Stefano Martiniani

Flow maps?

What if a world model could imagine the future from a completely different perspective? Introducing XVWM: given one view and an action, predict the future from another camera. A building block for theory of mind.
Collaboration with aimlabs.com
📄 arxiv.org/abs/2602.07277

PropMolFlow: property-guided molecule generation with geometry-complete flow matching - Nature Computational Science PropMolFlow is a flow-matching method for property-guided molecule generation that matches diffusion model performance while generating stable, valid structures more quickly and enabling the discovery...

📢Out now! @stemartiniani.bsky.social and colleagues present PropMolFlow, a flow-matching method for property-guided molecule generation. #MoleculeDiscovery #FlowMatching www.nature.com/articles/s43... #chemsky

🔓 rdcu.be/eZ5cG

a close up of a person 's hand holding a marker that says sharpie Alt: a close up of a person's hand holding a marker that says sharpie

New preprint!

So, say you're studying some critical transition. How do you catch its universality? Pair correlations? Boring!

We threw line segments at the system, looked at intersections with clusters, and uncovered static and dynamical universal behavior of MIPS!

arxiv.org/abs/2511.09444

Illustration of a 60-fold gyromorph's properties. Top row: Structure of the gyromorph. Left: Structure factor. Right: Pair correlation function. Bottom row: Evidence of a bandgap. Left: Scalar optical field inside the gyromorph. Right: Density of states depletion in the gyromorph.

New paper just out, as an editor's suggestion in PRL!

While looking for the ideal isotropic bandgap material, we actually discovered new structures.
These structures lie at the border between order and disorder, and that's good for optics!

More about their structure here,
tinyurl.com/3aej53ht

⚛️🧪

The transformative capability of quantum-accurate machine learning interatomic potentials Commentary: Many materials' properties and phase boundaries are generally not well known under extreme pressure and temperature conditions. This is a consequence of the scarcity of experimental inform...

The transformative capability of quantum-accurate machine learning interatomic potentials

Kim Review Commentary by Alfredo A. Correa; Sebastien Hamel

kimreview.org/commentaries...

All that structure matches does not glitter Generative models for materials, especially inorganic crystals, hold potential to transform the theoretical prediction of novel compounds and structures. Advancement in this field depends critically o...

If everyone does it, it must be right…right? Not quite. In “All That Structure Matches Does Not Glitter” #NeurIPS2025 we show CSP benchmarks miss polymorphs and datasets are duplicated. New deduped data, polymorph-aware splits, METRe & cRMSE. Harder tasks, better models!
www.arxiv.org/abs/2509.12178

Scientists Find Curvy Answer to Harnessing “Swarm Intelligence” Breakthrough offers way to develop AI to match flocking birds and schooling fish

Check out our latest paper in collaboration with Mathias Casiulis, Naomi Oppenheimer, and Matan Ben Zion on a simple geometric design rule to achieve robotic swarm intelligence. The paper is out today in the Proceedings of the National Academy of Sciences (PNAS).

www.nyu.edu/about/news-p...

Contrastive Self-Supervised Learning is Just Sphere Packing!
CLAMP (Contrastive Learning As Manifold Packing) recasts SSL as neural manifold packing with a physics-inspired repulsive-particle loss (like in jamming) and achieves new SOTA on ImageNet-100. arxiv.org/abs/2506.13717

🔬 Here comes the lab’s first neurobiology paper, made possible by the… | Stefano Martiniani 🔬 Here comes the lab’s first neurobiology paper, made possible by the amazing work of Dr. Jiyeon H., grad student Asit Pal, and collaborators, especially Andre Fenton (lead PI on the paper), Hans A. H...

www.linkedin.com/posts/smarti...

🔬 Here comes the lab’s first neurobiology paper, made possible by the… | Stefano Martiniani 🔬 Here comes the lab’s first neurobiology paper, made possible by the amazing work of Dr. Jiyeon H., grad student Asit Pal, and collaborators, especially Andre Fenton (lead PI on the paper), Hans A. H...

www.linkedin.com/posts/smarti...

Persistently increased expression of PKMζ and unbiased gene expression profiles identify hippocampal molecular traces of a long-term active place avoidance memory and ′shadow′ proteins Long-term memory formation transiently activates Ca2+-calmodulin kinase IIα (CaMKII) and atypical protein kinase C isoform iota/lambda (PKC𝜄/λ), whereas persistent activation of the other atypical PKC...

