search

MaterSim / PyXtal_FF

Machine Learning Interatomic Potential Predictions
85 stars 23 forks source link

ZBL PyXtal_FF and ZBL LAMMPS are inconsistent with small difference between inner and outer cutoff. #45

Closed yanxon closed 3 years ago

yanxon commented 3 years ago

@qzhu2017

I am testing ZBL PyXtal_FF with Si dataset.

For instance, when the inner cutoff = 2.5 and outer cutoff = 3.0:

Structure:  0
Max force difference:  2.7507164742677834
Structure 0 has difference in force larger than 1e-2
Max stress difference:  2.9181099398486587
Structure 0 has difference in stress larger than 1

Structure:  1
Max force difference:  2.46994635896664
Structure 1 has difference in force larger than 1e-2
Max stress difference:  2.1553954217836555
Structure 1 has difference in stress larger than 1

the inner cutoff = 2.5 and outer cutoff = 3.5:

Structure:  0
Max force difference:  0.23996299312963743
Structure 0 has difference in force larger than 1e-2

Structure:  1
Max force difference:  0.21533818936499882
Structure 1 has difference in force larger than 1e-2

Note: Stress is correct here.

the inner cutoff = 2.5 and outer cutoff = 4.0:

Structure:  0
Max force difference:  0.043034149352700446
Structure 0 has difference in force larger than 1e-2

Structure:  1
Max force difference:  0.03860776235454111
Structure 1 has difference in force larger than 1e-2

Note: The force error is less pronounce.

As the difference between outer and inner is greater, the error between pyxtal_ff and lammps becomes less pronounced. Nevertheless, the energy is always correct.

qzhu2017 commented 3 years ago

@yanxon Check my recent commit. Go to PyXtal_FF/examples/Si-LAMMPS, you can run the script as follows. Clearly, something is messed up with the zbl forces. Should be an easy fix. 

qzhu@cms Si-LAMMPS (master) $ python validate.py 
  0  -41.164 eV   20.792 GPa    0.000    0.000    0.000
  1  -40.567 eV   21.046 GPa    0.000    0.000    0.000
  2  -41.297 eV   20.267 GPa    0.000    0.000    0.000
  3  -41.539 eV   18.660 GPa    0.000    0.000    0.000
  4  -41.333 eV   19.804 GPa    0.000    0.000    0.000
  5  -41.349 eV   19.034 GPa    0.000    0.000    0.000
  6  -41.242 eV   19.493 GPa    0.000    0.000    0.000
  7  -41.288 eV   19.123 GPa    0.000    0.000    0.000
  8  -41.417 eV   18.865 GPa    0.000    0.000    0.000
  9  -41.246 eV   19.668 GPa    0.000    0.000    0.000
 10  -41.580 eV   19.202 GPa    0.000    0.000    0.000
 11  -41.481 eV   20.017 GPa    0.000    0.000    0.000
 12  -41.152 eV   20.199 GPa    0.000    0.000    0.000
 13  -41.348 eV   19.523 GPa    0.000    0.000    0.000
 14  -41.396 eV   19.366 GPa    0.000    0.000    0.000
 15  -40.854 eV   19.978 GPa    0.000    0.000    0.000
 16  -41.783 eV   18.815 GPa    0.000    0.000    0.000
 17  -41.789 eV   18.177 GPa    0.000    0.000    0.000
 18  -41.424 eV   19.139 GPa    0.000    0.000    0.000
 19  -40.430 eV   21.355 GPa    0.000    0.042    0.055
eng:  -40.42961328110127 -40.429798811356356
Forces from LAMMPS and PyXtal_FF
  -1.573    2.380   -0.234 ->   -1.573    2.380   -0.234 ->    0.000   -0.000    0.000
  -4.037   -0.948    2.192 ->   -4.020   -0.930    2.175 ->   -0.017   -0.018    0.017
   1.625    2.312   -2.010 ->    1.625    2.312   -2.010 ->   -0.000   -0.000    0.000
   2.381   -2.386   -1.735 ->    2.381   -2.386   -1.735 ->   -0.000    0.000    0.000
   1.491   -0.652    3.314 ->    1.491   -0.652    3.314 ->   -0.000    0.000   -0.000
  -2.411   -2.595    1.360 ->   -2.411   -2.595    1.360 ->    0.000    0.000   -0.000
   3.256    4.539   -3.311 ->    3.240    4.521   -3.294 ->    0.017    0.018   -0.017
  -0.733   -2.650    0.424 ->   -0.733   -2.650    0.424 ->    0.000    0.000    0.000

