Periodic Reporting for period 1 - HighHydrogenML (High-throughput Discovery of Catalysts for the Hydrogen Economy through Machine Learning)
Période du rapport: 2023-04-01 au 2025-07-31
Résumé du contexte et des objectifs généraux du projet
The 7th sustainable development goal of the United Nations Organization is to ensure access to affordable, reliable, sustainable, and modern energy for all, while the 13th goal is to take urgent action to combat climate change. Both objectives are closely linked to the use of clean and renewable energy sources for addressing the world’s rising energy needs while limiting environmental effects. However, further expansion of renewable energies (currently 28% of the world’s electricity) is limited because power provided by solar and wind energy (as opposed to coal, natural gas and nuclear) is intermittent and must be coupled with large energy storage capabilities. The most relevant large energy storage capabilities are pumping water uphill (pumped-storage hydroelectricity), batteries and hydrogen storage. While pumped-storage hydroelectricity and batteries have high efficiency, they do not scale to the energy storage needs of large grids powered by intermittent renewable energy due to land restrictions or cost, respectively. On the contrary, hydrogen energy storage offers a unique combination of scalability, long-term storage, and portability, leading to the so-called hydrogen economy.
In fact, the difficulty with hydrogen economy is not with storage but with the production of hydrogen from water and the generation of energy by the oxidation of hydrogen into water. The former process is controlled by the Hydrogen Evolution Reaction (HER), the cathodic reaction during the electrochemical process of dissociation of the water molecule into hydrogen and oxygen. Among the various alternatives for transforming hydrogen into electrical energy, polymer electrolytic membrane fuel cells are among the most interesting and widely applied options in transportation. The main limitation for the industrial application of this technology is the reduced kinetics of the cathode Oxygen Reduction Reaction (ORR). As in the case of the HER, the sluggish kinetics of the ORR can only be improved by increasing the reaction temperature or by using Pt catalysts. Thus, the search for efficient, cheap, durable, and non-toxic catalysts that can replace Pt for the HER and ORR is of dramatic importance for the hydrogen economy but has been unsuccessful so far.
The catalytic activity of a material is controlled by the electronic structure that can be modified by adding other elements to form an alloy or compound or by introducing defects in the crystal lattice. Furthermore, the electronic structure of the atoms on the surfaces is different from that of the atoms inside the material and it also varies between surfaces depending on the crystallographic orientation. Another mechanism to modify the catalytic activity of surfaces is based on elastic strain engineering (ESE), i.e. the application of large elastic deformations to modify the electronic structure. Nevertheless, a systematic application of ESE to search for catalysts for both HER and ORR has not been carried out.
The main objective of the project High-throughput Discovery of Catalysts for the Hydrogen Economy through Machine Learning (HighHydrogenML) is to develop a high-throughput strategy based on first principles calculations and artificial intelligence tools to discover intermetallic compounds whose catalytic activity can be tuned to reach an optimum catalytic performance for the HER and ORR by means of elastic strain engineering. The successful completion of these objectives will provide unique information for experimental synthesis of intermetallic compounds with high catalytic activity for the HER and ORR and could, therefore, open a new avenue for a feasible and efficient hydrogen economy. Moreover, the strategies and tools developed in this project can be applied later to many other catalytic processes of large industrial and/or environmental interest (such as ammonia production, carbon sequestration, etc.).
The specific objectives for achieving this goal are:
i) Construction by means of density functional theory of a reference dataset of adsorption energies of intermetallic compounds for the adsorption of H, O, and OH including the effect of elastic strains.
ii) Development of robust ML models for the prediction of adsorption energies of intermetallic compounds for the HER and ORR including the effect of elastic strains.
iii) Finding of intermetallic compounds that can achieve superior catalytic performance for the HER and ORR through the application of elastic strains.
