AI Capabilities

AI that generates, screens, and hunts for new materials

Three research papers and one autonomous search system — here's what each one does.

149,419

Candidates generated

8,800

Flagged likely metallic

3

Published research methods

Paper 1

Generating magnetic metal-organic frameworks

Metal-organic frameworks (MOFs) could replace rare-earth magnets, but their large unit cells make first-principles simulation too slow to search at scale. We fine-tuned a machine-learned interatomic potential (CHGNet) on 15,000+ first-principles computations from our Database, then used it to drive a site-substitution search across structural prototypes from the QMOF database — generating novel, highly magnetic MOF candidates far faster than DFT alone.

15,000+

DFT computations used to fine-tune CHGNet

149,419

Novel magnetic MOF candidates generated

12,500

DFT band gaps used to train the SchNet regressor

8,800

(5.9%)

Candidates flagged as likely metallic

Paper 2

Screening for metallic character

A magnetic MOF is only useful for spintronics if it also conducts. Generating candidates is one problem; knowing which ones are metallic is another. We screened all 149,419 candidates from the companion generation study with a SchNet regressor trained on 12,500 DFT band gaps, flagging 8,800 (5.9%) as likely metallic — a prioritization tool that narrows the field before anyone runs an expensive DFT calculation.

Agentic AI

An autonomous system that searches for materials on its own

We run agentic AI — autonomous systems that explore, act, and learn on their own — to search for new materials. Point it at a target property, and a swarm of AI agents proposes, tests, and refines candidates on its own. Today that loop is focused on finding a rare-earth-free permanent magnet, but the same system adapts to whatever material problem you bring us.

Every candidate still has to earn its place — validated against real physics before it's ever considered real.

Live evolution across generations

Paper 3

Characterizing qubit defects with machine learning

Superconducting qubits lose coherence to two-level system (TLS) defects buried in their building blocks, and finding them one by one with traditional fitting is slow — especially across a chip with hundreds of qubits. We trained a ResNet-based convolutional neural network on simulated two-tone spectroscopy images to read off all four TLS parameters — frequency, coupling strength, and both decay times — directly from the image, outperforming standard perturbation-theory fitting and accelerating the loop of designing better quantum chips.

4

TLS parameters predicted per scan — frequency, coupling, and both decay times

ResNet CNN

Trained on simulated two-tone spectroscopy images

Partner With Us

Get direct access to our data and models

Need access to our materials database, AI-generated candidates, or want to explore a joint discovery project? Reach out — we reply personally within one business day.