could you please provide an example to predict disordered structure?

[liuxiaotong15]

hi, I am trying to predict band gap of some disordered structures, but i didn't find any examples of that.

I use:

cg = CrystalGraphDisordered()
graph = cg.convert(test_structures[i])
model_output = model.predict_graph(graph).ravel()

I read some of your code and found that you change the atom number to a dict but I get an error finally:
ValueError: Failed to convert a NumPy array to a Tensor (Unsupported object type collections.defaultdict)

so could you provide an example of that?

thanks

[liuxiaotong15]

Even I load your model like this, it does not work...

model = MEGNetModel.from_file('/home/ubuntu/code/megnet/mvl_models/mf_2020/pbe_gllb_hse_exp_disorder/0/best_model.hdf5')
model_output = model.predict_structure(test_structures[i]).ravel()

Error stack

  File "/home/ubuntu/code/.../python3.8/site-packages/megnet/models/base.py", line 318, in predict_structure
    return self.predict_graph(graph)
  File "/home/ubuntu/code/.../python3.8/site-packages/megnet/models/base.py", line 344, in predict_graph
    inp = self.graph_converter.graph_to_input(graph)
  File "/home/ubuntu/code/.../python3.8/site-packages/megnet/data/graph.py", line 160, in graph_to_input
    expand_1st(self.atom_converter.convert(graph["atom"])),
  File "/home/ubuntu/code/.../python3.8/site-packages/megnet/data/crystal.py", line 130, in convert
    emb += np.array(self.embedding_dict[k]) * v
KeyError: 'Cu+'

Ordered structure work well...

[chc273]

@liuxiaotong15 can you remove the oxidation state and try again?

[liuxiaotong15]

@liuxiaotong15 can you remove the oxidation state and try again?

Thanks. It works on the disordered model you submitted after I add a remove_oxidation_states() to all structure. But still not work for all ordered models.

Can I test disordered structure on previous trained ordered model?
thanks

[liuxiaotong15]

@chc273 I think you use a dict to express an specie in disordered model, so the disorder structure can not be handled by previous trained model?

[liuxiaotong15]

I guess you change the atom in graph from
something like array([[71, 71, 71, 71, 15, 15, 15, 15]])
to something like

array([[defaultdict(<class 'float'>, {'Lu3+': 1.0}),
        defaultdict(<class 'float'>, {'Lu3+': 1.0}),
        defaultdict(<class 'float'>, {'Lu3+': 1.0}),
        defaultdict(<class 'float'>, {'Lu3+': 1.0}),
        defaultdict(<class 'float'>, {'P3-': 1.0}),
        defaultdict(<class 'float'>, {'P3-': 1.0}),
        defaultdict(<class 'float'>, {'P3-': 1.0}),
        defaultdict(<class 'float'>, {'P3-': 1.0})]], dtype=object)

so old model can't handle disordered structure?
Thanks

[chc273]

@liuxiaotong15 again please try remove the oxidation state in ordered structures too

[liuxiaotong15]

@chc273 I think you misunderstanding my point. I have a lot of trained model (by ordered structures). I want to test their performance on disordered structures? Is it possible to do that? thanks.

[liuxiaotong15]

@chc273 sorry for post so many msg. And now I think I have a more easy way to ask my question:
is it possible to replace the graph_converter from CrystalGraph() to CrystalGraphDisordered().
My old model trained by CrystalGraph(), but CrystalGraph() can't handle disordered structure I think. So I need to replace it to CrystalGraphDisordered() I guess.

[chc273]

@liuxiaotong15 you can construct a new model without the elemental embedding layer, but compute the elemental embedding yourself via linear combinations, e.g., what CrystalGraphDisordered does.

Some simple, but exactly the same example is here

https://github.com/materialsvirtuallab/megnet/blob/master/notebooks/model_reconstruct.ipynb

[liuxiaotong15]

@chc273 Yes, it is a good way. thanks.

By the way, I also tested training from graph generated by CrystalGraphDisordered.
It looks a converter should be added, I have submit a PR. The training process runs ok now.
Maybe you have some better place to add the converter.
#321

Merry Christmas and thank for a great MEGNet.

[liuxiaotong15]

@chc273 hi, today I dive into https://github.com/materialsvirtuallab/megnet/blob/master/notebooks/model_reconstruct.ipynb and find it is still not easy to handle the disordered structure. The 'CrystalGraphDisordered' bring in so many troubles...
I still advice you to provide some example on training and inference with 'CrystalGraphDisordered', many thanks...

[liuxiaotong15]

oh, my previous model is created by
MEGNetModel(nfeat_edge=10, nfeat_global=2, graph_converter=CrystalGraph(bond_converter=GaussianDistance(np.linspace(0, 5, 10), 0.5)))
no atom embedding layer in it.... default of nfeat_node is None....
That's why it is really hard for me to modify the model.........