✅ 8 Semiregular (Archimedean) Tilings – Combining two or more polygons while maintaining uniformity—each vertex has the same surroundings!

These patterns are key to geometry, architecture, and—most importantly for us—materials science! 🏗️🔬

#3DPrinting #MaterialsScience

🌍 What’s next?
Many applications:
- Expanding the Alexandria database 🏛️
- Designing materials with tailored properties 🔬
- Accelerating breakthroughs in energy storage & semiconductors

We’re just getting started!

🏆 Results we’re proud of:
- 8x more likely to generate stable structures than baselines (e.g., PyXtal with charge compensation)
- Fast: 1,000 novel structures/min ⚡
- Control over space group, composition, and stability
- Releasing 3 million compounds generated by the model 📥

💡 What makes it unique?
- Fully leverages Wyckoff positions (discrete + continuous parameters)
- Trained across the periodic table & 230 space groups
- Condition on critical properties like stability

🧪 Why this matters:
Materials are the foundation of modern technology—fueling everything from batteries to semiconductors.

However, generating stable 3D structures near the convex hull is challenging:

- Efficiency ⚡
- Symmetry ⚖️
- Stability 🏔️

Matra-Genoa adresses these challenges.

These compact stackings aren’t just for atoms—next time you see stacked oranges in a store, think crystallography! 🧠💡

Face-Centered Cubic (FCC, the golden structure in tic-tac-toe above):
 A cube with atoms on its faces. Look closely—it’s also alternating hexagon layers (A-B-C-A-B-C), filling ~74% space! ✨
Elements: Au, Cu, Al, etc.

Hexagonal Close-Packed (HCP): 
Layers of hexagons stacked A-B-A-B. Efficiently fills ~74% of space.
Elements: Mg, Ti, Zn, etc.

Body-Centered Cubic (BCC, the silver structure in tic-tac-toe above):
 A cube with an atom at its center. Simple but not most space-efficient (~52%). 🧊
Elements: Fe, Na, Cr, etc.

You’re absolutely right—we’ve been a bit too quick with the labeling. We’ll revise it for the next version. Thank you for pointing this out!

To the best of our knowledge, such systems are not included in the training set. Additionally, phonons or force constants are not used as training targets. Finally, results for most uMLIPs are not that good

Good point, it’s true that the training set for all the uMLIPs includes the primitive unit cell of these compounds. However, when calculating phonons, we use supercells containing approximately 200 atoms with slightly displaced positions.

5/5
🔑 Key takeaway: uMLIPs are ready for production use in phonon calculations! But choose wisely - not all models performing well on standard benchmarks will give you accurate phonon properties. Training data & architecture choices matter more than model complexity.

4/5
📊 The tested models fall into 3 clear tiers:
- Tier 1: MatterSim (excellent)
- Tier 2: SevenNet, MACE, CHGNet, M3GNet (good)
- Tier 3: ORB, OMat24 (needs work for phonons)

3/5
⚠️ Surprising finding: ORB & OMat24 excel at geometry optimization but struggle with phonons. Why? They predict forces directly instead of deriving them from energy gradients. This leads to issues with the small atomic displacements needed for phonon calculations.

2/5
🎯 Most accurate model? MatterSim steals the show for phonon predictions! It achieves errors even smaller than the difference between PBE & PBEsol functionals. What's interesting is it outperforms more complex equivariant networks while being based on simpler M3GNet architecture.

5/5
🔑 Key takeaway: uMLIPs are ready for production use in phonon calculations! But choose wisely - not all models performing well on standard benchmarks will give you accurate phonon properties. Training data & architecture choices matter more than model complexity.

3/5
⚠️ Surprising finding: ORB & OMat24 excel at geometry optimization but struggle with phonons. Why? They predict forces directly instead of deriving them from energy gradients. This leads to issues with the small atomic displacements needed for phonon calculations.

2/5
🎯 Most accurate model? MatterSim steals the show for phonon predictions! It achieves errors even smaller than the difference between PBE & PBEsol functionals. What's interesting is it outperforms more complex equivariant networks while being based on simpler M3GNet architecture.