Pourbaix Machine Learning Framework Identifies Acidic Water Oxidation Catalysts Exhibiting Suppressed Ruthenium Dissolution
Jehad Abed, Javier Heras-Domingo, Rohan Yuri Sanspeur, Mingchuan Luo, Wajdi Alnoush, Debora Motta Meira, Hsiaotsu Wang, Jian Wang, Jigang Zhou, Daojin Zhou, Khalid Fatih, John R. Kitchin, Drew Higgins, Zachary W. Ulissi and Edward H. Sargent, “Pourbaix Machine Learning Framework Identifies Acidic Water Oxidation Catalysts Exhibiting Suppressed Ruthenium Dissolution”, Journal of the American Chemical Society 146(23), 15740-15750 (2024). https://doi.org/10.1021/jacs.4c01353
This summary was written by a language model from the paper and has not been checked line by line. Rely on the published paper for details.
Question
Proton exchange membrane electrolyzers need an anode catalyst for the oxygen evolution reaction (OER) that survives strongly acidic conditions, and today that means scarce and costly IrO2. RuO2 is more active and Ru is more abundant and cheaper, but Ru over-oxidizes to soluble species during OER and the catalyst dissolves. Earlier doping and strain strategies improved activity but stability stayed at a few hours above 10 mA cm-2. The paper asks whether a machine learning screen of mixed rutile oxides for acid stability can point to RuO2-based compositions that dissolve more slowly while staying active (Abed et al. 2024).
Approach
A crystal graph convolutional neural network with hyperparameter tuning and dropout-based uncertainty (CGCNN-HD) was trained on 36,465 metal oxide bulk structures from the Materials Project to predict formation energy per atom, which the authors use as a proxy for Pourbaix decomposition energy (ΔGpbx). All structures were pre-scaled with a bond-statistics lattice scaling method instead of DFT relaxation, so the model works from unrelaxed hypothetical structures. To limit errors from non-ideal structures, predictions were shifted within a 95% confidence interval (μ - 2s) rather than using the dropout mean, accepting a few extra false positives for throughput. Substituting 46 elements into the rutile (space group 136) lattice gave 2070 hypothetical oxides; 195 survived an energy above hull cutoff of 0.5 eV. For Ru-containing candidates, metal-oxygen covalency from crystal orbital Hamilton population analysis served as a second stability descriptor.
Ru-Cr-Ti oxides were then made by sol-gel synthesis and characterized by XRD, HRTEM, EDS, XANES/EXAFS (including in situ XAS), STXM spectroptychography, ICP-AES dissolution measurements and 18O-labeled differential electrochemical mass spectrometry (DEMS). Surface Pourbaix diagrams and OER intermediates on rutile (110) and (101) slabs were computed with the automated WhereWulff workflow.
Key findings
The surrogate predicted ΔGpbx with a mean absolute error of 77 meV per atom from unrelaxed structures, checked against DFT on a subset of 676 oxides covering 26 elements (34% of the data set). Ti doping lowered ΔGpbx relative to RuO2 and gave the largest covalency gain (+10.7%), ahead of Sn (+9.05%) and Ge (8.6%); Cr also improved Pourbaix stability. Experimentally, 25 at.% Cr cut the overpotential at 10 mA cm-2 by about 70 mV relative to RuOx, while higher Cr contents formed less active Cr2O3. At most 20 at.% Ti could be incorporated, giving Ru0.6Cr0.2Ti0.2Ox, which kept the rutile structure with all cations near 4.
Adding Ti raised the overpotential by only about 12 mV, Tafel slopes were similar (58, 56 and 54 mV dec-1 for the Ti-doped, Cr-doped and RuOx samples), and Ru mass activities at 1.48 V vs RHE were 475, 433 and 79 A gRu-1, so Cr drove the activity gain (partly via about a 7-fold larger surface area). In a PEM electrolyzer at 100 mA cm-2, RuOx and Ru0.75Cr0.25Ox failed within roughly 40 h, whereas the Ti-containing catalyst ran 200 h with about a 5 mV rise in cell voltage; the abstract reports 267 mV overpotential and a degradation rate of 25 μV h-1. A three-electrode test ran 1100 h at 10 mA cm-2 with a 36 mV rise in the first 100 h and about 8 μV h-1 thereafter. Ru dissolved about 20 times more slowly than from RuOx, staying below 1 wt% loss. Post-reaction imaging showed a hydrous Cr-O overlayer, judged not to aid stability, while in situ XAS and STXM indicated a strong Ti-O network and higher metal-oxygen covalency (computed 93.62% versus 85.20% for RuO2). DEMS showed Ti suppressed lattice oxygen participation by about 66% relative to the Cr-only catalyst. In DFT, the (110) theoretical overpotentials were 0.52, 0.66, 0.58 and 0.39 V for RuO2, RuTiO4, RuCrO4 and RuCrTiOx, and the (101) RuCrO4 value was 0.24 V, attributed to a Ru-O-Cr dimer that weakens an intra OH*…O* hydrogen bond and destabilizes OH*.
The figure captions name a Ru0.6Cr0.4Ox sample where the text discusses Ru0.75Cr0.25Ox; this summary follows the text.
Significance
The work shows a full loop from a graph network stability screen through synthesis to long-duration electrolyzer testing, and it separates roles: Ti supplies stability through covalency and a rigid oxide framework that slows Ru dissolution and lattice oxygen evolution, while Cr supplies activity. It supports metal-oxygen covalency as a practical descriptor for acid-stable oxide design.