 Breakdown of Pyxtal_FF
energy_ml (eV): -40.42996946967095
energy_zbl(eV): 0.0001706583145955065
forces (eV/A)
  -1.573    2.380   -0.234 ->    0.000    0.000    0.000
  -4.004   -0.913    2.160 ->   -0.016   -0.017    0.016
   1.625    2.312   -2.010 ->    0.000    0.000    0.000
   2.381   -2.386   -1.735 ->    0.000    0.000    0.000
   1.491   -0.652    3.314 ->    0.000    0.000    0.000
  -2.411   -2.595    1.360 ->    0.000    0.000    0.000
   3.224    4.504   -3.279 ->    0.016    0.017   -0.016
  -0.733   -2.650    0.424 ->    0.000    0.000    0.000
stress_ml (GPa, xx, yy, zz, xy, xz, yz): [ 21.3135  21.1453  21.5978   1.5742  -5.9206  -3.7250]
stress_zbl(GPa, xx, yy, zz, xy, xz, yz): [  0.0197   0.0230   0.0199   0.0213  -0.0198  -0.0214]
yanxon commented 3 years ago

@qzhu2017

I fixed this issue in 657eab73589416ec6e8f11132b32b85930ca0bbf.

Let me see if I can update the test_all.py script too to include a very small difference between inner and outer cutoff for future convenience.

yanxon commented 3 years ago

The test case is added in 3b38f46c846f9b28d19fc2b9a5c49e38edd815c2.

qzhu2017 commented 3 years ago

@qzhu2017 The bug has not been fixed on my side after f65289589bca9dd5059aaa8c5a75a44673d84f7d. In PyXtal_FF/examples/Si-LAMMPS, I run

qzhu@cms Si-LAMMPS (master) $ python validate.py 
  0  -41.525 eV   19.459 GPa    0.000    0.000    0.000
  1  -41.322 eV   19.244 GPa    0.000    0.000    0.000
  2  -41.325 eV   20.346 GPa    0.000    0.000    0.000
  3  -41.008 eV   19.798 GPa    0.000    0.034    0.044
eng:  -41.00799588506293 -41.00812871860695
Forces from LAMMPS and PyXtal_FF
  -3.339   -2.387   -3.270 ->   -3.326   -2.372   -3.256 ->   -0.013   -0.015   -0.014
   4.284    2.315    3.168 ->    4.271    2.300    3.154 ->    0.013    0.015    0.014
  -0.279    0.182    1.095 ->   -0.279    0.182    1.095 ->   -0.000   -0.000   -0.000
  -1.527    1.850   -2.837 ->   -1.527    1.850   -2.837 ->    0.000   -0.000    0.000
  -0.434    0.084   -0.442 ->   -0.434    0.084   -0.442 ->   -0.000   -0.000    0.000
   1.728   -1.826   -1.479 ->    1.728   -1.826   -1.479 ->   -0.000    0.000    0.000
   0.210   -0.858    3.127 ->    0.210   -0.858    3.127 ->   -0.000    0.000   -0.000
  -0.643    0.640    0.638 ->   -0.643    0.640    0.638 ->    0.000   -0.000   -0.000

