5.3. Generate reference values¶
In all previous tutorials, the reference data was either already pre-calculated, or calculated by setting up a job in AMSinput and running it in the normal way before importing the results to ParAMS.
You can also use ParAMS to calculate the reference values. That is illustrated in this tutorial.
This can be useful if you, for example:
want to reevaluate (parts of) the training set with a different reference method,
have added a reaction energy without a reference value, or
only have a few reference jobs that are relatively quick to calculate.
Note
It is usually better to first run the reference calculations and then import them into ParAMS (i.e., to not follow this tutorial, but instead the Import training data (GUI) or Import training data (Python) tutorials). That way you can:
run the reference calculations on multiple nodes,
store the reference calculations and results in arbitrary places on disk,
inspect the reference results and correct any errors before the parametrization, and
use the results importer for PESScans which generate plots.
5.3.1. Prerequisites¶
Go through the Getting Started: Lennard-Jones tutorial.
Make a copy of the example directory
$AMSHOME/scripting/scm/params/examples/LJ_Ar_no_reference_data
.
5.3.2. The input files¶
5.3.2.1. Training set without reference values (training_set.yaml)¶
The reference values will be calculated when you start the parametrization.
Unlike the training_set.yaml file in the previous tutorial, this example’s training_set.yaml does not contain any reference values:
---
dtype: DataSet
version: '2022.101'
---
Expression: energy('Ar32_frame001')-energy('Ar32_frame002')
Weight: 1.0
Sigma: 0.054422772491975996
Unit: eV, 27.211386245988
---
Expression: energy('Ar32_frame003')-energy('Ar32_frame002')
Weight: 1.0
Sigma: 0.054422772491975996
Unit: eV, 27.211386245988
---
Expression: forces('Ar32_frame001')
Weight: 1.0
Sigma: 0.15426620242897765
Unit: eV/angstrom, 51.422067476325886
---
Expression: forces('Ar32_frame002')
Weight: 1.0
Sigma: 0.15426620242897765
Unit: eV/angstrom, 51.422067476325886
---
Expression: forces('Ar32_frame003')
Weight: 1.0
Sigma: 0.15426620242897765
Unit: eV/angstrom, 51.422067476325886
...
This tutorial shows you how you can calculate the reference values (in units of Unit) with params
.
5.3.2.2. Jobs and Engines (job_collection.yaml, job_collection_engines.yaml)¶
All Jobs that are mentioned in training set entries without a reference value, will be calculated before the parametrization starts with the help of a Reference Engine.
dftb;;kspace;;quality;GammaOnly;model;GFN1-xTB;
.This rather cryptic ID is a reference engine id.
Engine dftb kspace ......
.GFN1xTB_GammaPoint
(do not use spaces in the name).This affects all jobs with that particular reference engine.
To edit the engine used for a job:
Ar32_frame002
job in the Detail column (where it says SinglePoint + gradients)GFN1xTB;2526569033
. The numbers are a hash of the settings.Ar32_frame002
job has the new reference engine. You can rename it on the Engines panel if you prefer.In this way, you can create an arbitrary number of engines, or choose already-created engines.
