how do we tie all threads together?

[kjappelbaum]

I feel we're now going a bit all over the place and it will be hard to structure a coherent study. The initial "pitch" notes were structures around "forward classification" and "inverse design". Now we also added regression.

In total we have for the "forward" predictions

• polymers, MOFs, photovoltaics, and photoswitches as case studies. We do classification and regression for them. Where available we also compare different representations.

For the "inverse" route we currently only have the polymers.

[kjappelbaum]

I feel we do not want to have the best model for any given application - we do not tune enough (and it makes no sense to spend resources on that) - and the key message is probably LLM are so powerful that they can perform competitively with strong baselines on various chemistry and materials science applications without fine-tuning of representations or models.

We can probably show this in a really short paper of Andrew White style where we just

[kjappelbaum]

i.e. one could maybe organize it around one figure illustrating the results for classification, another one for regression, and then one for inverse design.

[pschwllr]

This is probably the most important point we should discuss.

We could initially submit a 4-page paper to:

• AI4Mat (NeurIPS workshop): https://sites.google.com/view/ai4mat (deadline Sept 26) and
• the virtual ML for Molecules workshop: https://moleculediscovery.github.io/workshop2022/ (deadline October 18)

Both of the workshops required the same NeurIPS style format for the submission.