Developing AI for true scientific discovery in the lab is challenging, but even more rewarding. Grateful to have so many who share this vision as we take this next big step! 🚀.

PS: The image was generated by Sora and I'm quite impressed by its quality!.

I'll be at #ICML2025 next week and would love to reconnect with friends and colleagues to chat about AI for Science — especially the fun and challenges of discovering novel materials in the lab using ML methods. We're also hiring at @radicalai_inc for several AI research and

RT @mshuaibii: Excited to share our latest releases to the FAIR Chemistry’s family of open datasets and models: OMol25 and UMA! @AIatMeta @….

We are hiring across ML roles! A lot of exciting AI projects are going on @radicalai_inc, feel free to reach out directly for any questions.

RT @josephfkrause: ML-first materials are here: we demonstrate that machine learning approaches to materials prediction achieves state-of-t….

RT @FallettaStefano: Excited for the release of our EGIP model! Pushing the boundaries of efficiency and accuracy 🚀.

Code and checkpoints will be public soon! .Feel free to reach out to me with any questions.

On K_SRME metric, we find that graph construction, precision, and displacements play a key role in accurate predictions. If we just increase the displacements to 0.05 A , we can use much faster models.

The backbone is built off of the GemNet architecture but focussed on efficiency through a combination of ablation studies and strategic pruning. The sum of predicted forces in the model is constrained to be zero.

- EGIP achieves SOTA results on the thermal conductivity prediction task (K_SRME). - It gets a similar Combined Performance Score (CPS) on Matbench Discovery to the recently released eSEN model while being significantly efficient. - It is the only direct force model in Top 15

🚨 Thrilled to share Efficient Geometric Interatomic Potential (EGIP) - an efficient and accurate ML model for materials!. Even though non-conservative MLIPs are not physically grounded, they offer significant efficiency. This work demonstrates that they are quite accurate on.

🚨 Checkout our new MD simulation package!. ✅ Completely written in PyTorch.✅ Handles batched simulations.✅ Easy to add new MLIPs. Led by @ganganabhijeet and @orionarcher_. Spoiler alert: New MLIP coming soon!. Code (MIT License):

RT @ask1729: 1/ Machine learning force fields are hot right now 🔥: models are getting bigger + being trained on more data. But how do we ba….

RT @chaitjo: Introducing All-atom Diffusion Transformers . — towards Foundation Models for generative chemistry, from my internship with th….

RT @Thom_Wolf: I shared a controversial take the other day at an event and I decided to write it down in a longer format: I’m afraid AI won….

RT @radicalai_inc: Discovering and developing a novel material used to take 15-25 years and over $100M for a single materials system. Not a….

RT @jehad__abed: Excited to unveil OCx24, a two-year effort with @UofT and @VSParticle! We've synthesized and tested in the lab hundreds of….

RT @LogConference: LoG 2024 paper decision notifications will be delayed until Saturday, 16 November. The notifications were originally me….

RT @luis_pupuis: Interested in the expressivity of higher-order attention? In our NeurIPS 2024 paper, we show that the Edge Transformer, or….