A breakthrough collaboration has shown that graphics processing unit (GPU) supercomputers can now solve the most demanding quantum chemistry calculations with both speed and accuracy. An international team led by computational chemists from NVIDIA, Sandbox AQ, and the Department of Energy's Pacific Northwest National Laboratory demonstrated that AI-oriented hardware can tackle complex chemical structures that were previously out of reach, potentially accelerating the discovery of new catalysts and materials.

What Makes These Quantum Chemistry Problems So Hard?

Quantum chemistry involves understanding how electrons interact in complex molecular systems. The two chemical structures the research team solved, FeMoco and cytochrome P450, are considered benchmarks for computational methods because they involve large numbers of interacting electrons and have real-world importance. FeMoco helps catalyze the conversion of atmospheric nitrogen to ammonia, a key component of fertilizer, while cytochrome P450 is an important liver enzyme. Until now, solving these structures has proven too complex and time-consuming for even the most advanced computing platforms.

The challenge lies in the sheer computational complexity. These systems require calculations with extremely high precision, which traditionally demanded expensive supercomputing resources. Researchers needed a way to harness the massive computational power of modern GPUs without sacrificing the mathematical accuracy that chemistry demands.

How Did Researchers Overcome the Precision Problem?

The key innovation was using a technique called mixed-precision computing on NVIDIA's Blackwell architecture GPU hardware. Rather than maintaining the highest level of mathematical precision throughout every calculation, the team strategically used lower precision where it wasn't necessary and reserved high precision for calculations that demanded it. This approach combined the Density Matrix Renormalization Group (DMRG) method, a numerical technique for solving difficult quantum chemistry problems, with techniques that emulate high-precision arithmetic through reduced-precision compute resources.

"Our study shows that AI-oriented hardware can do more than provide speed, it can also power chemically accurate, strongly correlated quantum chemistry at the frontier of what is computationally feasible," said Sotiris Xantheas, a computational chemist at PNNL.

Sotiris Xantheas, Computational Chemist at Pacific Northwest National Laboratory

The result was remarkable: the first quantum chemistry calculation using high-precision emulation that achieved chemical accuracy, meaning the results were precise enough to be useful for real chemistry applications. The team's approach successfully described complex, asymmetrical forces between particles with over 99% accuracy.

Steps to Apply This Method to Future Chemistry Problems

• Identify Target Systems: Select complex chemical structures that involve large numbers of interacting electrons and have practical applications in catalysis, bioinorganic chemistry, or materials science.
• Develop Custom Neural Networks: Design AI models that can learn from limited experimental data while respecting the physical laws governing the system, rather than treating the model as a "black box."
• Implement Mixed-Precision Computing: Leverage GPU hardware to strategically apply high precision only where chemical accuracy demands it, reducing computational cost while maintaining reliability.
• Validate Against Benchmarks: Test the AI model's predictions against known chemical properties and experimental results to ensure the method produces chemically meaningful answers.

The research team included Cole Brower, Samuel Rodriguez Bernabeu, Jeff Hammond, and John Gunnels from NVIDIA, Martin Ganahl from Sandbox AQ, and Andor Menczer from the Wigner Research Centre for Physics, along with Örs Legeza of Wigner and the Technical University of Munich, who specializes in developing DMRG methods.

What Does This Mean for Materials Science and Catalysis?

The implications are significant. Calculations once considered prohibitively difficult could become routine on next-generation GPU accelerator platforms. This opens practical pathways to solving problems in catalysis, bioinorganic chemistry, and materials science that were previously far harder to access.

"By demonstrating that mixed-precision DMRG with emulated high-precision arithmetic can reach chemical accuracy for challenging active spaces, we've opened a practical path to using next-generation Blackwell systems for problems in catalysis, bioinorganic chemistry, and materials science that were previously far harder to access," explained Örs Legeza.

Örs Legeza, Senior Author, Wigner Research Centre and Technical University of Munich

The work was published in the Journal of Chemical Theory and Computation and was supported by the Hungarian National Research, Development, and Innovation Office, the Hans Fischer Senior Fellowship program at the Technical University of Munich, and the Scalable Predictive methods for Excitations and Correlated phenomena (SPEC) initiative, a Department of Energy program.

This breakthrough demonstrates that the convergence of AI hardware and advanced scientific computing methods can unlock solutions to some of science's thorniest problems. As GPU technology continues to advance, researchers expect that even more complex chemical systems will become accessible, potentially accelerating the discovery of new materials and catalysts that could address energy challenges and other pressing scientific questions.