 Breakdown of Pyxtal_FF
energy_ml (eV): -41.008250609827485
energy_zbl(eV): 0.00012189122053740675
forces (eV/A)
  -3.315   -2.358   -3.243 ->   -0.012   -0.014   -0.013
   4.259    2.286    3.141 ->    0.012    0.014    0.013
  -0.279    0.182    1.095 ->    0.000    0.000    0.000
  -1.527    1.850   -2.837 ->    0.000    0.000    0.000
  -0.434    0.084   -0.442 ->    0.000    0.000    0.000
   1.728   -1.826   -1.479 ->    0.000    0.000    0.000
   0.210   -0.858    3.127 ->    0.000    0.000    0.000
  -0.643    0.640    0.638 ->    0.000    0.000    0.000
stress_ml (GPa, xx, yy, zz, xy, xz, yz): [ 19.7681  20.8087  19.5361  -0.5927   2.0311   3.1119]
stress_zbl(GPa, xx, yy, zz, xy, xz, yz): [  0.0145   0.0189   0.0167   0.0165   0.0155   0.0178]

The forces are not consistent.

qzhu2017 commented 3 years ago

@yanxon Here is the updated output from validate.py

qzhu@cms Si-LAMMPS (master) $ python validate.py 
  0  -41.525 eV   19.459 GPa    0.000    0.000    0.000
  1  -41.322 eV   19.244 GPa    0.000    0.000    0.000
  2  -41.325 eV   20.346 GPa    0.000    0.000    0.000
  3  -41.008 eV   19.798 GPa    0.000    0.034    0.044
eng:  -41.00799588506293 -41.00812871860695
Forces from LAMMPS and PyXtal_FF
 -3.3393  -2.3866  -3.2695 ->  -3.3264  -2.3719  -3.2557 ->  -0.0129  -0.0147  -0.0138
  4.2840   2.3147   3.1680 ->   4.2711   2.2999   3.1542 ->   0.0129   0.0147   0.0138
 -0.2785   0.1817   1.0952 ->  -0.2785   0.1817   1.0952 ->  -0.0000  -0.0000  -0.0000
 -1.5274   1.8504  -2.8370 ->  -1.5274   1.8504  -2.8370 ->   0.0000  -0.0000   0.0000
 -0.4340   0.0841  -0.4425 ->  -0.4340   0.0841  -0.4425 ->  -0.0000  -0.0000   0.0000
  1.7282  -1.8263  -1.4795 ->   1.7282  -1.8263  -1.4795 ->  -0.0000   0.0000   0.0000
  0.2098  -0.8583   3.1269 ->   0.2098  -0.8583   3.1269 ->  -0.0000   0.0000  -0.0000
 -0.6428   0.6405   0.6383 ->  -0.6428   0.6405   0.6383 ->   0.0000  -0.0000  -0.0000

 Breakdown of Pyxtal_FF
energy_ml (eV): -41.008250609827485
energy_zbl(eV): 0.00012189122053740675
forces (eV/A)
  -3.315   -2.358   -3.243 ->   -0.012   -0.014   -0.013
   4.259    2.286    3.141 ->    0.012    0.014    0.013
  -0.279    0.182    1.095 ->    0.000    0.000    0.000
  -1.527    1.850   -2.837 ->    0.000    0.000    0.000
  -0.434    0.084   -0.442 ->    0.000    0.000    0.000
   1.728   -1.826   -1.479 ->    0.000    0.000    0.000
   0.210   -0.858    3.127 ->    0.000    0.000    0.000
  -0.643    0.640    0.638 ->    0.000    0.000    0.000
stress_ml (GPa, xx, yy, zz, xy, xz, yz): [ 19.7681  20.8087  19.5361  -0.5927   2.0311   3.1119]
stress_zbl(GPa, xx, yy, zz, xy, xz, yz): [  0.0145   0.0189   0.0167   0.0165   0.0155   0.0178]