In fact, the difficulty with hydrogen economy is not with storage but with the production of hydrogen from water and the generation of energy by the oxidation of hydrogen into water. The former process is controlled by the Hydrogen Evolution Reaction (HER), the cathodic reaction during the electrochemical process of dissociation of the water molecule into hydrogen and oxygen. Among the various alternatives for transforming hydrogen into electrical energy, polymer electrolytic membrane fuel cells are among the most interesting and widely applied options in transportation. The main limitation for the industrial application of this technology is the reduced kinetics of the cathode Oxygen Reduction Reaction (ORR). As in the case of the HER, the sluggish kinetics of the ORR can only be improved by increasing the reaction temperature or by using Pt catalysts. Thus, the search for efficient, cheap, durable, and non-toxic catalysts that can replace Pt for the HER and ORR is of dramatic importance for the hydrogen economy but has been unsuccessful so far.
The catalytic activity of a material is controlled by the electronic structure that can be modified by adding other elements to form an alloy or compound or by introducing defects in the crystal lattice. Furthermore, the electronic structure of the atoms on the surfaces is different from that of the atoms inside the material and it also varies between surfaces depending on the crystallographic orientation. Another mechanism to modify the catalytic activity of surfaces is based on elastic strain engineering (ESE), i.e. the application of large elastic deformations to modify the electronic structure. Nevertheless, a systematic application of ESE to search for catalysts for both HER and ORR has not been carried out.
The main objective of the project High-throughput Discovery of Catalysts for the Hydrogen Economy through Machine Learning (HighHydrogenML) is to develop a high-throughput strategy based on first principles calculations and artificial intelligence tools to discover intermetallic compounds whose catalytic activity can be tuned to reach an optimum catalytic performance for the HER and ORR by means of elastic strain engineering. The successful completion of these objectives will provide unique information for experimental synthesis of intermetallic compounds with high catalytic activity for the HER and ORR and could, therefore, open a new avenue for a feasible and efficient hydrogen economy. Moreover, the strategies and tools developed in this project can be applied later to many other catalytic processes of large industrial and/or environmental interest (such as ammonia production, carbon sequestration, etc.).
The specific objectives for achieving this goal are:
i) Construction by means of density functional theory of a reference dataset of adsorption energies of intermetallic compounds for the adsorption of H, O, and OH including the effect of elastic strains.
ii) Development of robust ML models for the prediction of adsorption energies of intermetallic compounds for the HER and ORR including the effect of elastic strains.
iii) Finding of intermetallic compounds that can achieve superior catalytic performance for the HER and ORR through the application of elastic strains.
Travail effectué depuis le début du projet jusqu’à la fin de la période considérée dans le rapport et principaux résultats atteints jusqu’à présent
The research activities of the project HighHydrogenML were organized in four main activities.
1. DFT calculations to determine adsorption energies.
DESCRIPTION OF ACTIVITIES: The goal here was to generate a dataset of adsorption energies for surface slabs of a large number of binary intermetallic compounds with different compositions and lattices (for instance, A3B fcc, A3B hpc, AB bcc, etc.). Adsorption energies were computed for different adsorbates (H, O, and OH) on distinct adsorption sites (fcc AAB, fcc AAA, hcp AAA, hcp AAB, on-top A, and on-top B) and on minimum energy surfaces. In addition, different elastic strains (biaxial tension and biaxial compression) were applied to assess their effect on adsorption energies. The calculations were carried out using density functional theory approximations with a novel GPU implementation of the Quantum Espresso software, which provides huge reductions in computer time. Cut-off energy, k-point sampling, and smearing were decided as a result of convergence tests for each configuration. The GPU cluster of IMDEA was used for initial optimization of the simulations but most calculations were carried out using the computational resources of the Spanish Supercomputing Network and of PRACE (Partnership for advanced computing in Europe).
MAIN ACHIEVEMENTS: A database inlcuding the adsorption energies of H, O, and OH on different adsorption sites of 24 pure metals and 332 binary intermetallic compounds with stoichiometries AB, A2B, and A3B taking into account the effect of biaxial strains.
