(Invited) Machine Learning-Based Models and Simulations for Accelerated Discovery of Thermochemical Water-Splitting Media

Machine Learning-Assisted Development of High-Performance Ethanol Synthesis Catalysts via CO2 Hydrogenation.

The discovery and development of high-performance catalysts, which is crucial across all catalysis areas, requires advanced technologies and innovative approaches. Recently, machine learning (ML) has shown promise in accelerating this process, but its capability and examples of discovery of truly novel catalysts have remained limited. In this study, we describe an ML approach that goes beyond the traditional element pool, incorporating elements that have not been previously studied, to develop highly efficient catalysts for ethanol synthesis via CO2 hydrogenation. Starting with an initial data set of 58 catalysts (274 data points obtained at reaction temperatures ranging from 240-400 °C), we conducted 24 iterations of a closed-loop discovery system (ML predictions + experimental validation), testing a total of 555 catalysts (2477 data points), and building a large experimental data set. More than 50 catalysts with superior activity were discovered through this data-driven approach. The multielemental Pd(0.8)-Au(0.3)/K(2.5)-Sr(1)-Fe(20)-Zn(4)-Cd(2)-Yb(1)-Re(1)/CeO2(25%)-ZrO2 catalyst, where the numbers in parentheses represent weight percent (wt %), was identified as the most effective catalyst for ethanol synthesis (ethanol space-time yield: 8.2 mmol gcat-1 h-1 with a CO2 conversion of 57.6% and an ethanol selectivity of 23.2% under reaction conditions of 360 °C, 4 MPa, 12 L gcat-1 h-1, H2/CO2 = 3/1). Comprehensive characterizations, including in situ/operando techniques such as X-ray absorption spectroscopy (XAS), ambient-pressure X-ray photoelectron spectroscopy (AP-XPS), and diffuse reflectance infrared Fourier transform spectroscopy (DRIFTS), enable us to highlight the critical roles of each constituting element in improving ethanol synthesis efficiency.

Machine Learning Meets Photocatalytic Water Splitting for Sustainable Hydrogen Production

Photocatalytic water splitting has garnered immense interest as a sustainable pathway for clean hydrogen production by directly converting solar energy into chemical fuel. However, challenges related to intricate charge carrier dynamics, surface redox kinetics, and the vast search space for multicomponent catalysts continue to constrain the systematic development of efficient systems. Machine learning (ML) has emerged as a transformative tool to address these bottlenecks by enabling the accurate prediction of electronic properties, the identification of promising heterostructures, and the optimization of reaction conditions while reducing reliance on traditional trial‐and‐error methods. By capturing complex nonlinear correlations among structural descriptors and catalytic performance, ML facilitates the exploration of high‐dimensional design spaces that are essential for advancing solar‐to‐fuel conversion research. This review provides a comprehensive overview of how ML supports systematic materials innovation to realize stable and high‐efficiency systems for sustainable hydrogen evolution. As such, the integration of ML with experimental and theoretical methodologies is expected to establish a predictive and systematic framework for photocatalyst development, thereby accelerating progress toward scalable solar‐to‐hydrogen energy conversion.

Machine learning approaches for electrocatalyst design in water splitting: a review for green hydrogen production

The production of green hydrogen through water splitting requires highly efficient electrocatalysts, but the current trial-and-error-based synthesis or discovery is time-consuming, costly and resource-intensive. Machine learning (ML) provides a powerful, data-driven alternative that can model complex structure-activity relationships across large chemical spaces at orders-of-magnitude speed. This review systematically overviews the life cycle of the electrocatalyst research and development application of ML. First, the thermodynamic and kinetic principles of the hydrogen and oxygen evolution reactions are summarised, along with some well-adopted and accepted activity descriptors. Then we explore data sources, featurization approaches, and algorithms, and discuss the model space, from a simple interpretable model to a graph neural network to a generative model, in the context of the ML toolkit. Strategic applications are discussed for high-throughput virtual screening of alloys and single-atom catalysts, as well as multifunctional activity prediction for overall water splitting, and stability optimisation under operating conditions. The topic of emerging frontiers is highlighted, including high-entropy alloys, amorphous materials, and linking atomic-scale understanding to device-level performance through integration with density functional theory. Finally, the problems of data scarcity, model interpretability and the discrepancy between computational predictions and industrial implementation are discussed, along with future directions for closed-loop discovery and self-driving laboratories. Incorporating ML into electrocatalyst design and combining it with autonomous experimentation will revolutionise this process, from simulation to energy solution, dramatically speeding it up.

( Invited ) Machine Learned Force Field Accelerated Modeling of Synthetic Pathways for Fe–N–C Fuel Cell Catalysts

Iron–nitrogen–carbon (Fe–N–C) catalysts are among the most promising platinum-group-metal free electrocatalysts for the oxygen reduction reaction in proton-exchange membrane fuel cells. Despite significant experimental progress, rational improvement of Fe–N–C catalytic performance remains limited by incomplete understanding of the synthesis pathways governing active-site formation. In particular, the kinetics associated with high-temperature pyrolysis and Fe impregnation steps are difficult to access experimentally and with traditional atomistic simulation approaches at scale. Here, we leverage machine-learning foundation models to efficiently resolve transition states along candidate synthetic pathways for active site formation within graphitic domains. The foundation models are fine-tuned to Fe–N–C chemical environments using density functional theory reference data, enabling accurate and computationally tractable exploration of complex reaction landscapes. This approach allows systematic comparison of competing pathways and direct evaluation of kinetic barriers associated with active-site formation. By combining transition-state resolution with the scalability of foundation models, this work establishes a rapid, predictive framework for assessing how precursor chemistry and local coordination influence Fe–N–C site formation. The results demonstrate the potential of foundation-model-driven workflows to accelerate the design of carbon-based electrocatalysts and provide mechanistic guidance for synthetic optimization.

Machine-Learning-Guided Optimization of Rapid-Lamp-Annealed Hematite Photoanodes for Solar Water Splitting

Hematite (α-Fe2O3) photoanodes for solar water splitting suffer from poor charge transport and significant sample-to-sample irreproducibility. Here we show that rapid lamp annealing (RLA), in which a graphite susceptor enables rapid volumetric heating, produces hematite films with substantially higher and more reproducible photocurrents than conventional furnace annealing, eliminating the population of inactive electrodes that arises under prolonged calcination. Across 171 solution-deposited samples, RLA films achieved an average photocurrent of ∼0.98 mA cm−2 at 1.23 V vs RHEapproximately four times higher than furnace-annealed sampleswith all RLA samples remaining active. PEIS and XRD descriptors were analyzed using various machine learning models with hierarchical feature clustering and SHapley Additive exPlanations (SHAP), achieving a test R2 of 0.84. SHAP identified bulk and interfacial resistances (R2, R3) and bulk capacitance (C2) as dominant descriptors, with RLA samples consistently occupying the low-resistance, high-C2 region of descriptor space. A bisected-substrate experiment confirmed that this improvement is intrinsic to the annealing method rather than precursor variability. Scanning electron microscopy confirmed that RLA preserves the rod-like nanoparticle morphology of the β-FeOOH precursor, whereas furnace annealing drives grain coalescencedirectly accounting for the improved charge-transport descriptors. These results establish RLA as a mechanistically understood route to high-performance hematite photoanodes and demonstrate that the ML-SHAP framework provides physically interpretable process guidance from datasets of modest size.

Clemson OPEN Clemson OPEN