
PhD candidate at @McGillU and @Mila_Quebec Working on AI for Science
Joined June 2022
- Tweets15
- Following323
- Followers352
- Likes49
Pinned Tweet
Super excited to share what I've been working on with @thevineetjain: we develop a method to steer generative models for materials using tree search! We'll be presenting the paper at @NeurIPSConf in Sydney.
ArXiv 🔗: arxiv.org/abs/2611.17462
The idea (🧵) :
The result: we get more out of our base models by scaling up inference-time compute. That means more stable crystals, more crystals with rare properties, and better Pareto fronts for multi-objective targets
1
1
44
@thevineetjain and I worked on this with our awesome collaborators at @Mila_Quebec: @tara_aksa, @sekoumarkaba, and Siamak Ravanbakhsh. Check out the paper, and if you're in Sydney, let's chat in person!
Paper: arxiv.org/abs/2611.17462
Code: github.com/danielTLevy/mater…
1
48
If you're interested in connections between AI and physics and are at NeurIPS, come check out the poster for our paper Friday afternoon! Great stuff presented by @sekoumarkaba and @KushaSareen
Prediction of physical configurations is almost always trained with proxy losses like MSE and cross-entropy.
But physics already gives us the right objective: (free-)energy minimization.
In our NeurIPS paper, we show that using approximate energies as losses can drastically boost performance at no extra cost 👇🏾
🧵 1/7
Daniel Tian Levy retweeted
"Equivariance matters even more at larger scales" ~ arxiv.org/abs/2510.09768
All the more reason we need scalable architectures with symmetry awareness. I know this is an obvious ask but I'm still confident that scaling and inductive bias need not be at odds.
This paper (alongside arxiv.org/abs/2410.23179) is convincing evidence that believing "equivariance is dead/not necessary" and "scaling is all you need" might be myopic (ofc, no one has made this *strong* claim but it still seems to be an existing "community myth" of sorts)
Stay tuned to this space – we're dropping something cool on this topic veryyy soon ;)
If you're at #ICLR come check out our poster on symmetric crystal generation! We're at poster #07.
Symmetry is the fundamental property of crystals, yet generative models don't yield crystals with realistic symmetries
We solved that with SymmCD and can get crystals from any of the 230 space groups
Learn more at our #ICLR poster w/@dnllvy @sibasmarak
arxiv.org/abs/2502.03638
🧵
Daniel Tian Levy retweeted
Symmetry is the fundamental property of crystals, yet generative models don't yield crystals with realistic symmetries
We solved that with SymmCD and can get crystals from any of the 230 space groups
Learn more at our #ICLR poster w/@dnllvy @sibasmarak
arxiv.org/abs/2502.03638
🧵
Daniel Tian Levy retweeted
⚛️🤗 Announcing LeMaterial ⚛️🤗
@huggingface & @entalpic_ai are teaming up to release LeMaterial -- an open source initiative aiming to facilitate (AI for) materials discovery !
Datasets, hash function, tools to explore the chemical space & more !
huggingface.co/blog/lemateri…
I'm at #NeurIPS2024 this week, presenting our work SymmCD at the #ai4mat workshop!
openreview.net/forum?id=V7x2…
Come check it out, and feel free to message me to meet up, especially if you want to chat about generative modelling for materials, crystals, and symmetry!
Daniel Tian Levy retweeted
Thrilled to launch EquiAdapt! 🚀
A library for equivariant adaptation of neural networks, including SAM by @AIatMeta, making them robust to transformations without losing efficiency: github.com/arnab39/equiadapt
w/ @ArnabMondal96 @DanieBenes @sekoumarkaba @GagnonJikael
🧵
Daniel Tian Levy retweeted
Want to make your model equivariant to certain known transformations without changing the network architecture? Please attend our contributed talk at 11AM @neur_reps to know how we achieve this by adding a simple shallow network that learns to canonicalize the input. #NeurIPS2022















