7.10.1. Architecture Quick Reference¶
7.10.1.1. Diagram¶
The following diagram is intended as a quick reference to aid in the building / debugging of personal workflows with ParAMS, listing main input-output interactions between classes. Boxes bordering from the left are inputs to the ones on the right. For more information about a specific class, see its respective section.
+--------------------+--------------------+--------------------+--------------------+
| plams.Settings() | | | |
+--------------------+ JCEntry() | JobCollection() | |
| plams.Molecule() | | | |
+--------------------+--------------------+--------------------| |
| | |
+--------------------+ DataSet() | |
| extractor.py | | |
+--------------------+--------------------| |
| | |
|ParameterInterface()| |
| | |
+--------------------| Optimization() |
| | |
| Optimizer() | |
| | |
+--------------------| |
| | |
| ParallelLevels() | |
| | |
+--------------------| |
| | |
| Loss() | |
| | |
+--------------------| |
+--------------------+--------------------+--------------------+
| plams.Settings() | | |
+--------------------+ Engine() | EngineCollection() |
| | |
+--------------------+--------------------+
7.10.1.2. Simple usage of some classes¶
Launch $AMSBIN/amspython, and import all modules from scm.params:
>>> from scm.params import *
There are three main components to the package:
The Job Collection, which represents a set of chemical systems stored as plams.Molecule objects alongside AMS settings specific to each system. The combination of system and settings clearly defines a computational job. An example for a job could be a geometry optimization of a methane molecule.
A Data Set, which holds the definition of all properties that are relevant for the parameter optimization. These could be forces, energy and geometry of the aforementioned methane job. The Data Set is responsible for the extraction and comparison of all properties to their reference values, providing a loss value metric.
A parameter interface, which serves as a translation layer between an AMS engine and a list of parameters. Any available parameter interface in ParAMS can be used for re-parameterization.
Components interact with each other as follows:
>>> print(jc) # jc is a JobCollection instance
---
ID: Water01
ReferenceEngineID: None
AMSInput: |
properties
gradients
End
system
Atoms
O 0.0000000000 0.0000000000 0.5937200000
H 0.0000000000 -0.7654400000 -0.0083600000
H 0.0000000000 0.7654400000 -0.0083600000
End
End
task geometryoptimization
...
>>> print(ds) # ds is a DataSet instance
---
Expression: energy('Water01')
Weight: 1.0
ReferenceValue: -0.09210963293745549
---
Expression: forces('Water01')
Weight: 1.0
ReferenceValue: |
array([[-4.41310133e-11, 1.24966204e-10, -1.38738596e-05],
[-2.19091605e-11, -2.01034720e-05, 6.93698245e-06],
[-2.19091605e-11, 2.01036055e-05, 6.93698245e-06]])
...
>>> params = LennardJonesParameters() # params is a ParameterInterface instance
>>> params.names # parameter names
['eps', 'rmin']
>>> print(params.x) # parameter values
[0.0003, 3.0]
>>> jobresults = jc.run(params) # calculate all jobs in `jc` with our parameters
>>> loss = ds.evaluate(jobresults)
>>> print(loss) # How close are the results calculated with `params` to their reference values?
2082.088800689759
>>> params.x = [0.0004, 4.] # change the parameter values ...
>>> jobresults = jc.run(params) # ... re-calculate ...
>>> loss = ds.evaluate(jobresults) # ... and re-evaluate.
2066.4096600168773
Note how in the above code block we are manually assigning different parameter values to our params instance. This is the entry point to the fourth component in ParAMS - the optimizer. Piecing it all together is a top-level Optimization class, which is managing the parameter fitting process based on user-defined settings. A simplified overview of the ParAMS parameterization loop is depicted in the figure below.
For a hands-on dive into the workings of the package, we recommend checking out the Tutorials: Getting Started section (and trying them yourself). If at any point you stumble upon a class, whose functionality is not entirely clear, make sure to look it up in the respective Python Classes and Functions subsection. Most of these provide additional examples on the component’s functionality on top of introducing the complete API.