Scientists at EPFL have unraveled the details of the first crucial step in the oxygen evolution reaction, a bottleneck for clean hydrogen production, using advanced simulations and machine learning techniques.
In our search for cleaner energy sources, hydrogen stands out, as it can store and deliver energy without producing carbon emissions because burning hydrogen only creates water. This makes it one of the cleanest fuels available.
One of the ways hydrogen is produced is through a process called "water splitting", which breaks water into hydrogen and oxygen using sunlight. Water splitting can make hydrogen widely available, but despite decades of research, it is still too inefficient.
The bottleneck in hydrogen production
The main culprit is the oxygen evolution reaction (OER), the stage in water splitting where oxygen gas is produced by removing electrons and protons from water molecules. OER limits the speed of water splitting, and it is the focus of scientists trying to improve the efficiency of the process.
OER takes place at the interface between a light-absorbing material like bismuth vanadate (BiVO₄) and water, and it is notoriously slow. During the reaction, BiVO₄ helps to extract electrons and protons from water molecules, which produces oxygen gas. In other words, BiVO₄ is the "workhorse" surface where this crucial, slow step happens.
The key step: proton-coupled electron transfer
The bottleneck has always been the very first step: moving a proton and an electron together-a process called proton-coupled electron transfer (PCET). In this process, a proton and an electron move together to help break apart water molecules, making it a key step for producing hydrogen and oxygen.
Despite progress in understanding the thermodynamics of PCET, pinning down its exact mechanism has been difficult. Previous studies have often overlooked the chaotic movement of water molecules at the surface of BiVO₄ or relied on methods that couldn't reach the time or accuracy needed. This has left a gap in our understanding of why BiVO₄ works as it does, and how to improve it.
Advanced simulations reveal the dance
Now, two researchers at EPFL, Yong-Bin Zhuang and Alfredo Pasquarello have found an answer to the problem. By combining long-timescale molecular dynamics simulations with machine learning potentials-algorithms trained to mimic high-level quantum calculations-the scientists captured the full dance of atoms and electrons at the BiVO₄-water interface. Zeroing in on the very first step of the OER, the initial PCET event, they found that the proton moves first, followed by the electron, and this sequence sets the reaction's pace.
The scientists built a detailed atomic model of the BiVO₄-water interface and used machine learning to predict the forces between atoms with high accuracy. This allowed them to run simulations much longer than would be possible with standard quantum calculations-long enough (up to 30 nanoseconds) to reach stable, converged results and sample hundreds of thousands of atomic configurations.
By carefully tracking key variables like the position of the proton and the shifting location of the "hole" (the absence of an electron), they could watch the entire PCET process unfold. They also used multiple independently trained machine learning models to ensure their results were robust.
What makes hydrogen production slow?
The simulations revealed something important: the slowest, rate-determining step is the direct transfer of a proton from a water molecule adsorbed on the BiVO₄ surface to a neighboring oxygen atom on the surface. The electron (or "hole") only hops to its new site after the proton has moved. The study also showed that this direct proton transfer, rather than an indirect route involving extra water molecules, dominates the process. This insight fits well with recent experiments that highlighted proton transfer as the bottleneck for the reaction.
Knowing that proton transfer limits the reaction, scientists can now focus on finding ways to speed up this specific step-perhaps by modifying the surface of BiVO4 or using additives to stabilize key structures. The study also sets a new standard for simulating complex reactions at interfaces, showing that machine learning can bridge the gap between accuracy and computational cost.