2. Construction of machine learning (ML) models.
DESCRIPTION OF ACTIVITIES: The goal was to design, train, and validate an ML model able to predict the adsorption energy of H, O, and OH on intermetallic compounds subjected to strains. Three main aspects were considered to converge to the most suitable ML model: descriptors, ML method, and optimization of the ML workflow. State-of-the-art descriptors, as well as new descriptors that were devised from our dataset (for instance, employing approaches like SISSO), were tested for finding the representation that best describes the structure-property relationships of our catalysts. A second important task was the selection of the most appropriate ML method (e.g. gaussian process regression, kernel-based methods, decision trees, etc.) to accurately predict the adsorption energies. This was achieved by finding the best set of hyperparameters (i.e. parameters that need to be fixed before training. For instance, the type of kernel or the regularization scheme) for each type of method to later compare their performance by training models considering different training set size. Here it was also explored the use of explainable machine learning tools for understanding model outputs and for application in the search of new catalysts.
MAIN ACHIEVEMENTS: Extra trees models with high accuracy in the prediction of adsorption energies of H, O, and OH and a novel explainable artificial intelligence strategy based on counterfactual explanations for the design and discovery of new materials.
3. Screening of adsorption energies using the ML models.
DESCRIPTION OF ACTIVITIES: We used the developed ML model to perform a high throughput screening for finding novel candidates for catalysis of the hydrogen evolution reaction (HER) and the oxygen reduction reaction (ORR). The search was focused on binary intermetallic compounds that were experimentally proved, fulfilled certain safety requirements (non-toxic, non-radioactive, non-enviromental hazardous), and that were stable according to Materials Project database. In addition, the screening included the influence of elastic strains.
MAIN ACHIEVEMENTS: A list of potential new candidates for the catalysis of the HER and the ORR that were proposed to experimental colleagues for their experimental realization with the objective of comparing their catalytic performance to that of Pt and to validate our findings regarding the effects of elastic strain engineering on the catalytic properties of materials.
4. Development of an explainable artificial intelligence (XAI) strategy for materials discovery and design.
DESCRIPTION OF ACTIVITIES: We developed a novel XAI strategy for the design and discovery of materials based on the generation of counterfactual explanations.
MAIN ACHIEVEMENTS: The XAI strategy that can be applied to different problems beyond catalysis to design and discover new materials.
1. DFT calculations to determine adsorption energies.
DESCRIPTION OF ACTIVITIES: The goal here was to generate a dataset of adsorption energies for surface slabs of a large number of binary intermetallic compounds with different compositions and lattices (for instance, A3B fcc, A3B hpc, AB bcc, etc.). Adsorption energies were computed for different adsorbates (H, O, and OH) on distinct adsorption sites (fcc AAB, fcc AAA, hcp AAA, hcp AAB, on-top A, and on-top B) and on minimum energy surfaces. In addition, different elastic strains (biaxial tension and biaxial compression) were applied to assess their effect on adsorption energies. The calculations were carried out using density functional theory approximations with a novel GPU implementation of the Quantum Espresso software, which provides huge reductions in computer time. Cut-off energy, k-point sampling, and smearing were decided as a result of convergence tests for each configuration. The GPU cluster of IMDEA was used for initial optimization of the simulations but most calculations were carried out using the computational resources of the Spanish Supercomputing Network and of PRACE (Partnership for advanced computing in Europe).
MAIN ACHIEVEMENTS: A database inlcuding the adsorption energies of H, O, and OH on different adsorption sites of 24 pure metals and 332 binary intermetallic compounds with stoichiometries AB, A2B, and A3B taking into account the effect of biaxial strains.
2. Construction of machine learning (ML) models.
DESCRIPTION OF ACTIVITIES: The goal was to design, train, and validate an ML model able to predict the adsorption energy of H, O, and OH on intermetallic compounds subjected to strains. Three main aspects were considered to converge to the most suitable ML model: descriptors, ML method, and optimization of the ML workflow. State-of-the-art descriptors, as well as new descriptors that were devised from our dataset (for instance, employing approaches like SISSO), were tested for finding the representation that best describes the structure-property relationships of our catalysts. A second important task was the selection of the most appropriate ML method (e.g. gaussian process regression, kernel-based methods, decision trees, etc.) to accurately predict the adsorption energies. This was achieved by finding the best set of hyperparameters (i.e. parameters that need to be fixed before training. For instance, the type of kernel or the regularization scheme) for each type of method to later compare their performance by training models considering different training set size. Here it was also explored the use of explainable machine learning tools for understanding model outputs and for application in the search of new catalysts.
