You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Accelerate fp calculation by splitting it into two steps Detailed Description
fp calculations in dpgen usually consumes most resources during the whole training process. Dpgen requires many static calculations which need to starts from a bad initial wavefunction. The minimization from the initial wavefunction will consume a lot of time in the traininig process. The performance can be accelerated by several times if we minimize the wavefunction to an accuracy of meV quickly, pass the WAVECAR, and do a second minimization accurately. In the first minimization, the LREAL, PREC, and, ENCUT can be tuned to accelerate the performance. The improvement is a typical feature for traininig the potential since the usual applications of fp, e.g., structural minimization and aimd only perform one minimization from the initial wavefunction at the first ionic step. Therefore, this is not a feature request for VASP but for DPGEN.
Further Information, Files, and Links
Attached is a performance test of a CeZrO system. The DFT calculation with rare earth element like Ce will greatly slow down the fp calculation. By splitting the fp calculation into two steps, the performance can be accelerated by 6 times, which I think is an great improvement for the training process. Performance_test_zjy.pdf
The text was updated successfully, but these errors were encountered:
For now, this can be realized by modifying the command tag in fp in the machine.json. At least, it would be benificial if the user can be aware of this potential improvement from the manual.
Summary
Accelerate fp calculation by splitting it into two steps
Detailed Description
fp calculations in dpgen usually consumes most resources during the whole training process. Dpgen requires many static calculations which need to starts from a bad initial wavefunction. The minimization from the initial wavefunction will consume a lot of time in the traininig process. The performance can be accelerated by several times if we minimize the wavefunction to an accuracy of meV quickly, pass the WAVECAR, and do a second minimization accurately. In the first minimization, the LREAL, PREC, and, ENCUT can be tuned to accelerate the performance. The improvement is a typical feature for traininig the potential since the usual applications of fp, e.g., structural minimization and aimd only perform one minimization from the initial wavefunction at the first ionic step. Therefore, this is not a feature request for VASP but for DPGEN.
Further Information, Files, and Links
Attached is a performance test of a CeZrO system. The DFT calculation with rare earth element like Ce will greatly slow down the fp calculation. By splitting the fp calculation into two steps, the performance can be accelerated by 6 times, which I think is an great improvement for the training process.
Performance_test_zjy.pdf
The text was updated successfully, but these errors were encountered: