In the NPT simulation, it expands to gas above the melting points

[pobo95]

When I tested NPT simulation above the melting points

it expands to gas but it didn't when I used QUIP ML potential

here is my trianing parameters

E0s = average
num_interactions = 2
correlation = 2
max_L = 1
r_max = 5.0
hidden_irreps = 64x0e + 64x1o

full command
mace_run_train --name="mace" --train_file="train.xyz" --valid_fraction=0.05 --test_file="validate.xyz" --E0s="average" --model="MACE" --num_interactions=2 --hidden_irreps='64x0e + 64x1o' --max_L=1 --correlation=2 --r_max=5.0 --forces_weight=1000 --energy_weight=10 --batch_size=5 --valid_batch_size=2 --max_num_epochs=5000 --compute_stress=”true” --start_swa=450 --scheduler_patience=5 --patience=15 --eval_interval=3 --ema --swa --swa_forces_weight=10 --default_dtype="float64" --device=cuda --seed=123 --restart_latest --save_cpu --error_table=PerAtomRMSEstressvirials

It shows
RMSE energy per atom : 1.1 meV/atom
RMSE force per atom : 81.8 meV/A
RMSE stress per atom : 0.01 meV/A^3
for test file

I attach my traing and validate files too

train_validate.zip

Do you have any clues to slove it?

Thanks

[gabor1]

yes. what's happenning here is that your isolated atom energies are not correct. you are using the "average" method for getting E0s. if you calculate isolated atom energies, and tell MACE to use those instead, you will not see the evaporation. You are saying that "QUIP ML", by which I guess you mean SOAP/GAP, didn't evaorate - perhaps you used calculated isolated atom energies for E0s there ?

[pobo95]

Thank you for your quick answer.
That's correct I calculated isolated atom energys for E0s when I used SOAP/GAP. I will try trianing again with exact E0s.

Does isolated atom energy matter when calculating pressure?

[gabor1]

the model is completely different when you include real E0s, so yes, your stress predictions will change.

[pobo95]

Thank you so much. It helped me a lot. I will retrain my potentials with E0s. I think I've solved the problem thank you :)

[gabor1]

I belived massage passing process can describe binding energies.

it can, but only if you give E0s explicitly. it cannot guess the correct E0s, there is absolutely nothing in your training data that can determine what the isolated atom energies are (unless you have the isolated atoms in the training data set and are set up to be treated in a special way, but we don't especially look for it currently)

[bernstei]

If you include configs with isolated atoms and set config_type=IsolatedAtom (or something like that), and you don't pass any --E0s flag, I believe it uses those configs.

[pobo95]

I got it, Thank you maybe I can test include isolated atoms in my training data :)

[pobo95]

If you include configs with isolated atoms and set config_type=IsolatedAtom (or something like that), and you don't pass any --E0s flag, I believe it uses those configs.

@bernstei I just check it in the utils.py file thank you for remind.
BTW when I use --E0s={32:-0.02859140,51:-0.03305661,52:-0.03357790} flag it only read the last one (52:-0.03357790) in mace = 0.3.4 v, python = 3.9.7 v. Maybe somthing wrong in my system.