Insider Brief
- SandboxAQ released AQCat25-EV2, a quantitative AI model trained on the AQCat25 dataset to accelerate catalyst discovery across industries such as energy, chemicals, and agriculture.
- The model incorporates spin polarization effects to achieve full periodic-table coverage, providing accuracy close to quantum-mechanical methods at up to 20,000× faster speed.
- Developed on NVIDIA DGX Cloud using more than 500,000 GPU-hours on H100 GPUs, AQCat25-EV2 is available on Hugging Face, with SandboxAQ planning broader deployment through NVIDIA’s ALCHEMI platform.
PRESS RELEASE — SandboxAQ today announced the release of AQCat25-EV2, a powerful new quantitative AI model trained on the AQCat25 dataset, representing an advancement in computational catalyst discovery. AQCat25-EV2 enables researchers to accelerate the discovery of catalysts for applications in energy, chemical production, agriculture, automotive and consumer goods with unprecedented speed and accuracy.
Novel catalyst discovery offers the promise to revolutionize all sectors of the physical economy, since catalysts underpin more than 80% of all commercial goods produced globally. But designing better catalysts is presently limited by the low throughput of mostly-laboratory methods, which typically process fewer than 100 catalysts per week. At this throughput, only incremental changes upon existing catalysts are economically feasible.
Quantitative AI models can accelerate this throughput by several orders of magnitude. Until now, such models have been confined to accurately describing only a subset of the elements used in industrial catalyst discovery. By including the quantum effect of spin polarization, AQCat25-EV2 accurately expands that range to all industrially relevant elements for the first time, now enabling full coverage of the entire periodic table of elements. It is therefore the first heterogeneous catalyst model which can be consistently and robustly applied, thus enabling step change, rather than incremental advances, to industrial catalyst discovery broadly.
AQCAT25-EV2 predicts energetics with an accuracy approaching physics-based quantum-mechanical methods at speeds up to 20,000X faster. This combination of accuracy and speed makes large-scale, high-accuracy virtual screening across all industrially relevant elements finally practical, eliminating a key bottleneck that has constrained materials innovation for decades.
Trained on the AQCat25 dataset with 13.5 million high-fidelity quantum chemistry calculations across 47,000 intermediate-catalyst systems, AQCat25-EV2 is the only large-scale catalytic AI model to include the magnetic data of spin polarization, crucial to increasing the accuracy of model predictions since many abundant metals (such as cobalt, nickel, and iron) are spin polarized.
“AQCat25-EV2 is among the first models that will allow screening in silico on a wide set of chemistries with unprecedented precision and speed, opening the door to novel catalysts and applications,” said Dr. Bob Maughon, former CTO at Saudi Arabian chemical manufacturing company SABIC. “For critical, unsolved industry problems, from CO2 reduction to advanced battery materials, this technology will be an indispensable tool for accelerating discovery and securing better, more sustainable chemical solutions.”
AQCat25-EV2 was developed to provide researchers with confidence that they’ve identified the most promising target candidates, not just a narrow subset. Prior to AQCat25-EV2, the typical process involved looking for a limited number of interesting target candidates, since running additional simulations was costly and time consuming. Once a candidate was discovered, and screening stopped, other, better candidates remained undiscovered.
“A wide range of industries face critical unsolved problems in catalysis today, all of which are expected to directly benefit from AQCat25-EV2,” said Aayush Singh, who leads Catalytic Sciences at SandboxAQ. “These include plastic recycling (depolymerization), CO2 reduction to fuels & chemicals, hydrogen for fuel cells, methane (flare gas) to methanol, syngas to ethanol (higher alcohols), and more. For these industries, we’re fundamentally de-risking the R&D process across the entire spectrum of materials science.”
The AQCat25 dataset and AQCat25-EV2 were developed on NVIDIA DGX™ Cloud, leveraging more than 500,000 GPU-hours on NVIDIA H100 Tensor Core GPUs. SandboxAQ will be using the NVIDIA ALCHEMI platform following this release, broadening accessibility and offering bleeding-edge performance to all researchers worldwide. SandboxAQ is also a member of the NVIDIA Inception program.
AQCat25-EV2 is launching today and is available on Hugging Face alongside a new manuscript describing its technical foundation and validation studies. To learn more visit www.sandboxaq.com/aqcat25 or contact aqcat@sandboxquantum.com.
