Luis Antunes

Dream up a brand-new material in 30 seconds, no PhD required. What once required years of training and expensive lab equipment now happens with a text prompt. CrystaLLM from @_lantunes and collaborators just transformed Ouro into every citizen scientist's materials lab. Simply describe a chemical system and CrystaLLM generates the crystal structure as a CIF file. Visualize your creation instantly, then connect it to other user-added services to explore its properties and potential. Now anyone can propose hypotheses, generate structures, and iterate in real-time. The future of materials discovery is collaborative, open, and a lot more fun. s/o @modal_labs for supporting open science, GPU-accelerated hosting, and startup credits.

@lateinteraction Here's an example of a small, specialized LM for crystal structure generation: nature.com/articles/s4146…. The small version of the model has 25M parameters. It's not trained on natural language, but a specialized language for describing crystal structures.

I’m surprised we’re not really seeing research into “micro-LMs” or something like so. You can probably build out a whole ecosystem of software architectures based on aggressively specialized 30M-parameter LMs or something, right? Wonder how small they can get and be useful.

@taylordsparks @rgraucrespo @keeeto2000 @lonepair @K_Shankland Thank you so much, @taylordsparks!

@rgraucrespo @_lantunes @keeeto2000 @lonepair @K_Shankland Congrats Luis! He's done remarkable work during his PhD.

@vedviki @rgraucrespo @keeeto2000 @lonepair @K_Shankland Thank you, Vikram!

@rgraucrespo @_lantunes @keeeto2000 @lonepair @K_Shankland Congratulations Luis, well deserved 👏👏

@proupin72 @rgraucrespo @keeeto2000 @lonepair @K_Shankland Thank you!

@rgraucrespo @_lantunes @keeeto2000 @lonepair @K_Shankland By chance I was reading his/your article on crystaLLM. Congrats!

@rgraucrespo @keeeto2000 @lonepair @K_Shankland Thank you @rgraucrespo and @keeeto2000 for being unwavering partners on this voyage! And a big thanks once again to @lonepair and @K_Shankland for examining the viva!

Big shoutout and heartfelt appreciation to the team behind CrystaLLM. I tried the web application, and the results are awesome. 👏

@SterlingBaird1 @keeeto2000 @rgraucrespo @crystalmooc Thank you @SterlingBaird1! It will be awesome to see the community build upon and improve the model, now that the code has been released.

@_lantunes @keeeto2000 @rgraucrespo 🔥Well done, Luis! Agreed with @crystalmooc that the web application at crystallm.com is superb👌 I gave it a try with some superconductivity candidate compositions we screened some time ago. The tool looks promising! See the video below

CrystaLLM source code, weights and datasets have been released! We've also updated the preprint with benchmarking results, an SI, & more. Code: github.com/lantunes/Cryst… Preprint: arxiv.org/abs/2307.04340 In collaboration with @keeeto2000 & @rgraucrespo

@QAISALI12 @keeeto2000 @rgraucrespo Unfortunately, the model does not currently generate disordered structures. However, this is something we plan to support in the future.

@chc273 @keeeto2000 @rgraucrespo Thank you!

@_lantunes @keeeto2000 @rgraucrespo Nice work! @_lantunes

Can LLMs be used to generate plausible crystal structures? Read an accessible overview of CrystaLLM, a generative model of crystal structures: blog.materialis.ai/crystal-struct…. In collaboration with @keeeto2000 & @rgraucrespo

Want to know how the latest large language models might help to generate new materials? Luis developed a great new method, Crystallm and has written an equally great and accessible blog on how it works:

@jrib_ @rgraucrespo @keeeto2000 @mkhorton It's a bit involved to explain by tweet, but basically I execute some JS code when the page loads that maps touch events to mouse events. I'd be happy to share the code.

@_lantunes @rgraucrespo @keeeto2000 That’s great! I’m sure @mkhorton will be happy to hear. What’s your workaround? Two finger swiping works somewhat on mobile but could definitely be improved.

@jrib_ @rgraucrespo @keeeto2000 The model is deployed on @modal_labs serverless GPUs, which I highly recommend!

@rgraucrespo @_lantunes @keeeto2000 Very cool! Two questions: how are you deploying the model and what was your experience with crystaltoolkit? Anything we can improve?

Luis Antunes, @_lantunes, a PhD student at the Chemistry Department of @UniofReading has developed a "chatGPT-like" large language model for crystal structures. The AI model generates detailed crystal structures given a chemical composition.

Can we predict thermoelectric transport properties from composition? Read an accessible overview of our recent work on this topic: blog.materialis.ai/predicting-the…. Test the model yourself: thermopower.materialis.ai. In collaboration with @keeeto2000 & @rgraucrespo

Great new work by @keeeto2000 @rgraucrespo @_lantunes @UniRdg_Chem @QMULSEMS - 'Predicting #thermoelectric transport properties from composition with attention-based #deeplearning' - bit.ly/3ZMqB5s #machinelearning #materials #AI #energy #openscience #datascience #HPC

The latest version of CellPyLib supports Block Cellular Automata, which appear often in @stephen_wolfram's recent series on The Second Law. Learn more here: cellpylib.org/block_ca.html.