Ar32_frame002
job in the Detail column (where it says SinglePoint + gradients)GFN1xTB_GammaPoint
).GFN1xTB_GammaPoint
engineUnlike the previous tutorial, here each entry in the job_collection.yaml has a ReferenceEngineID. For example, the first entry is
1---
2Engines: job_collection_engines.yaml
3dtype: JobCollection
4version: '2022.101'
5---
6ID: 'Ar32_frame001'
7ReferenceEngineID: dftb;;kspace;;quality;GammaOnly;model;GFN1-xTB;
8AMSInput: |
9 properties
10 gradients Yes
11 End
12 system
13 Atoms
14 Ar 5.1883477539 -0.4887475488 7.9660568076
15 Ar 5.7991822399 0.4024595652 2.5103286966
16 Ar 6.1338265157 5.5335946219 7.0874208384
17 Ar 4.6137188191 5.9644505949 3.0942810122
18 Ar 8.4186778390 7.6292969115 8.0729664423
19 Ar 8.3937816110 8.6402371806 2.6057806799
20 Ar 7.5320205143 1.7666481606 7.7525889818
21 Ar 8.5630139885 2.0472039529 2.6380554086
22 Ar 2.6892353632 7.8435284207 7.7883054306
23 Ar 2.4061636915 7.5716025415 2.4535180075
24 Ar 2.2485171283 2.9764130946 7.8589298904
25 Ar 3.0711058946 1.8500587164 2.5620921469
26 Ar 7.6655637500 -0.4865893003 0.0018797080
27 Ar 7.7550067215 -0.0222821825 4.8528637785
28 Ar 7.7157262425 4.6625079517 -0.3861722152
29 Ar 7.7434900996 5.2619590353 4.2602386226
30 Ar 3.4302237084 -0.2708640738 0.6280466620
31 Ar 2.8648051689 0.6106220610 6.1208342905
32 Ar 3.2529823775 5.7151788324 -0.2024448179
33 Ar 2.0046357208 4.9353027402 5.4968740217
34 Ar 0.9326855213 8.0600564695 -0.3181225099
35 Ar -0.5654205469 8.5703446434 5.8930973456
36 Ar -0.9561293133 2.1098403312 -0.0052667919
37 Ar -0.8081417664 3.2747992855 5.5295389610
38 Ar 5.5571960244 7.5645919074 0.1312355350
39 Ar 4.4530832384 7.6170633330 5.4810860433
40 Ar 5.1235367625 2.7983577675 -0.3161069611
41 Ar 5.2048439076 2.9885672135 4.5193274119
42 Ar -0.2535891591 0.0134355189 8.3061692970
43 Ar 0.5614183785 -0.1927751317 3.2355155467
44 Ar -0.0234943080 5.0313863031 8.0451075074
45 Ar -0.4760138873 6.2617510830 2.5759742219
46 End
47 Lattice
48 10.5200000000 0.0000000000 0.0000000000
49 0.0000000000 10.5200000000 0.0000000000
50 0.0000000000 0.0000000000 10.5200000000
The ReferenceEngineID refers to an engine in job_collection_engines.yaml
, which is an Engine Collection. Each entry has a unique ID, and an AMSInput block containing calculation settings for the AMS engine.
---
Jobs: job_collection.yaml
dtype: EngineCollection
version: '2022.101'
---
ID: 'dftb;;kspace;;quality;GammaOnly;model;GFN1-xTB;'
AMSInput: |
Engine dftb
kspace
quality GammaOnly
End
model GFN1-xTB
EndEngine
...
In this example, there is only one reference engine. It has the ID
dftb;;kspace;;quality;GammaOnly;model;GFN1-xTB;
. The ID could be any
string. It does not affect the results, but should describe the reference
engine. Each job in the job collection with this ReferenceEngineID will be
evaluated with this reference engine.
The AMSInput affects the calculation. In this example, it sets up a GFN1-xTB engine with Γ-point sampling. The AMSInput will be added verbatim to the input to the reference job.
Note
If you do not have DFTB license, change the Engine
block in job_collection_engines.yaml
to
Engine ForceField
Type UFF
EndEngine
to instead use a UFF force field as the reference method.
5.3.3. Calculate the reference values¶
Select the Generate Reference task:
Save and run the task:
calc_ref_values.params
Once complete the GUI will automatically load the reference values:
Since the reference data has been automatically loaded, you may now immediately continue to setting up an optimization. You can switch the task back to Optimization, change any settings you like (for details see Getting Started: Lennard-Jones), and then save the job. The reference values will be saved within the new job folder as normal.
To run the reference jobs and generate the reference data, change to the example directory and run:
"$AMSBIN/params"
This runs the reference calculations, adds the reference data to the training set, and saves it in results/training_set.ref.yaml
To use the results in an optimization simply set one up as described in Getting Started: Lennard-Jones and ensure that the DataSet%Path
key points to results/training_set.ref.yaml
or a copy of it.
5.3.4. Output files for reference calculations and data¶
In the results folder, you will find:
the individual job results saved in
results/reference_jobs
the task input saved in
results/settings_and_initial_data
a file
training_set.ref.yaml
containing the reference values.
