-
Notifications
You must be signed in to change notification settings - Fork 648
The question about in non-periodic systems about DeepMD-kit、DPGEN、DPA-3 #5190
1:When training a force field using deepmd-kit, are there any modifications or points of attention required in the parameter settings compared to the training settings I previously used for periodic systems, as shown in the attached image?
周期性体系input
2:When using dpgen, for the exploration step, I previously employed LAMMPS for molecular dynamics simulations in periodic systems. What would be suitable for non-periodic systems (given that I use Gaussian for FP calculations)?
(PS: The CH4 case I found on GitHub still seems to consider its periodic structure and uses LAMMPS for the exploration step.)
3:Are there any datasets in DPA-3 suitable for non-periodic systems?
All reactions
Replies: 2 comments
For DeepMD-kit, you don't need to change your model architecture or learning parameters for non-periodic systems. The key is to set the pbc flag to false in your training and validation data sections, and include an empty nopbc file in each system directory to indicate non-periodic boundary conditions. The box property is not required for non-periodic data. The rest of your configuration (neurons, learning rate, loss weights, etc.) can remain as you used for periodic systems, but always ensure your dataset samples local atomic environments well and passes data quality checks (no NaN/Inf/extreme values, consistent units, etc.) see docs.
For dpgen, the exploration step (model deviation detection) is still typically done with LAMMPS, but you can set model_devi_nopbc: true to run MD with open boundary conditions for non-periodic systems. Gaussian is supported for FP calculations (fp_style: gaussian), but not as an MD engine for exploration. If classical MD is not suitable for your system, you may need to use short MD trajectories, increase frame recording frequency, or consider custom scripts for exploration. There is currently no built-in workflow for using Gaussian as the exploration engine see docs.
For DPA-3, there are no datasets explicitly labeled as non-periodic, but the infrastructure (via dpdata) supports non-periodic molecular data if you format and label it correctly. You can use dpdata's LabeledSystem and related methods to convert Gaussian or other molecular data into DeepMD-kit training format see docs.
Let me know if you want example scripts or more details on any step.
To reply, just mention @dosu.
Docs are dead. Just use Dosu.
Leave Feedback Ask Dosu about deepmd-kit Share Dosu with your team
All reactions
One correction to the Dosu answer: DeePMD training input does not use a pbc: false field in the training/validation JSON. Periodicity is a property of each data system. For an NPY system, create an empty nopbc file in the system directory; DeePMD then treats the box as optional and does not use periodic images. See the non-periodic data format and examples/nopbc. Virial/stress labels are normally not meaningful or consumed for a non-periodic molecular system.
For DP-GEN, model_devi_nopbc: true is the intended switch. Its LAMMPS exploration input uses boundary f f f; Gaussian remains the FP labeling engine, not the exploration MD engine. Use NVE/NVT rather than a box-changing ensemble, and make sure an isolated cluster cannot simply evaporate or hit a fixed boundary during exploration.
For pretrained DPA-3 data, molecular/non-periodic coverage is now available through the OMol25 branch and DPA3-Omol-Large. Whether it is suitable depends on element and chemical-environment coverage; fine-tune it on your Gaussian-labeled data and validate on held-out structures relevant to the intended reactions.
Coding agent: Codex
Codex version: codex-cli 0.149.0
Model: gpt-5.6-sol
Reasoning effort: xhigh