Queries

• How to build useful machine-representations of atoms?
  • $R^1$ would just be a 1D vector embedding, likely of little use. But are there taxonomies of representations (including a matrix per atom?)

Missing

• Splits for datasets, validation / testing.
• Balancing datasets.
• Evaluation metrics, loss functions,
• ROC classification performance metric

Ideas

• Train and show the results of atoms for vectors in a website? With some button to load each dataset?

• List of examples of successful applications of ML in chemistry

  1. Anything for which there are useful datasets from experiments or from calculation:
     • orbital energies,
     • geometries,
     • formation energy,
     • energy band-gap

• Which models are related to "Causal Inference"? Are they useful? For example The Mythos of Model Interpretability states (removed a reference which is to Judea Pearl's Causality book, and may be worth checking):

The task of inferring causal relationships from observational data has been extensively studied. Causal inference methods, however, tend to rely on strong assumptions and are not widely used by practitioners, especially on large, complex data sets.