 Breakdown of LAMMPS
energy_ml (eV): -41.00825062318803
energy_zbl(eV): 0.0002547247607291858
Forces
  -3.315   -2.358   -3.243 ->   -0.025   -0.028   -0.027
   4.259    2.286    3.141 ->    0.025    0.028    0.027
  -0.279    0.182    1.095 ->    0.000    0.000    0.000
  -1.527    1.850   -2.837 ->    0.000    0.000    0.000
  -0.434    0.084   -0.442 ->    0.000    0.000    0.000
   1.728   -1.826   -1.479 ->    0.000    0.000    0.000
   0.210   -0.858    3.127 ->    0.000    0.000    0.000
  -0.643    0.640    0.638 ->    0.000    0.000    0.000
stress_ml (GPa, xx, yy, zz, xy, xz, yz): [ 19.7681  20.8087  19.5361  -0.5927   2.0311   3.1119]
stress_zbl(GPa, xx, yy, zz, xy, xz, yz): [  0.0301   0.0394   0.0348   0.0344   0.0323   0.0370]

It seems that the zbl output from PyXtal is about only half of the values from lammps for this structure.

qzhu2017 commented 3 years ago

@yanxon Can you resolve this issue soon. It seems that there is a fundamental problem that both energy/forces from pyxtal_FF are only about half of the outputs from lammps.

yanxon commented 3 years ago

@qzhu2017 I am looking at it. I'm not sure what's wrong yet. Hopefully, I can spot something. Sometimes, it's just a silly mistake.

yanxon commented 3 years ago

@qzhu2017 Can we please have a quick meeting today, say 5-10 mins? I think I know what's going on.

qzhu2017 commented 3 years ago

https://unlv.webex.com/meet/qiang.zhu

On Sat, May 15, 2021 at 6:27 PM Howard Yanxon @.***> wrote:

@qzhu2017 https://github.com/qzhu2017 Can we please have a quick meeting today, say 5-10 mins? I think I know what's going on.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/qzhu2017/PyXtal_FF/issues/45#issuecomment-841750411, or unsubscribe https://github.com/notifications/unsubscribe-auth/AHAUZ5RM5AFLHGE5F2GJSFLTN4NQRANCNFSM445CXFUA .

-- [image: Photograph] http://www.skarif2.io/ Qiang Zhu Assistant Professor University of Nevada Las Vegas • Physics and Astronomy Phone ⋅ +1 702 895 1710 Email ⋅ @. Address ⋅ BPB 232, 4505 S. Maryland Parkway @.,-115.1454793,17z/data=!3m1!4b1!4m5!3m4!1s0x80c8c5aacc5bd7bf:0x57992e2f948155c0!8m2!3d36.1092077!4d-115.1432906> [image: github] https://github.com/qzhu2017[image: linkedin] https://qzhu2017.github.io/

yanxon commented 3 years ago

@qzhu2017

  1. I successfully installed lammps-python and ran validate.py. I got this result:

    3  -37.583 eV   50.343 GPa    0.029    1.140    1.615
eng:  -37.582944530794485 -37.611991540348335
Forces from LAMMPS and PyXtal_FF
-5.1019  -3.9989  -5.1461 ->  -4.6704  -3.5021  -4.6808 ->  -0.4315  -0.4968  -0.4653
6.4002   3.9275   4.8345 ->   5.9687   3.4307   4.3692 ->   0.4315   0.4968   0.4653
-0.1972   0.2448   1.5855 ->  -0.1972   0.2448   1.5855 ->  -0.0000  -0.0000  -0.0000
-2.1801   2.4429  -3.6152 ->  -2.1690   2.4330  -3.6052 ->  -0.0111   0.0099  -0.0100
-0.6702   0.1857  -0.7116 ->  -0.6702   0.1857  -0.7116 ->  -0.0000  -0.0000   0.0000
2.4327  -2.4534  -1.7964 ->   2.4327  -2.4534  -1.7964 ->  -0.0000   0.0000   0.0000
0.2616  -1.1736   4.1861 ->   0.2506  -1.1636   4.1761 ->   0.0111  -0.0099   0.0100
-0.9453   0.8250   0.6633 ->  -0.9453   0.8250   0.6633 ->   0.0000  -0.0000  -0.0000

