Machine learning and quantum chemistry unite to simulate catalytic dynamics

Catalysts play an integral role in modern manufacturing. More than 80% of all manufactured products, from pharmaceuticals to plastics, rely on catalytic processes at some stage of production. In particular, transition metals stand out as highly effective catalysts. This is because its partially filled D orbitals allow for easy exchange of electrons with other molecules. However, this property requires an accurate description of the electronic structure, which challenges them to model accurately.

Designing efficient transition metal catalysts that can be implemented under realistic conditions requires more than a static snapshot of the reaction. Instead, you need to capture dynamic images. How molecules move, interact at various temperatures and pressures, and atomic motion fundamentally shapes catalytic performance.

To address this challenge, the lab of Professor Laura Gagrialdi, Faculty of Molecular Engineering (Uchicago PME) and Chemistry Department, has developed a powerful new tool that utilizes electronic structure theory and machine learning to simulate transition metal catalytic dynamics with both accuracy and speed.

“For the past decade, machine learning possibilities have been a major advance in ways that simulate molecular dynamics and provide speed and scalability. However, accurately capturing the electronic structure of transition metal catalysts has been an unsolved challenge.

The results are published in Proceedings of the National Academy of Sciences.

Machine learning enables faster simulations

Over the past decade, the Gagliardi group has developed Multi-configuration Pair Density Functional Theory (MC-PDFT), a quantum chemical method that can describe the complex electronic structures of transition metal reactions. MC-PDFT offers high accuracy, but is extremely slow to simulate the dynamics of catalytic systems. This is an important step in predicting how a catalyst will behave under realistic conditions.

To address this challenge, teams can turn to machine-learned interatomic potentials (ML potentials) and capture molecular dynamics with incredible efficiency. ML potentials have been widely applied to materials science, but have not been successfully combined with multiplexing methods such as MC-PDFT.

The reason lies in the long-standing obstacle: consistency of labeling. Machine learning models require unique and reliable property labels, such as energy and forces derived from wave functions, for any molecular geometry along the reaction path. For multiplexed quantum chemistry methods, assigning such labels uniquely remained an open question.

To overcome this challenge, student Aniruddha Seal, co-advising PhD Gagliardi and Professor Andrew Ferguson, has developed a new algorithm to generate a consistent wavefunction of new geometry as a weighted combination of wavefunctions from previously sampled molecular structures. The closer the new geometry is to a known geometry, the more its wavefunction will resemble that of a known structure. This approach allows unique, consistent wave functions to be assigned to all points along the reaction path, allowing accurate training of ML-Potentials with multi-islancer data.

“Think of mixing paint into the palette,” explained Seal. “If you want to create a green shade that is closer to blue, use more blue paint and a little yellow. If you want a shade that is tilted to yellow, it's a balance flip. The closer the target color is to one of the base paints, the more impact it will have on the mix. WASP works the same way.

This innovation forms the basis for Weighted Active Space Protocol (WASP), a framework that combines the accuracy of MC-PDFT with the efficiency of machine learning, developed through close collaboration with the Italian Institute of Technology and the Parinero Group in Genoa, combining the expertise in electronic structure theory with the possibilities of machine learning. WASP offers dramatic speedup: multiplexing accuracy simulations that can be completed in minutes that once took months.

Impact: Bridging accuracy and efficiency in catalyst design

By integrating accuracy and speed, WASP opens the door to designing catalysts that can withstand realistic conditions such as high temperatures and high pressures. Transition metals are the heart of countless large-scale processes, but their complexity has made the catalyst rational design difficult.

The main example is the Harbor Bosch process, in which iron functions as a catalyst for converting iron and hydrogen to ammonia. Despite being developed over a century ago, this iron catalyst still controls ammonia production around the world. With WASP, researchers now have tools to explore alternatives that can increase efficiency, reduce by-products, and reduce environmental costs.

So far, WASP has been successfully demonstrated for thermally activated catalysts, that is, heat driven reactions. The next frontier adapts this method to photoactivation reactions, which are essential for the design of new photocatalysts. Photocatalysts have a great promise in technology, from water treatment to energy production.

The new tool is published at https://github.com/gagliardigroup/wasp.