The fitting residuals of the fourth-order force constant of Alamode are larger, so how should they be lower

[liudadie12138]

@ttadano
Thanks for your answer, so I can think that the above process for calculating the thermal conductivity after reforming is correct,
Also, I have one last question, I should do opt after fitting the fourth-order force constant, cv, this is my opt file
&general
PREFIX = ct221
MODE = optimize
NAT = 160; NKD = 3
KD = Na Nb O
/

&interaction
NORDER = 3 # 1: harmonic, 2: cubic, ..
NBODY = 2 3 3
/

&cutoff
- None 10.0 9.0
/

&optimize
LMODEL = ols
NDATA = 40
DFSET = DFSET_AIMD_random
FC2XML = NaNbO221.xml

CV = 0 # switch off CV
L1_RATIO = 1.0 # LASSO
L1_ALPHA = 4.64159e-05 # Add this line.
CV_MINALPHA = 1.0e-8
CV_MAXALPHA = 0.01
CV_NALPHA = 30

STANDARDIZE = 1
CONV_TOL = 1.0e-9 # was CONV_TOL = 1.0e-8 in CV

/

&cell
1.889726
10.9833793640 0.0000000000 0.0000000000
0.0000000000 11.1092567444 0.0000000000
0.0000000000 0.0000000000 15.4741468430
/
My previous LMODEL=enet fit residuals were 30%, which is not ideal
I tried to modify LMODEL = ols, it is still being calculated, if this fitting effect is not good, which parameters can I adjust to improve the fitting effect,
Then I would like to ask how to determine the cutoff here, the recommended value you gave in the tutorial is none 15 9, is it possible that this value can be explicitly obtained from the log like the second-order force constant, or is it an empirical value that needs to be tested?

[ttadano]

If both the training error and validation error are as large as 30%, you will need to make the Taylor expansion model more flexible by increasing the cutoff radii or by increasing NORDER.

If the training error is as small as a few percent, the validation error can be reduced by increasing the training data.

[liudadie12138]

Thank you for your answer, I'm going to try it now.