| 26 Jun 2024 |
 mjwaters | MLIPs cost human time, keep that in mind. | 18:48:30 |
| mjwaters | but linear scaling. | 18:48:46 |
| mjwaters | they need training data, it's usually good to have at least one structure per parameter. | 18:51:13 |
| mjwaters | so more chemical complexity will mean more training data. | 18:51:57 |
| mjwaters | In reply to @mjwaters:matrix.org
they need training data, it's usually good to have at least one structure per parameter. this is an order of magnitude estimate | 18:53:05 |
| mjwaters | So you'll need a DFT data set to train on. | 18:53:34 |
| mjwaters | typically smaller structures that capture enough interactions that they are still useful when representing bulk. | 18:54:43 |
 Hasith Vattikuti | Do you have any good references on hand for training MLIPs, I've worked with ML models before so I have a few ideas of how it might work, but seeing established practices would probably be much better than anything I think of in a few minutes | 18:57:09 |
| mjwaters | mschaefer: Has a great list. | 18:59:52 |
| 27 Jun 2024 |
 mschaefer | In reply to @mjwaters:matrix.org
mschaefer: Has a great list. oh that list is super outdated at this point | 20:46:58 |
| mschaefer | In reply to @hasithv:matrix.org
Do you have any good references on hand for training MLIPs, I've worked with ML models before so I have a few ideas of how it might work, but seeing established practices would probably be much better than anything I think of in a few minutes I mean are you looking at practical aspects or the fundamentals of how models work? The reading list was mostly about the latter.
For the former you can look at the reviews from the Klaus Robert Müller or Jörg Behler groups but there are probably dozens out there at this point | 20:51:50 |
| mschaefer | the practical recommendations are somewhat model dependent and what you want to use it for. So it's hard to give general recommendations other than make sure you follow datascience 101 practices, understand the model (and it's limitations) and the application you have in mind. Also always make sure your dataset sufficiently samples the configurational space you are interested in | 20:53:46 |
| Hasith Vattikuti | I guess something helpful would be to see an example of an mlip implementation from generating the dataset to training the model, do you know of any work that documents this? | 20:56:16 |
| mjwaters | I have a Zr/O data set shared online if you want something realistic to play with. Alternatively, you can run MD with the ASE EMT potential and some metal elements to make some cheap practice data. | 21:28:09 |
| mjwaters | mschaefer: Which code is the easiest to start with? | 21:28:56 |
| mjwaters | How is your code going? | 21:29:16 |
| Hasith Vattikuti | In reply to @mjwaters:matrix.org
I have a Zr/O data set shared online if you want something realistic to play with. Alternatively, you can run MD with the ASE EMT potential and some metal elements to make some cheap practice data. Yes, this would be be a perfect reference! Can you please send a link to the dataset? And, if it is available, the calculator you built as well? | 21:33:10 |
| mjwaters | I didn't build a calculator for that work. | 21:46:16 |
| mjwaters | Hasith Vattikuti: https://zenodo.org/records/8264492 | 22:04:52 |
| mschaefer | In reply to @mjwaters:matrix.org
mschaefer: Which code is the easiest to start with? gotta say, I have not used many packages other than our own but every time I do I am quite happy with what we have
https://github.com/apax-hub/apax | 22:09:09 |
| mschaefer | In reply to @mjwaters:matrix.org
How is your code going? going pretty well i would say. We got plenty of stuff done in the past 6 or so months. finally writing a paper about it we'll hand it in some time after I return to Germany | 22:11:04 |
| mjwaters | good to hear! | 22:11:39 |
| mjwaters | If I start another MLIP project, I'll be trying out apax. | 22:12:46 |
| mschaefer | let me know what you think :D | 22:14:09 |
| 28 Jun 2024 |
 Tristan Maxson | In reply to @mjwaters:matrix.org
I have a Zr/O data set shared online if you want something realistic to play with. Alternatively, you can run MD with the ASE EMT potential and some metal elements to make some cheap practice data. I made some cheap EMT-D3 data once that I should zenodo for this purpose. | 04:45:25 |
| Tristan Maxson | Slapping D3 on it makes MLIPs sweat a little more | 04:48:18 |
 alin m elena🍀 | I will use ase mace_mp as starting point will make your life easier | 09:12:44 |
| alin m elena🍀 | https://github.com/ACEsuit/mace-mp/releases/tag/mace_mp_0b | 09:13:30 |
| alin m elena🍀 | or chgnet or m3gnet or any other foundation models | 09:14:11 |
 Nils Phillip Winkelmann (TUDa) | Is there a way to get auto completion working? (Using pycharm) | 09:14:20 |
