FLARE_Tutorial_2021.ipynb produces incorrect potential energies
Closed this issue · 3 comments
Hi MIR team,
I was trying to run FLARE_Tutorial_2021.ipynb in colab (https://colab.research.google.com/drive/18_pTcWM19AUiksaRyCgg9BCpVyw744xv). The aspirin example runs smoothly, indicating that the installation of flare_pp is fine.
The Al example example, however, produces an incorrect potential energy curve although the temperature curve is good.
Do you have any suggestions?
Best,
Eric
The output of test_otf.run() is as follows
/usr/local/lib/python3.7/dist-packages/ase/io/extxyz.py:303: UserWarning: Skipping unhashable information adsorbate_info
'{0}'.format(key))
Precomputing KnK for hyps optimization
Done precomputing. Time: 0.0034034252166748047
Hyperparameters:
[2.0e+00 9.7e-02 5.0e-02 6.0e-04]
Likelihood gradient:
[-3.70118607e+00 0.00000000e+00 -4.65226400e+03 0.00000000e+00]
Likelihood:
507.64399120059767
Hyperparameters:
[ 1.99920443e+00 9.70000000e-02 -9.49999684e-01 6.00000000e-04]
Likelihood gradient:
[ 6.34674994 0. 294.90930057 0. ]
Likelihood:
-265.5908856426073
Hyperparameters:
[ 1.99979561e+00 9.70000000e-02 -2.06912782e-01 6.00000000e-04]
Likelihood gradient:
[1.11347928e+00 0.00000000e+00 1.27994395e+03 0.00000000e+00]
Likelihood:
151.22101744005056
Hyperparameters:
[ 1.99995271e+00 9.70000000e-02 -9.44565431e-03 6.00000000e-04]
Likelihood gradient:
[ 4.2125249 0. 244.53980128 0. ]
Likelihood:
780.3038128903622
Hyperparameters:
[ 2.00000141e+00 9.70000000e-02 -6.47705624e-03 6.00000000e-04]
Likelihood gradient:
[ 1.45795924e+01 0.00000000e+00 -3.47710594e+04 0.00000000e+00]
Likelihood:
743.1828488459068
Hyperparameters:
[ 1.99995374e+00 9.70000000e-02 -9.38239733e-03 6.00000000e-04]
Likelihood gradient:
[ 4.33104119 0. -55.22303239 0. ]
Likelihood:
780.3098644361403
Hyperparameters:
[ 1.99995448e+00 9.70000000e-02 -9.39403540e-03 6.00000000e-04]
Likelihood gradient:
[4.31238784 0. 1.27964709 0. ]
Likelihood:
780.3101727506503
Hyperparameters:
[ 1.99995537e+00 9.70000000e-02 -9.39382734e-03 6.00000000e-04]
Likelihood gradient:
[4.31761004 0. 1.0165541 0. ]
Likelihood:
780.3101660407056
Hyperparameters:
[ 1.99995454e+00 9.70000000e-02 -9.39402122e-03 6.00000000e-04]
Likelihood gradient:
[4.3103268 0. 1.27027891 0. ]
Likelihood:
780.3101724742858
Hyperparameters:
[ 1.99995448e+00 9.70000000e-02 -9.39403409e-03 6.00000000e-04]
Likelihood gradient:
[4.3180535 0. 0.91302183 0. ]
Likelihood:
780.3101725680785
Hyperparameters:
[ 1.99995448e+00 9.70000000e-02 -9.39403537e-03 6.00000000e-04]
Likelihood gradient:
[4.30461414 0. 1.87590886 0. ]
Likelihood:
780.3101736194021
Hyperparameters:
[ 1.99995448e+00 9.70000000e-02 -9.39403537e-03 6.00000000e-04]
Likelihood gradient:
[4.30962056 0. 0.95950237 0. ]
Likelihood:
780.3101729329906
Hyperparameters:
[ 1.99995448e+00 9.70000000e-02 -9.39403537e-03 6.00000000e-04]
Likelihood gradient:
[4.31000367 0. 1.03232979 0. ]
Likelihood:
780.3101727639694
Hyperparameters:
[ 1.99995448e+00 9.70000000e-02 -9.39403537e-03 6.00000000e-04]
Likelihood gradient:
[4.31051909 0. 1.59868165 0. ]
Likelihood:
780.3101735763009
Precomputing KnK for hyps optimization
Done precomputing. Time: 0.01186990737915039
Hyperparameters:
[ 1.99995448e+00 9.70000000e-02 -9.39403537e-03 6.00000000e-04]
Likelihood gradient:
[ 12.48799445 0. -1710.43981304 0. ]
Likelihood:
1679.021783954544
Hyperparameters:
[ 2.00725532e+00 9.70000000e-02 -1.00936738e+00 6.00000000e-04]
Likelihood gradient:
[ 5.407869 0. 563.3832408 0. ]
Likelihood:
-557.1375466746391
Hyperparameters:
[ 2.00066878e+00 9.70000000e-02 -1.07229350e-01 6.00000000e-04]
Likelihood gradient:
[-3.44192440e+00 0.00000000e+00 4.96793293e+03 0.00000000e+00]
Likelihood:
686.1362494999759
Hyperparameters:
[ 1.99997620e+00 9.70000000e-02 -1.23695363e-02 6.00000000e-04]
Likelihood gradient:
[4.79527246e+00 0.00000000e+00 1.60591521e+04 0.00000000e+00]
Likelihood:
1651.3921042322436
Hyperparameters:
[ 1.99995573e+00 9.70000000e-02 -9.56535442e-03 6.00000000e-04]
Likelihood gradient:
[ 11.77653005 0. 172.39427354 0. ]
Likelihood:
1679.1509752202005
Hyperparameters:
[ 1.99995669e+00 9.70000000e-02 -9.54967573e-03 6.00000000e-04]
Likelihood gradient:
[11.85817338 0. 9.30494455 0. ]
Likelihood:
1679.152404069011
Hyperparameters:
[ 1.99995789e+00 9.70000000e-02 -9.54871056e-03 6.00000000e-04]
Likelihood gradient:
[11.8466115 0. -2.89073442 0. ]
Likelihood:
1679.152406529191
Hi @eric-yu-zhu
This is one small revision you can make, that is to add 'force_only': False to the otf_params dict. Then everything should work. We recently made the change that we set force_only to True by default, therefore, in the tutorial the energy labels from DFT are not used for training and the energy prediction of the model is awful. We will update this tutorial soon.
Also, we recently merged flare_pp to flare, and you can check the latest code and tutorials here: https://github.com/mir-group/flare/tree/development
You can check the flare pre-release 1.1.2 for the best compatibility of dependencies.
Hi Yu, this is to confirm that the problem is fixed by adding "force_only" parameter. Everything works nicely now. Thanks a lot!
Eric