Breakdown of Pyxtal_FF
energy_ml (eV): -37.64178279348804
energy_zbl(eV): 0.029791253139706875
forces (eV/A)
-4.206   -2.967   -4.180 ->   -0.465   -0.535   -0.501
5.504    2.896    3.868 ->    0.465    0.535    0.501
-0.197    0.245    1.585 ->    0.000    0.000    0.000
-2.159    2.424   -3.596 ->   -0.010    0.009   -0.009
-0.670    0.186   -0.712 ->    0.000    0.000    0.000
2.433   -2.453   -1.796 ->    0.000    0.000    0.000
0.240   -1.154    4.167 ->    0.010   -0.009    0.009
-0.945    0.825    0.663 ->    0.000    0.000    0.000
stress_ml (GPa, xx, yy, zz, xy, xz, yz): [ 49.1432  50.9964  48.6411  -1.1376   2.7436   4.5140]
stress_zbl(GPa, xx, yy, zz, xy, xz, yz): [  0.6208   0.8148   0.7165   0.6821   0.6667   0.7381]

Breakdown of LAMMPS
energy_ml (eV): -37.64178281119801
energy_zbl(eV): 0.058838262604393066
Forces
-4.206   -2.967   -4.180 ->   -0.896   -1.032   -0.966
5.504    2.896    3.868 ->    0.896    1.032    0.966
-0.197    0.245    1.585 ->    0.000    0.000    0.000
-2.159    2.424   -3.596 ->   -0.021    0.019   -0.019
-0.670    0.186   -0.712 ->    0.000    0.000    0.000
2.433   -2.453   -1.796 ->    0.000    0.000    0.000
0.240   -1.154    4.167 ->    0.021   -0.019    0.019
-0.945    0.825    0.663 ->    0.000    0.000    0.000
stress_ml (GPa, xx, yy, zz, xy, xz, yz): [ 49.1432  50.9964  48.6411  -1.1376   2.7436   4.5140]
stress_zbl(GPa, xx, yy, zz, xy, xz, yz): [  1.1999   1.5736   1.3839   1.3133   1.2882   1.4214]

  2. Now, I take the "random" structure 3 and convert it to lammps data structure and the corresponding cif file. Then, I ran it with lmp_mpi using the in_file, and I get the same results as the validate.py, which is expected.

  3. Then you can run the base_potential.py in https://github.com/qzhu2017/PyXtal_FF/blob/master/examples/Si-LAMMPS/validate/base_potential.py. The base_potential.py is the one that I currently have implemented in PyXtal_FF. Here is the result:

    Energy:  0.05881695165257561
Force:
[[-0.895870 -1.031627 -0.966130]
[ 0.895870  1.031627  0.966130]
[ 0.000000  0.000000  0.000000]
[-0.021165  0.018975 -0.019159]
[ 0.000000  0.000000  0.000000]
[ 0.000000  0.000000  0.000000]
[ 0.021165 -0.018975  0.019159]
[ 0.000000  0.000000  0.000000]]
Stress:  [ 0.007486  0.009819  0.008635  0.008196  0.008038  0.008871]

Conc: I don't think there is anything wrong with the base_potential.py. Perhaps, there is something wrong with the Calculator.

yanxon commented 3 years ago

@qzhu2017

If you run this script:

import torch

model = torch.load("Si-snap-zbl/12-12-checkpoint.pth")
print(model['des_info'])

You will get this:

{'system': ['Si'], 'type': 'SNAP', 'Rc': 5.0, 'weights': {'Si': 1.0}, 'N': None, 'N_train': None, 'N_test': None, 'ncpu': 1, 'force': True, 'stress': True, 'cutoff': 'cosine', 'base_potential': {'inner': 1.0, 'outer': 2.0}, 'parameters': {'lmax': 3, 'rfac': 0.99363, 'normalize_U': False}}

The inner/outer is 1.0/2.0, not 1.5/2.0.

If you update the NN model, this should solve the issue.

qzhu2017 commented 3 years ago

@yanxon I just fixed it as follows