Melting temperature from Magpie

Dear Matminer team,

While trying to featurize ceramic materials using the class matminer.featurizers.composition.ElementProperty with the preset called magpie, I found that it gives incorrect melting temperature (e.g., 406 K for Al2O3, of which melting T is ~2345 K), which is due to its way of estimating elemental properties.
I guess Matminer team is also aware of this issue, but am reporting this, just in case.
Thanks much!

Hey there,

We are aware of this, but thank you for letting us know. The MagPie featurizer is able to generate attributes from arbitrary compositions by taking statistics over elemental features. For example, the Al2O3 “mean melting temperatuere” is just the weighted average of the Al and O elemental melting temperatures ((933x2 + (54x3))/5 = 406), not the actual melting temperature of Al2O3. Of course, generating the real melting temperature would be more useful… but we do not have a featurizer to do this for arbitrary compositions, so it’s not as useful for machine learning

Thanks,
Alex

1 Like

Thanks Alex for the explanation!

1 Like

Sure thing!

All that being said, if you have a model/featurizer useful for predicting melting temperatures of arbitrary compounds, we may be interested in including it in matminer!

Thanks,
alex

Hi all,

Just want to add that the Magpie features are almost never meant to be a final estimate of a property, but rather just one descriptor that is used as part of a larger model to predict a property. Thus, I would discourage ever interpreting Magpie predictors as “predictions”, but rather as building blocks that provide one piece of information.

Rather, you might want to do something like use our “automatiner” code to automatically train a full ML model to make predictions using a melting point data set (which would need to be provided).