How to train a model on QM9 dataset?

[Jevon-Du]

Hi guys,

I want to train a model to predict HOMO energy using the QM9 dataset but I'm having trouble finding relevant documentation. Can you guide me on how to proceed?

[misko]

Hi!,
I don't have a detailed answer for you, but I think a rough approach might be as follows,

1. Convert QM9 into a usable data format (ASEdb or ASELMDB)
2. Add an output for HOMO energy , or easier, just use the already existing energy scalar to represent HOMO energy
3. Run training

For (1) there are two existing issues in the repo which we have not solved yet that might provide some insight, #788 , #787 , https://fair-chem.github.io/core/ase_dataset_creation.html

For (3) I think this might be a good start. Depending on what data format you use you can ignore parts relating to LMDB. To use ASEdb format you change the config slightly to specify format: ase_db (as is done in the example linked)

Hope this helps! If you make some progress and get stuck please reach out here, I would be happy to help! if you are interested we could also use your approach in the tutorial, and save other people asking the same question some time :)

[itsuwari]

I have done some fine-tuning on QM9 dataset, found no distinct advantage over GFN-xTB2 and other fine-tuned models. It could be EqV2 was trained for periodic systems and it's not the best choice for representing the interaction among light atoms.

[lavenderwfy]

I have done some fine-tuning on QM9 dataset, found no distinct advantage over GFN-xTB2 and other fine-tuned models. It could be EqV2 was trained for periodic systems and it's not the best choice for representing the interaction among light atoms.

Hi there! I'm currently engaged in similar work as well. When it comes to the task of completing data format conversion, specifically converting the QM9 dataset into a usable data format like the ASEdb or ASELMDB, I was wondering if you could be so kind as to share your relevant code with me? It would be of great help to my work, and I really appreciate it in advance.