Can we train the CHGNet model with multiple DFT optimized structures (Not MD simulation data)?

[Souvik-ml]

I have trained CHGNet model with multiple DFT optimized structures (Not MD simulation, just optimization). However, when I am using the trained CHGNet model for optimization of those structures which were a part of training data, the predicted energy not even close. For example,
chgnet_ft = CHGNet.from_file('/content/bestE_epoch98_e52_f101_s162_mNA.pth.tar')
struc = Structure.from_file("/content/POSCAR")
opt = StructOptimizer(model=chgnet_ft)
relax = opt.relax(struc)
relax

the final CHGNet optimized energy is -0.280197 eV

where as my DFT (VASP) calculated optimized energy is -76.41950045 eV.

There must be something wrong in my approach. But I'm not getting where is the problem.

if someone can help me understanding this, I'll be very grateful.

Thanks