⏱️ Lectura: 11 min
This week CuspAI launched its AI Materials Foundry, a network that promises to connect generative models, simulation computing, and manufacturing so that a material designed by artificial intelligence stops being trapped in a database and reaches a real laboratory.
📑 En este artículo
The news, reported by Las Vegas Sun, comes at a time when labs like DeepMind, Microsoft Research, and Meta have already published massive models and datasets to accelerate materials design. The question this article answers is what really changes when, instead of an isolated model, a full network of partners is launched.
TL;DR
- CuspAI launched AI Materials Foundry, a global network to accelerate AI-driven materials design.
- The network connects research, computing, and manufacturing to validate AI-generated candidate materials.
- In 2023, DeepMind’s GNoME predicted 2.2 million new crystal structures; 380,000 turned out to be stable.
- Microsoft introduced MatterGen, a generative model that designs materials from scratch instead of just screening them.
- Meta released the OMat24 dataset with more than 100 million DFT calculations to train materials models.
- The bottleneck is still physical synthesis and lab validation, not candidate generation.
- With Materials Project, ASE, and MACE-MP, any developer can try candidate screening today.
What happened: CuspAI launches AI Materials Foundry
CuspAI, a startup focused on generative artificial intelligence applied to materials discovery, announced the launch of AI Materials Foundry: according to the Las Vegas Sun report, it’s a global network designed to speed up the transition of an AI-generated candidate material into a validated, manufacturable material.
The core proposal isn’t a new AI model but a collaboration infrastructure: connecting those who generate candidates with those who have the computing power to screen them, and with those who can synthesize and measure them in a physical laboratory. That chain (generate, screen, validate, manufacture) has historically been the bottleneck of materials science.
CuspAI competes in a space where corporate labs with far more resources also operate, so the partner network is, in practice, its way of matching computing and lab capacity without building every piece in-house.
Context and history
Discovering new materials (a lighter alloy, a solid electrolyte for batteries, a catalyst for carbon capture) traditionally combined chemical intuition with years of lab trial and error. Quantum simulation, particularly density functional theory (DFT), made it possible to calculate a structure’s stability without synthesizing it, but at a high computational cost per candidate.
The recent leap came from deep learning. In 2023, DeepMind published its GNoME (Graph Networks for Materials Exploration) system in Nature, which predicted 2.2 million new crystal structures, of which about 380,000 were considered stable enough to be real synthesis candidates. It was a leap in scale: before GNoME, the materials community knew about 48,000 stable structures after decades of accumulated experimental and computational work.
Microsoft Research went a step further with MatterGen, a diffusion model that isn’t limited to evaluating existing structures: it directly generates new compositions and crystal structures, conditioned on desired properties like low toxicity or a specific band gap. The Microsoft Research blog describes it as a reversal of the traditional process: instead of searching a database, the model designs from scratch.
Meta, for its part, released the OMat24 (Open Materials 2024) dataset in 2024, with more than 100 million open DFT calculations to train materials potential models, available through Meta AI. That volume of open data is largely what made it possible for smaller companies like CuspAI to train and fine-tune their own models without repeating the most expensive computational work from scratch.
This kind of effort didn’t come out of nowhere: in 2011, the United States launched the Materials Genome Initiative, a government program that pushed for combining open data, computational simulation, and experimentation to shorten the time it takes to bring a new material from the lab to the market. The generative models of the 2020s are, in large part, the maturation of that same idea using deep learning tools that didn’t yet exist at that scale in 2011.
Technical details and performance
A typical AI-assisted materials discovery pipeline has four stages, and CuspAI’s proposal is explicitly to organize them as a network instead of leaving them scattered among teams that don’t communicate with each other.
First, a generative model proposes candidate structures: chemical composition and crystal arrangement. Second, a machine learning potential (models like MACE-MP or M3GNet) approximates that structure’s energy in milliseconds, without solving each atom’s Schrödinger equation. Third, the candidates that survive that first filter move on to DFT validation, much more precise but also much slower. Fourth, only the candidates that pass both filters reach lab synthesis, where it’s finally measured whether the real material behaves as the model predicted.
flowchart TD
A["Data and scientific literature"] --> B["Generative AI model"]
B --> C["Screening with ML potentials"]
C --> D["DFT validation"]
D --> E["Lab synthesis"]
E --> F["Validated material"]
subgraph AIMF["AI Materials Foundry"]
B
C
D
endThe way to verify whether a candidate is worth synthesizing is a concrete metric from solid-state physics: the distance to the formation energy convex hull, expressed in electronvolts per atom (eV/atom). The closer to zero, the more thermodynamically stable the structure is relative to other possible combinations of the same elements. Databases like Materials Project expose that value directly for every calculated structure, and it’s the filter that any team, inside or outside a network like CuspAI’s, uses before spending lab time on a candidate.
| Approach | How it works | Advantage | Limitation |
|---|---|---|---|
| Traditional experimental synthesis | Lab trial and error, direct physical characterization | Physically verified result from the start | Slow and costly per candidate |
| DFT simulation | Quantum calculation of a proposed structure’s energy and stability | Physical precision without synthesizing anything | Computationally expensive for exploring thousands of candidates |
| ML potentials (MACE, M3GNet) | Approximate DFT energy by training on prior data | Orders of magnitude faster than DFT | Can be wrong on structures very different from their training set |
| Generative models (GNoME, MatterGen, CuspAI’s network) | Propose new structures instead of just evaluating existing ones | Expand the search space beyond known databases | Require DFT and lab validation before the result can be trusted |
How to start testing it
You don’t need to be part of CuspAI’s network to experiment with this same chain of tools: much of the AI-assisted materials discovery stack is open.
