Accelerated optimization of pure metal and ligand compositions for light-driven hydrogen production
Maya Bhat, Eric M. Lopato, Zoe C. Simon, Jill E. Millstone, Stefan Bernhard and John R. Kitchin, “Accelerated optimization of pure metal and ligand compositions for light-driven hydrogen production”, Reaction Chemistry & Engineering 7(3), 599-608 (2022). https://doi.org/10.1039/d1re00441g
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
Photocatalytic hydrogen evolution with nanoparticle (NP) cocatalysts formed in situ from metal salts involves many interacting components: a molecular photosensitizer (PS), a sacrificial reductant, a stabilizing ligand, the metal salt, water and organic solvent. Varying one component at a time is slow and misses interactions. The authors asked whether a design of experiments (DOE) framework coupled to a 108-well parallel photoreactor could quickly find the optimal PS, metal and ligand concentrations for four single metals (Au, Cu, Ni and Fe), and whether trends could be generalized across them (Bhat et al. 2022).
Approach
Wells (440 µL) contained the iridium photosensitizer [Ir(Fmppy)2dtbbpy]PF6 and metal salts in DMSO, with thiolated poly(ethylene glycol) methyl ether (PEGSH, 1 kDa) as the stabilizing ligand and triethanolamine (TEOA) in water. A liquid handling robot filled the plates, which were illuminated with blue LEDs (440 to 450 nm) for about 15 hours per round; H2 was quantified from the darkening of a chemosensitive detection tape. Six wells per plate held a molecular internal standard (Eosin Y with a cobaloxime catalyst), used to rescale results between plates.
The activity was modeled with an ordinary least squares fit of a quadratic response surface in metal, ligand and PS concentration (linear, cross and squared terms). The first designs were Box-Behnken (15 wells per metal, three center points). The model’s predicted optimum became the center of the next design with tighter bounds, and iteration stopped once the predicted optimum changed by less than 1 µmol H2. After PS was fixed, the two-dimensional metal and ligand space was sampled with central composite inscribed designs. The maximum H2 produced in a well was used as the target, since it correlated well with the maximum production rate but needed no smoothing of noisy time series. Data and analysis lived in a shared Google Drive workspace with Colab notebooks and two Python packages written for the project (doeanalysis and gespyranto). Best wells were examined by TEM and XPS.
Key findings
Earlier work used 0.25 mM PS, but once ligand and metal were tuned the H2 curves for different metals became indistinguishable, a sign of a PS-limited regime. Two DOE iterations pushed the predicted PS optimum to the upper bound (1 mM), so PS was scanned separately: activity rose steeply up to about 0.25 to 0.4 mM and gained little beyond roughly 0.4 to 0.5 mM for all four metals. PS was then fixed at 0.5 mM, which conserves the expensive iridium complex and reduces the problem to two dimensions.
Convergence in the metal and ligand space took up to six iterations per metal. Predicted optima were 1.41 mM Au with 0.81 mM PEGSH (20.17 µmol H2), 1.55 mM Cu with 0.85 mM PEGSH (15.55 µmol), 1.27 mM Ni with 0.97 mM PEGSH (21.40 µmol) and 1.22 mM Fe with 1.08 mM PEGSH (20.54 µmol). The best measured wells gave 22.77, 17.92, 24.30 and 22.71 µmol H2 respectively. The model underpredicted the best observations, but within the roughly 3 µmol H2 uncertainty estimated from the internal standard wells. Au and Cu had similar response surfaces (more sensitive to metal than to ligand concentration, with broad peaks), while Ni and Fe were more sensitive to ligand concentration. The standard error of the predicted optimum fell to about 0.2 µmol H2 by the last iteration for every metal.
TEM after 45 hours of illumination showed discrete, unaggregated particles for all four metals, consistent with PEGSH stabilization. XPS showed metallic Au and Cu, which the authors attribute to their higher reduction potentials, while the Ni sample contained metallic Ni and NiO and the Fe sample contained iron sulfide. The iron sulfide particles were about as active as Au and stayed stable for more than 45 hours. In total the study used fewer than 1,000 wells, compared with an estimated 16,000 for a replicated grid at 0.1 mM resolution with five PS levels, a reduction of more than a factor of ten.
Significance
The paper shows that a simple quadratic response surface model, iterated with classical DOE designs on a parallel photoreactor, can locate composition optima for multicomponent photocatalytic systems with an order of magnitude fewer experiments than grid screening. It gives practical operating guidance for this reactor (0.5 mM PS is sufficient for metal loadings up to about 1.2 mM, and about 1 mM metal with 1 mM PEGSH is a good region for high activity) and supports the use of thiolated ligands instead of supports to keep in situ NPs active. The authors also present the shared, digitized data workflow as a model for multi-institution collaboration, and suggest extending the framework to bimetallic colloids and other stabilizing agents.