Nice joke but not for long :)

Wait for the release of #elagente Pre-signup at elagente.ca Like a #Toronto subway line we are building and testing the scalable infrastructure but it is coming closer and closer every day.

@variniabernales.bsky.social @thematterlab.bsky.social #chemsky #compchemsky

EGMOF demonstrates how hybrid architectures can bridge data scarcity and structural complexity, marking a step toward universal, data-efficient AI for materials discovery.
[6/6]

This two-step, interpretable design allows EGMOF to:

- Achieve 95% validity and 84% hit rate for hydrogen uptake targets
- Work robustly even with just 1,000 training samples
- Generalize across 29 computational and experimental datasets: including CoreMOF, QMOF, and even text-mined datasets
[5/6]

Instead of directly generating complex structures, EGMOF introduces a descriptor-based modular workflow:

- Prop2Desc: a diffusion model that maps target properties to chemically meaningful descriptors

- Desc2MOF: a transformer that reconstructs full MOF structures from those descriptors.
[4/6]

To address this, EGMOF (Efficient Generation of Metal–Organic Frameworks): a hybrid diffusion–transformer framework is developed that rethinks how AI approaches materials generation.
[3/6]

Designing new materials with desired properties is one of the hardest problems in materials science. While generative AI has revolutionized image and text creation, applying it to materials remains difficult. Property-labeled datasets are extremely limited, and obtaining them is costly.
[2/6]

Kudos to the authors: @gkwt.bsky.social, Edwin Yu, @narukiyoshikawa.bsky.social, @valencekjell.com, and @aspuru.bsky.social
[4/4]

We also created new molecular design tasks that are explicitly dependent on the stereochemical information of the molecules: chiral molecule rediscovery, ligand-protein docking and circular dichroism spectra optimization.
[3/4]

We compared several string-based generative modeling approaches and found that including stereochemical detail can help in tasks where 3D shape matters, though it also makes the search space more complex.
[2/4]

Even better, our Hessian readout head is simple and can be added to any of your favorite equivariant MLIPs ✅

The brilliant team:
@andreasburger.bsky.social, Luca Thiede, Nikolaj Rønne, @variniabernales.bsky.social, Nandita Vijaykumar, @tvegge.bsky.social, Arghya Bhowmik, @aspuru.bsky.social
[4/4]

Compared to MLIPs with autograd, we achieve:

➡️ 2x lower error
➡️ 70x faster inference, more efficient parallelism, and better scaling with system size ⚡
➡️ Consistently higher success rates on all downstream tasks
[3/4]

In our new preprint, "Shoot from the HIP: Hessian Interatomic potentials without derivatives", we show that we can directly predict symmetry-preserving Hessians using an equivariant Hessian redout head.

📃 arxiv.org/abs/2509.21624
🖥️ github.com/BurgerAndrea...
[2/4]

Kudos to authors: Danial Motlagh, Robert A. Lang, Paarth Jain, @ja-camga.bsky.social, William Maxwell, Tao Zeng, @aspuru.bsky.social, and Juan Miguel Arrazola.
[8/8]

- Resource estimates for anthracene-based chromophores highlight the path toward quantum-enabled discovery of new singlet fission materials.
[7/8]

- Key observables such as electronic state populations can be extracted simply by measuring the electronic register, bypassing the input–output bottleneck common in many quantum algorithms.
[6/8]