Hi!
First of all, thank you for releasing the MatRIS foundation potentials! I am reaching out for access to the raw, intermediate MatRIS-OMat24 checkpoints prior to their fine-tuning on the sAlex and MPTrj datasets.
I am currently working on molecular dynamics simulations of molten salts. While evaluating the production MatRIS models (and also other production OAM models such as the UMA models from FAIRCHEM, I noticed that their performance was much worse on my systems than the ones trained on OMat24 alone. Because molten salts feature highly dynamic coordination environments and out-of-equilibrium liquid structures, OMat24 dataset provides an ideal baseline, I suspect the raw OMat24 pre-trained weights retain a broader generalizability for these disordered liquid phases, whereas the fine-tuned OAM versions may suffer from a slight regression or "concept forgetting" regarding high-temperature liquid configurations.
Would you be open to hosting or sharing the raw, pre-trained checkpoints (on figshare, huggingface, whatever you prefer)?
Thank you very much for your time and support!
Best regards,
Daniel
Hi!
First of all, thank you for releasing the MatRIS foundation potentials! I am reaching out for access to the raw, intermediate MatRIS-OMat24 checkpoints prior to their fine-tuning on the sAlex and MPTrj datasets.
I am currently working on molecular dynamics simulations of molten salts. While evaluating the production MatRIS models (and also other production OAM models such as the UMA models from FAIRCHEM, I noticed that their performance was much worse on my systems than the ones trained on OMat24 alone. Because molten salts feature highly dynamic coordination environments and out-of-equilibrium liquid structures, OMat24 dataset provides an ideal baseline, I suspect the raw OMat24 pre-trained weights retain a broader generalizability for these disordered liquid phases, whereas the fine-tuned OAM versions may suffer from a slight regression or "concept forgetting" regarding high-temperature liquid configurations.
Would you be open to hosting or sharing the raw, pre-trained checkpoints (on figshare, huggingface, whatever you prefer)?
Thank you very much for your time and support!
Best regards,
Daniel