zincware / ipsuite Goto Github PK

If you change AddData the ConfigurationSelection will rerun, but because the output is just a list of indices, e.g. [1, 2, 3] the UpdateAtoms might not be rerun automatically!

flowchart TD
        node1["AddData"]
        node2["ConfigurationSelection"]
        node3["UpdateAtoms"]
        node1-->node2
        node2-->node3

We need lazy loading here, because apax can not be installed. This break IPS at the moment.

The ModelAnalysis Node only analyses the force magnitude. We should add a metric and plot for force components (all in one plot should be sufficient)

Some calculators, e.g. CP2K or XTB require a conda setup for tests.
Maybe we can parametrize and run these tests in a dedicated runner.

    test_data = []
    train_data = []
    test_data.append(
        ips.configuration_selection.RandomSelection(data=geopt, n_configurations=50)
    )

    train_data.append(
        ips.configuration_selection.RandomSelection(data=geopt, n_configurations=50, exclude=test_data)
    )

support something like this via:

        if self.exclude is not None:
            if self.exclude_configurations is None:
                self.exclude_configurations = {}
            if not isinstance(self.exclude, list):
                self.exclude = [self.exclude]
            for exclude in self.exclude:
                for key in exclude.selected_configurations:
                    if key in self.exclude_configurations:
                        self.exclude_configurations[key].extend(
                            exclude.selected_configurations[key]
                        )
                    else:
                        self.exclude_configurations[key] = exclude.selected_configurations[
                            key
                        ]

Replace the prediction dataclass by a fields.Atoms

DVC actually supports comparing images

• https://dvc.org/doc/user-guide/experiment-management/visualizing-plots#supported-plot-file-formats

IPS doesn't have docs.

Default to None and do dynmically in post_init + post_load.

It should be easy to have a Benchmark Dataset, QM9, MD17, ... as a repository that one can clone and then work with.
Currently the limiting factor is ZnH5MD periodic pbc and chaning size of atoms.

forces are alreay per atom - check the analyseprediction node for force errors

Can't do IndexSelection(index=[-1])

Have the following scheme:
MD with UncertaintyCheck starting from different configurations
UncertaintyCheck stops, then use ItemSelection(item=-1) and run DFT.
Retrain Model on all the data points

Repeat as often as you wish.

As discussed:

self.status = f"Temperature Check {temperature} < {self.max}"
# error
self.status = f"Temperature Check failed: last {temperature} > {self.max}"
# und dann
def __str__(self):
   return self.status 

We could provide Node Groups, that define multiple nodes together. E.g. metrics could be predict, metrics, force decomp all together.

Discussion: Furthermore, we don't really need PredictEithModel as it is basically Singlepoint where the calc is the model. Then we pass true/pred to the metrics node. This would probably make uncertainty easier as well, because we don't need the prediction class anymore. This would be easier with Node Groups.

If you switch from a MLModel (or any Node) that uses params.yaml to one that does not use params.yaml at all, the parameter in params.yaml remain and can cause issues (ZnTrack Problem)

The IPS Analysis module starts to get really crowded.
Because these Nodes are some of the most important Nodes of IPS I think we should put them in some sort of order.

It's also important to mention, if the Node location is changed inside IPS, the respective projects using these Nodes will have a changed cmd zntrack run ipsuite.analysis.PredictWithModel --name PredictWithModel in the dvc.yaml file and therefore, it will be recomputed.

I'm open for suggestions on how to organize these Nodes inside IPS.

support:

data: HasAtoms | List[HasAtoms] | Atoms | list[atoms]

Clone and run them in the test suite

#47 is missing tests

Have a PyPack Node as a pure-Python drop-in replacement for the Packmol Node.

Maybe we can add http://m3g.iqm.unicamp.br/packmol/nmols.shtml to get a density approximate?

Maybe a thermostat that can do that?

Would replace

Sometimes you want to run a ProcessSingleAtom for multiple configurations.
Currently this is very difficult so I'd propose:

method = ips.analysis.BoxScaleAnalysis(data=None)

with project:
   data = ips.AddData()
   volume_scan = ips.Map(data=data, method=method)

project.run()

This should be zntrack.zn.outs

In addition to #81 or in general for all checks allow:

• relative change: if e.g. the value is off by more then 50 % from a rolling mean, stop
• relative change (outlier) take into account the fluctuations

• keep one wfn file and make it an attribute that can be used in other nodes
• use ZnH5MD

Have a minimal amount of configurations that must pass to trigger ThresholdSelection

Check if uncertainties are available and then use them in the plots

Instead of hard coding the inputs for mace create a yaml input file:

e.g. convert --hidden_irreps='128x0e + 128x1o' to

hidden_irreps: 128x0e + 128x1o

gap seems to write train_atoms.extxyz.idx which is not in the .gitignore

You can't import XTB

ImportError: xtb C extension unimportable, cannot use C-API

if nequip (and I guess torch) is imported.

-> We need lazy importing of modules

Hello Maintainers,

I am working with a dataset of around 5000 molecular configurations, which are not necessarily generated through MD simulations. I am interested in employing the "Atomic Energy Selection" methodology, as outlined in this paper, to assemble a dataset for training my machine learning model.

In addition to this, the other methods available in this repository, such as the kernel selection method, seem to align well with my use-case requirements.

To better understand the implementation and to ensure I am utilizing these methodologies correctly, it would be greatly beneficial if you could provide an example code for the Atomic Energy Selection and Kernel Selection methodologies.

I have also spoken about this issue to Samuel Tovey via email.

We could add a Node, that on a condition, e.g. AL doesn't produce new configurations stops the graph by raising an Error.

We should check why it is so slow:
is it ZnTrack or other modules?

• add an option to bin and randomly select from bins instead of the current method
• add uniform force selection

          should we address this in an other PR?

Originally posted by @Tetracarbonylnickel in #81 (comment)

Just like with data we can also provide trained models for faster tests

avoid

Would this be required in some scenarios?

Add an energy + uncertainty plot to the uncertainty selection node where selected points are marked

As mentioned in #104 the SOAP parameters currently contain the periodic key. This should not be a parameter but defined by the data.

ConfigurationSelection_1_kernel:
    soap:
        _type: soap_parameter_dataclass
        value:
            l_max: 7
            n_jobs: -1
            n_max: 7
            periodic: true
            r_cut: 9.0
            rbf: gto
            sigma: 1.0
            weighting: null

Make all nodes to be imported from ips.nodes.<name>

The can still be in their respective modules but use module overwrite and lazy loading.