5.3.4.1. The reference_jobs folder¶
For example, the file reference_jobs/Ar32_frame001/Ar32_frame001.in
contains the input to the Ar32_frame001
job, which combines input from the
job collection and engine
collection:
properties
gradients yes
End
system
Atoms
Ar 5.1883477539 -0.4887475488 7.9660568076
Ar 5.7991822399 0.4024595652 2.5103286966
Ar 6.1338265157 5.5335946219 7.0874208384
Ar 4.6137188191 5.9644505949 3.0942810122
Ar 8.4186778390 7.6292969115 8.0729664423
Ar 8.3937816110 8.6402371806 2.6057806799
Ar 7.5320205143 1.7666481606 7.7525889818
Ar 8.5630139885 2.0472039529 2.6380554086
Ar 2.6892353632 7.8435284207 7.7883054306
Ar 2.4061636915 7.5716025415 2.4535180075
Ar 2.2485171283 2.9764130946 7.8589298904
Ar 3.0711058946 1.8500587164 2.5620921469
Ar 7.6655637500 -0.4865893003 0.0018797080
Ar 7.7550067215 -0.0222821825 4.8528637785
Ar 7.7157262425 4.6625079517 -0.3861722152
Ar 7.7434900996 5.2619590353 4.2602386226
Ar 3.4302237084 -0.2708640738 0.6280466620
Ar 2.8648051689 0.6106220610 6.1208342905
Ar 3.2529823775 5.7151788324 -0.2024448179
Ar 2.0046357208 4.9353027402 5.4968740217
Ar 0.9326855213 8.0600564695 -0.3181225099
Ar -0.5654205469 8.5703446434 5.8930973456
Ar -0.9561293133 2.1098403312 -0.0052667919
Ar -0.8081417664 3.2747992855 5.5295389610
Ar 5.5571960244 7.5645919074 0.1312355350
Ar 4.4530832384 7.6170633330 5.4810860433
Ar 5.1235367625 2.7983577675 -0.3161069611
Ar 5.2048439076 2.9885672135 4.5193274119
Ar -0.2535891591 0.0134355189 8.3061692970
Ar 0.5614183785 -0.1927751317 3.2355155467
Ar -0.0234943080 5.0313863031 8.0451075074
Ar -0.4760138873 6.2617510830 2.5759742219
End
Lattice
10.5200000000 0.0000000000 0.0000000000
0.0000000000 10.5200000000 0.0000000000
0.0000000000 0.0000000000 10.5200000000
End
End
task singlepoint
Engine dftb
kspace
quality GammaOnly
End
model GFN1-xTB
EndEngine
The normal AMS output can be found in the same folder: the logfile, standard outputfile, and the binary ams.rkf and dftb.rkf files.
5.3.4.2. The training_set.ref.yaml file¶
The training_set.ref.yaml file contains the calculated reference values. For example, it starts with
---
dtype: DataSet
version: '2022.101'
---
Expression: energy('Ar32_frame001')-energy('Ar32_frame002')
Weight: 1.0
Sigma: 0.054422772491975996
ReferenceValue: 0.20395942701637979
Unit: eV, 27.211386245988
---
Expression: energy('Ar32_frame003')-energy('Ar32_frame002')
Weight: 1.0
Sigma: 0.054422772491975996
ReferenceValue: 0.22060005303998803
Unit: eV, 27.211386245988
---
Expression: forces('Ar32_frame001')
Weight: 1.0
Sigma: 0.15426620242897765
ReferenceValue: |
These were the reference values that were used in the Getting Started: Lennard-Jones tutorial.
5.3.5. Generate reference values in Python¶
See run.py
, where first the GenerateReference job is run, and then the resulting reference values are used to run an optimization. See also ParAMSJob and ParAMSResults.
#!/usr/bin/env amspython
from scm.plams import *
from scm.params import *
import os
def main():
init()
inputfile = os.path.expandvars("$AMSHOME/scripting/scm/params/examples/LJ_Ar_no_reference_data/params.in")
job = ParAMSJob.from_inputfile(inputfile, name="genref")
job.run()
# use all settings from the LJ_Ar example except for the path to the training set
inputfile = os.path.expandvars("$AMSHOME/scripting/scm/params/examples/LJ_Ar/params.in")
opt_job = ParAMSJob.from_inputfile(inputfile, name="optimization")
opt_job.training_set = job.results.get_training_set_ref_path()
opt_job.run()
loss = opt_job.results.get_loss(source="latest")
print(f"Final loss: {loss}")
finish()
if __name__ == "__main__":
main()