To calculate a structure’s energy with a pretrained ML potential, without installing a full DFT code, this is enough:
pip install ase mace-torchfrom ase.build import bulk
from mace.calculators import mace_mp
celda = bulk("Si", "diamond", a=5.43)
celda.calc = mace_mp(model="medium", device="cpu")
print(celda.get_potential_energy())This script builds a silicon crystal cell in its diamond phase, assigns it the pretrained MACE-MP potential, and calculates its potential energy in seconds, without running a single DFT iteration. It’s the same kind of fast screening used by large-scale discovery pipelines, just running on a laptop.
To search for candidates already calculated and filtered by stability, the Materials Project API allows direct queries:
from mp_api.client import MPRester
with MPRester("YOUR_API_KEY") as mpr:
candidatos = mpr.materials.summary.search(
elements=["Li", "Fe", "O"],
energy_above_hull=(0, 0.03),
fields=["material_id", "formula_pretty", "energy_above_hull"]
)
for material in candidatos[:5]:
print(material.material_id, material.formula_pretty, material.energy_above_hull)The query searches for lithium, iron, and oxygen compounds, a typical family of battery cathodes, with a distance to the hull under 0.03 eV per atom, a common threshold for considering a material thermodynamically close to stability. The result is a short list of real candidates, with their database identifier, ready to review before thinking about synthesis.
💡 Tip: the Materials Project API requires a free key, obtained by registering at materialsproject.org; without a key, queries fail with an authentication error.
Impact and analysis
The underlying argument behind AI Materials Foundry is that generating candidates has stopped being the main problem. Models like GNoME and MatterGen showed that proposing millions of structures is feasible; what remains scarce is lab capacity to synthesize and characterize those candidates at the same speed. Connecting multiple partners in a network aims to spread that bottleneck across more facilities in parallel, instead of each company competing to build its own lab from scratch.
The fields most often cited for this kind of acceleration are next-generation batteries, carbon capture materials, and semiconductors, three areas where a shorter trial-and-error cycle has a direct impact on research and development costs.
Betting on a network instead of a closed model is also a business bet: if CuspAI gets multiple manufacturers and labs to use the same AI layer to evaluate candidates, every successful synthesis that happens within that network feeds real experimental data back into the model, something no public dataset like OMat24 can offer on its own, because it doesn’t capture the failures and adjustments that only show up in an active lab.
⚠️ Heads up: a generative model predicting a stable structure doesn’t guarantee it’s synthesizable with current lab methods; the real survival rate of candidates from prediction to successful synthesis is still the number no demo shows.
That’s the honest limit of this whole approach: a generative model can hallucinate structures that are stable on paper but unreproducible in practice, either because they require extreme synthesis conditions or because the ML potential that evaluated them was wrong outside its training range. No partner network eliminates that risk; at best, it catches it faster, because more hands are validating in parallel.
What’s next
What remains to be seen is whether CuspAI publishes verifiable results (materials synthesized and characterized, not just predicted) that specifically came out of this network, and whether independent labs confirm those findings, the way external teams replicated and questioned a fraction of GNoME’s structures. Third-party validation, more than the launch announcement, is what will determine whether AI Materials Foundry accelerates materials science or just accelerates the marketing around it.
📖 Summary on Telegram: View summary
Try it yourself: install ase and mace-torch with pip and run the script above to calculate your first crystal structure energy today, without waiting for CuspAI’s Foundry to open public access.
Frequently Asked Questions
What is CuspAI?
It’s a startup focused on generative artificial intelligence applied to materials discovery, competing in a space where large corporate labs like DeepMind and Microsoft Research also operate.
What exactly is AI Materials Foundry?
According to the announcement covered by Las Vegas Sun, it’s a global network that connects research, computing, and manufacturing to take AI-generated candidate materials to real-world validation, not just an isolated model or dataset.
How is it different from GNoME or MatterGen?
GNoME and MatterGen are a specific system and model for predicting or generating structures. The Foundry, as described, is a collaboration layer among different partners that uses that type of model as just one more piece of the chain, not as the final product.
Can I experiment with these tools without being a materials researcher?
Yes. With a free Materials Project key and the open ASE and MACE-MP libraries, you can calculate structure energies and query candidates already filtered by stability, as shown in this article’s examples.
Does AI replace lab synthesis?
No. It reduces how many candidates need to be synthesized and in what order to test them, but the material’s physical manufacturing and characterization still happen in a real lab, and it remains the slowest step in the whole chain.
Where is the first practical impact expected?
The fields most frequently cited in the industry are next-generation batteries, carbon capture materials, and semiconductors, all areas where shortening the trial-and-error cycle directly reduces research and development costs.
References
- Las Vegas Sun: original announcement of CuspAI’s AI Materials Foundry launch.
- DeepMind: official blog on GNoME and the discovery of 2.2 million crystal structures.
- Microsoft Research: announcement of MatterGen, the generative materials design model.
- Meta AI: official page of the lab that released the open OMat24 dataset.
- Materials Project: open database and API of material structures and properties.
- ASE (Atomic Simulation Environment): official documentation of the library used in the code examples.
📱 Enjoying this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.
Imagen destacada: Foto de Vitaly Gariev en Unsplash
0 Comments