v2-OpenLAM model? #4695
-
|
In DPA3-v2-OpenLAM model, it says it has integrated multiple DFT datasets into a single and the potential is trained on it. If so, as different DFT datasets have different settings, some techniques exist for handling this. Can I have some docs relevant to this? If it have accomplished it through something similar to SevenNet, which can choose 'fieldity' in inference mode, how to do the same in DPA3? Related 7net reference: https://pubs.acs.org/doi/full/10.1021/jacs.4c14455 |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment
-
|
DPA3-v2-OpenLAM adopts the same multi-task scheme as the DPA-2 model. Different tasks train the model using corresponding datasets labeled with varying levels of DFT (Density Functional Theory). You may select any of the available task heads following the instructions in this link. Note: Even if two datasets are labeled with the same XC functional, they are typically not combined and are instead trained using different task heads. This is because their DFT labeling hyperparameters—such as energy cutoff and k-spacing—differ. You may check the labeling hyperparameters by searching the head name on https://www.aissquare.com/ Note2: to use dpa3-v2-openlam you need to install v3.1.0a |
Beta Was this translation helpful? Give feedback.
DPA3-v2-OpenLAM adopts the same multi-task scheme as the DPA-2 model. Different tasks train the model using corresponding datasets labeled with varying levels of DFT (Density Functional Theory). You may select any of the available task heads following the instructions in this link.
Note: Even if two datasets are labeled with the same XC functional, they are typically not combined and are instead trained using different task heads. This is because their DFT labeling hyperparameters—such as energy cutoff and k-spacing—differ. You may check the labeling hyperparameters by searching the head name on https://www.aissquare.com/
Note2: to use dpa3-v2-openlam you need to install v3.1.0a