Large value of RMSE while fine-tuning 7net-0

[DGKang234]

Dear Developers,

I attempt to fine-tuning 7net-0 model with the vasp results of the single point calculated randomly scattered atoms.
During the fine-tuning (I followed the fine-tuning tutorial) the RMSE of each epoch is very high:

Epoch 10/10  Learning rate: 0.000400
------------------------------------------------------------------------------------------------------------------------
       Energy_RMSE (eV/atom)   Force_RMSE (eV/Å)   Stress_RMSE (kbar)   Energy_MAE (eV/atom)   Force_MAE (eV/Å)   Stress_MAE (kbar)   Energy_Loss   Force_Loss   Stress_Loss   TotalLoss  
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Train             178.687156           81.152464         68301.612801              82.339354          16.575042         8404.350185      0.823344     0.165701     84.043455    1.829479
Valid             183.634165           91.979268         76253.958309              92.050893          19.477570        10124.414716      0.920459     0.194726    101.244100    2.127626
Epoch 10 elapsed: 0:00:23.23

Could you please give advice to point out what I am doing wrong?

P.S.

I used INCAR for the single point;

PREC = Accurate       # Precision of the calculation
ENCUT = 520           # Plane-wave cutoff energy

# Electronic Relaxation
EDIFF = 1E-6 #5          # Energy convergence criterion for electronic self-consistency
ALGO = Normal           # Algorithm for electronic minimisation

# Ionic Relaxation
IBRION = 2            # Conjugate gradient algorithm for ionic relaxation
NSW = 0 #1000             # Maximum number of ionic steps
EDIFFG = -0.03        # Force convergence criterion (negative sign indicates force)
ISIF = 2              # Relaxation of atomic positions, calculation of stress tensor

# Output Settings
LCHARG = .FALSE.      # Do not write charge density
LWAVE = .FALSE.       # Do not write wavefunctions

Cheers,

Dr. Kang

[YutackPark]

Hi, thanks for reaching out. Here's my suggestions.

1. Re-try with below fine-tuning preset. You can either use terminal interface (sevenn train input.yaml and sevenn preset) or python interface you used in the tutorial.

https://github.com/MDIL-SNU/SevenNet/blob/main/sevenn%2Fpresets%2Ffine_tune_le.yaml

2. I'm concerned about 'scattered' atoms you mentioned in your training dataset. That could be very different from the original training dataset of 7net-0 (bulk crystal structure).
   In this case, you should increase your learning rate or increase the number of the training data. However, the best approach is using both the MPtrj dataset and yours (commonly referred as 'replay method' in ML community)

3. Finally, if you haven't done yet and your computational resource is enough, try with the latest models without fine-tuning. The performance gap between 7net-0 and latest models are huge. We have observed the latest models perform better than fine-tuned 7net-0, in some cases.

[DGKang234]

Hi Dr Park,
Thank you for the reply.

1. I have not tried to use .yaml for training. I will try fintuning with .yaml format: should I just edit the path of the trianing data? (apologise for the newbie question)

2. If I train with MPtrj and my training data (scattered atoms/liquid of atoms) should I train from scratch or finetuning would be okay?

3. yes, I will use the recent model to train/finetune.

Thank you.
Dr. Kang

[YutackPark]

I have not tried to use .yaml for training. I will try fintuning with .yaml format: should I just edit the path of the trianing data? (apologise for the newbie question)

I think you can start from there. But to the best of my knowledge, there's no magic hyperparameter set that works for every system. So maybe you need some adjustment based on your result.

If I train with MPtrj and my training data (scattered atoms/liquid of atoms) should I train from scratch or finetuning would be okay?

You still fine-tuning the model. This is a reference of the recent, related work: https://arxiv.org/pdf/2506.15223
Note: The authors are planning to make contribution to this repository that implements the method demonstrated in the paper.

[DGKang234]

I love to use the reEWC (https://github.com/kskjs1203/SevenNet-reEWC) for the finetuning.
I wonder when would it be available in this repository?

in the mean time, I want to work it out to make the potential to describe the designated systems with fair accuracy which I will follow your suggestion; replay method: omat potential <-- finetune w/ omat data + finetuning data.

[kskjs1203]

Thank you for your interest in reEWC. I’m the first author of the work, and since the paper has now been accepted, I’m planning to merge the reEWC implementation into this repository within about a month.

Thank you again for your interest and patience. If you have any further questions, please feel free to reach out to me by email. I’ll post an update here once the reEWC integration is complete.