MAIN ACHIEVEMENTS: Extra trees models with high accuracy in the prediction of adsorption energies of H, O, and OH and a novel explainable artificial intelligence strategy based on counterfactual explanations for the design and discovery of new materials.
3. Screening of adsorption energies using the ML models.
DESCRIPTION OF ACTIVITIES: We used the developed ML model to perform a high throughput screening for finding novel candidates for catalysis of the hydrogen evolution reaction (HER) and the oxygen reduction reaction (ORR). The search was focused on binary intermetallic compounds that were experimentally proved, fulfilled certain safety requirements (non-toxic, non-radioactive, non-enviromental hazardous), and that were stable according to Materials Project database. In addition, the screening included the influence of elastic strains.
MAIN ACHIEVEMENTS: A list of potential new candidates for the catalysis of the HER and the ORR that were proposed to experimental colleagues for their experimental realization with the objective of comparing their catalytic performance to that of Pt and to validate our findings regarding the effects of elastic strain engineering on the catalytic properties of materials.
4. Development of an explainable artificial intelligence (XAI) strategy for materials discovery and design.
DESCRIPTION OF ACTIVITIES: We developed a novel XAI strategy for the design and discovery of materials based on the generation of counterfactual explanations.
MAIN ACHIEVEMENTS: The XAI strategy that can be applied to different problems beyond catalysis to design and discover new materials.
Progrès au-delà de l’état des connaissances et impact potentiel prévu (y compris l’impact socio-économique et les conséquences sociétales plus larges du projet jusqu’à présent)
The main results of this project that go beyond the state of the art are:
i) A dataset of adsorption energies of H, O, and OH on 24 pure metals and 332 binary intermetallic compounds including the effect of elastic strains. This result has an impact in academia since the dataset is open source and accessible through Zenodo or Catalysis-Hub and anyone can make use of the data, for instance, to train their own machine learning models.
ii) Machine learning models for the prediction of adsorption energies for H, O, and OH with an accuracy higher than the state-of-the-art. The original datasets and the routine to obtain the models are available in Zenodo and GitHub. Hence, people in academia can take advantage of this freely available tools in order to carry out their own research.
iii) The validation of candidates found through the high-throughput screening and a confirmation, through a collaboration with experimental colleagues, of the effect of elastic strains that push the performance of such candidates closer to that of Pt. Although the found candidates are not yet better than Pt, they do offer a good catalytic activity with a fraction of the cost of Pt-based catalysts. Therefore, they serve as a valuable example for both the academic and industrial sectors for fostering further research into finding new catalysts.
iv) An explainable artificial intelligence (XAI) strategy for the discovery and design of new materials. The proposed strategy is a valuable tool that can be exploited both by the academic and industrial sectors not only in catalysis, but in any area where novel materials are required. This XAI strategy can be used to found new materials with desired properties and, at the same time, obtain explanations on why the found material is better than others.
i) A dataset of adsorption energies of H, O, and OH on 24 pure metals and 332 binary intermetallic compounds including the effect of elastic strains. This result has an impact in academia since the dataset is open source and accessible through Zenodo or Catalysis-Hub and anyone can make use of the data, for instance, to train their own machine learning models.
ii) Machine learning models for the prediction of adsorption energies for H, O, and OH with an accuracy higher than the state-of-the-art. The original datasets and the routine to obtain the models are available in Zenodo and GitHub. Hence, people in academia can take advantage of this freely available tools in order to carry out their own research.
iii) The validation of candidates found through the high-throughput screening and a confirmation, through a collaboration with experimental colleagues, of the effect of elastic strains that push the performance of such candidates closer to that of Pt. Although the found candidates are not yet better than Pt, they do offer a good catalytic activity with a fraction of the cost of Pt-based catalysts. Therefore, they serve as a valuable example for both the academic and industrial sectors for fostering further research into finding new catalysts.
iv) An explainable artificial intelligence (XAI) strategy for the discovery and design of new materials. The proposed strategy is a valuable tool that can be exploited both by the academic and industrial sectors not only in catalysis, but in any area where novel materials are required. This XAI strategy can be used to found new materials with desired properties and, at the same time, obtain explanations on why the found material is better than others.