SandboxAQ has announced the general availability of AQCat, an AI model designed to predict the physics of materials and screen catalysts at a scale that has traditionally required significant computing resources and specialized expertise. 

 

Available through Claude Science via the Model Context Protocol (MCP), AQCat lets researchers evaluate catalyst candidates using plain-English prompts, with no coding required.

 

The company says AQCat can identify promising catalysts with accuracy approaching established high-accuracy methods while running up to 20,000 times faster.

 

The stakes are enormous. Catalysts accelerate chemical reactions and make them viable at industrial scale. They are used in more than 80% of manufactured goods and more than 90% of commercially produced chemicals, underpinning products ranging from fertilizer and fuel to plastics, glass and textiles.

 

Yet catalyst discovery remains a slow, expensive process. Conventional laboratory approaches can test fewer than 100 candidate materials a week, limiting researchers' ability to search the vast number of possible materials.

 

AQCat is designed to change that equation by tackling one of the earliest and most important calculations in catalyst discovery: adsorption energy. The measurement indicates how strongly molecules bind to a catalyst's surface and provides an early signal of whether a material is worth pursuing.

 

Instead of committing expensive modeling or laboratory resources to candidates one at a time, researchers can use AQCat to screen thousands of possibilities and concentrate their efforts on the most promising materials.

 

AQCat's key distinction is its treatment of magnetism.

 

Density Functional Theory (DFT), widely regarded as the benchmark for these calculations, can deliver highly accurate results but requires substantial computational time and resources. Faster machine-learning approaches can reduce that burden, but may fail to capture magnetic effects that are important for industrially relevant materials.

 

AQCat is designed to be "spin-aware," accounting for magnetic behavior in materials including iron, cobalt and nickel. That could make abundant and relatively inexpensive metals with complex magnetic properties more practical targets for high-throughput catalyst discovery.

 

SandboxAQ trained AQCat on AQCat25, a public dataset containing 13.5 million high-fidelity DFT calculations spanning 47,000 intermediate-catalyst systems and every industrially relevant element.

 

"In catalysis research, a key bottleneck has been the number of surfaces you can realistically evaluate within compute and scientist resource constraints. Recovering DFT-quality adsorption energies at a faster speed changes the scale of questions my group can ask, and it opens up design spaces that were previously out of reach. 

 

"Putting models behind a natural-language interface lowers the barrier for students and collaborators to participate in computational work," said Joe Gauthier, Assistant Professor of Chemical Engineering, Texas Tech University.

 

The natural-language interface is central to the launch. Rather than requiring researchers to write code or operate specialized computational workflows, AQCat allows them to interact with the model conversationally through Claude.

 

"Adsorption energy is the critical first calculation in most catalyst screening workflows, because it tells you how strongly a molecule binds to a surface and helps you prioritize what to study next. Many AI models overlook the magnetic physics that matter in earth-abundant industrial catalysts. 

 

"We built AQCat to be spin-aware from the ground up, so it can recover near-DFT accuracy for magnetically complex systems, including iron, cobalt, and nickel, while running thousands of times faster than traditional approaches. Making it generally available on Claude means any researcher can run that calculation at scale, in plain language, without touching a line of code," said Aayush Singh, Head of Science, Catalysis, SandboxAQ.

 

The technology could be applied to research ranging from green hydrogen and sustainable aviation fuel to fertilizer production and plastics recycling. The common strategy is the same: use AI to narrow a massive design space before researchers spend significant resources on DFT calculations or physical experiments.

 

That could shift catalyst discovery from a process constrained by how many candidates researchers can afford to test into one increasingly constrained by how many ideas they can generate.

 

Geoff Ling, Founding Director of the Biotech Office at DARPA, said, "In science, the biggest risk is often the unknown. Anything that helps researchers reduce uncertainty is incredibly valuable, and SandboxAQ's MCP tools in Claude do that for both materials science and drug discovery."