We show that memory persistence is encoded in gene expression manifolds, not single gene changes. Shadow memory proteins like PKMzeta & KIBRA leave no single-gene signature, but reshape network structure. New from our lab + Hofmann + Fenton lab: www.biorxiv.org/content/10.1...

Emergent universal long-range structure in random-organizing systems Self-organization through noisy interactions is ubiquitous across physics, mathematics, and machine learning, yet how long-range structure emerges from local noisy dynamics remains poorly understood. ...

🚀 Satyam and Guanming’s “Emergent Universal Long Range Structure in Random-Organizing Systems” shows noise correlations create long-range structure, from 🧩 hyperuniform materials to 🤖 ML, and that SGD’s flat minima bias is universal. 👇 arxiv.org/abs/2505.22933 #SoftMatter #ML

The Martiniani Lab

Left to right: Dr. M. Casiulis, Dr. (as of today!) A. Shih , S Rawat, Dr. J. Han, Dr. K. McClain, E. House, Dr. G. Zhang, ..., Dr. P. Hoellmer, T. Egg, S. Anand, A. Pal, P. Suryadevara, G. Wolfe, Dr. M. Martirossyan, (Dr. F. Morone)

Stabilization of recurrent neural networks through divisive normalization Stability is a fundamental requirement for both biological and engineered neural circuits, yet it is surprisingly difficult to guarantee in the presence of recurrent interactions. Standard linear dyna...

🚀 New paper on stabilizing recurrent neural circuits! Normalization keeps recurrent networks in check. When it fails: ⏳ critical slowing, 🎲 variability ➡️ 🌪️ oscillations➡️💥 instability. Important for understanding brain functions and building AI. www.biorxiv.org/content/10.1...

📄 More info: openreview.net/pdf?id=ka2jx...

🧪 Fully open-source on GitHub: github.com/FERMat-ML/OM...

🙏 Thanks to all contributors! 💻 Trained on EmpireAI and NYU, UF, & UMN HPC.

🚀 Thrilled to introduce Open Materials Generation (OMatG), a state of the art framework for generative design of inorganic crystalline materials! Accepted at #ICML2025 & Spotlight at #AI4Mat @ICLR2025!

🔬 OMatG unifies flow matching & score-based diffusion, outperforming FlowMM and FlowLLM!

RamO(N) 🤣

as.nyu.edu/departments/...

On Tuesday, March 25, Stefano Martiniani will give an #AI for Science Seminar on “Learning as Manifold Packing” in room 414 AGH, hosted by the Data Driven Discovery Initiative (DDDI) and the Center for Innovation in Data Engineering and Science (IDEAS). Join us!
web.sas.upenn.edu/da...

From 2010 to 2016 (latest data I have ), NIH research contributed to EVERY drug approved by the FDA

Nice!

Do systems where the equations are known but cannot be solved in less than exponential time count? If so just take Schroedinger's equation for an interacting many-body system. Perfect description of the problem with no solution :)

Oh and if you think "but surely hydrodynamics was derived from atomistic theories", think again, a lot of hydrodynamics (e.g. Flick's law) was derived phenomenologically (in neuro language, normatively)

So do we need to know the state of every cell in the brain to know the state of the brain? I hope not! We can only understand and predict with coarse grained theories.

In principle this would work but the size of the system of equations that one would have to solve and the time required to solve them makes it 1/ impossible 2/ a dumb proposition because we have a coarse-grained theory (fluid dynamics and continuum mechanics) which are better suited for this problem

I think more to your point, and not unrelated to my previous example, take the lift of a plane. Someone fixated with microscopic details would argue that to model a plane one would have to run a molecular dynamics simulation of every atom in the plane and the surrounding air.

Do systems where the equations are known but cannot be solved in less than exponential time count? If so just take Schroedinger's equation for an interacting many-body system. Perfect description of the problem with no solution :)

CDS is hiring a Clinical Professor of Data Science.

Teach ML, programming, and specialized courses in our 60 5th Ave building.

Renewable contracts with promotion opportunities.

Apply by April 1, 2025.

For details, see: apply.interfolio.com/155349

#MachineLearning #ML #AIjobs

ColabFit Exchange is another great dataset curation effort that I'd like to boost.

Great work by @stemartiniani.bsky.social and team to curate the most diverse materials database in the world!