How to lower the RESIDUAL and fix the problem with α selection

[Ztao0210]

dear developer,

During my calculation of anharmonic force constants using ALAMODE combined with AIMD, I encountered an issue of excessively high RESIDUAL values. I attempted to extend the AIMD simulation steps: the number of sampling snapshots was increased from the original 80 to 160, with a sampling interval of 50 fs and random displacement amplitude set to 0.04 Å. This adjustment reduced the RESIDUAL from 62.2275% to 40.5289%. I further expanded the cutoff radius to 30 Bohr and included fifth-order anharmonic force constants in the fitting.

However,` the cutoff radius of 30 Bohr now exceeds the lattice length of the supercell (24.216819646 Bohr). I am unsure whether this setup is physically valid.

In` addition, during the CV cross-validation stage, repeated calculations consistently yielded the minimum CVSCORE at alpha = 1.58489e-05. In theory, this alpha value should produce the minimal subsequent RESIDUAL, yet the actual result contradicts this expectation. After obtaining the optimal alpha of 1.58489e-05 via CV cross-validation, I manually modified the corresponding parameter in the opt.in file to 1.58489e-06; the resulting RESIDUAL dropped to 24.071%. This inconsistency confuses me greatly.

Could` you advise on the subsequent workflow to obtain physically reliable anharmonic force constants?

I am also attaching input files and output files below.

1opt.log

2opt.log

3opt.log
Best regards,
Tao

[ttadano]

The large fitting error can be due to some of the following reasons.

• The cutoff radii too small. Clue: the fitting error decreases dramatically after increasing the cutoff radii.
• The material is strongly anharmonic or/and the AIMD temperature is too high. Clue: the fitting error decreases dramatically by increasing NORDER
• The accuracy of the FC2 or DFT forces are not very good. If none of the above two reduces the fitting error dramatically, please review if the accuracy of the DFT forces are good enough (Pseudopotential, Cutoff energy, etc.)

BTW, if you change the Taylor expansion model (NORDER, NBODY, cutoff), you will need to redo a cross-validation to get an optimal alpha. I've found that you used the same L1_ALPHA = 2.51189e-05 for 1opt.log and 2opt.log.

Also, the fitting errors shown by the "RESIDUAL (%):" section in the log file correspond to the training error, while the optimal L1_ALPHA gives the smallest validation score.