Show Posts
1
General Questions and Answers / Questions about mace trainingset
« on: December 23, 2025, 18:57 »
Hi QuantumATK Support Team,
I’m following the QuantumATK MACE machine-learning tutorial, and I have a few questions about how the TrainingSet(s) used in the tutorial were prepared.
How was the “tutorial TrainingSet” generated?
The tutorial provides a pre-made TrainingSet file. From what I can tell, it looks like crystal random displacements (CRD) may have been used to generate structures.
However, I’m not sure whether the dataset was created via the MTP tutorial-style workflow (e.g., Step 1: “MomentTensorPotential training” / “reference calculator” to label energies/forces/stresses), or if a different workflow was used.
Also, was the TrainingSet created with “Recalculate training data = True” (i.e., structures generated first and then re-labeled with a specified reference calculator), or were the labels produced directly at the time the structures were generated?
Could you clarify the exact workflow and tools used to produce the provided TrainingSet?
How was the amorphous-only TrainingSet produced (and with what DFT settings)?
I noticed another TrainingSet that seems to contain only amorphous structures. Could you share how those amorphous configurations were generated?
Were they obtained from AIMD trajectories (e.g., melt–quench), or were they produced by geometry optimization from some initial amorphous guess (or another sampling method)?
For the DFT labeling step: what k-point sampling was used (Γ-only vs. Monkhorst–Pack, and the specific mesh if applicable)?
If AIMD was used: were the AIMD steps also run at Γ-only and then re-labeled later with a denser k-mesh, or were they labeled with the final k-point settings from the beginning?
Any details you can provide (e.g., whether stresses were included, exact k-point mesh, basis set / pseudopotentials, XC functional, and whether re-labeling was performed) would be very helpful so I can reproduce the tutorial dataset preparation reliably.
Thank you
I’m following the QuantumATK MACE machine-learning tutorial, and I have a few questions about how the TrainingSet(s) used in the tutorial were prepared.
How was the “tutorial TrainingSet” generated?
The tutorial provides a pre-made TrainingSet file. From what I can tell, it looks like crystal random displacements (CRD) may have been used to generate structures.
However, I’m not sure whether the dataset was created via the MTP tutorial-style workflow (e.g., Step 1: “MomentTensorPotential training” / “reference calculator” to label energies/forces/stresses), or if a different workflow was used.
Also, was the TrainingSet created with “Recalculate training data = True” (i.e., structures generated first and then re-labeled with a specified reference calculator), or were the labels produced directly at the time the structures were generated?
Could you clarify the exact workflow and tools used to produce the provided TrainingSet?
How was the amorphous-only TrainingSet produced (and with what DFT settings)?
I noticed another TrainingSet that seems to contain only amorphous structures. Could you share how those amorphous configurations were generated?
Were they obtained from AIMD trajectories (e.g., melt–quench), or were they produced by geometry optimization from some initial amorphous guess (or another sampling method)?
For the DFT labeling step: what k-point sampling was used (Γ-only vs. Monkhorst–Pack, and the specific mesh if applicable)?
If AIMD was used: were the AIMD steps also run at Γ-only and then re-labeled later with a denser k-mesh, or were they labeled with the final k-point settings from the beginning?
Any details you can provide (e.g., whether stresses were included, exact k-point mesh, basis set / pseudopotentials, XC functional, and whether re-labeling was performed) would be very helpful so I can reproduce the tutorial dataset preparation reliably.
Thank you