The urgent need to transition to a green economy has generated significant interest in ammonia (NH3) as a hydrogen carrier. To make this possible, efficient routes for extracting hydrogen via NH3 decomposition must be found. NH3 thermal decomposition on iron-based catalysts has received much interest, as they offer an attractive balance between catalytic efficiency and sustainability. However, an atomic-level understanding of the reaction mechanisms, particularly under high-temperature operando conditions, has not been fully elucidated. Traditional models, often based on static representations of the catalyst, fail to capture the dynamic nature of surface reactions, which is crucial for explaining its activity and long-term stability. In this thesis, we present a comprehensive computational investigation of NH3 decomposition on Fe and iron-cobalt (FeCo) catalysts, addressing these challenges via molecular dynamics simulations powered by machine learning (ML) and enhanced sampling techniques. We first investigated NH3 decomposition on Fe(110) and Fe(111) surfaces, representing the two limiting cases. Our simulations revealed that dynamic effects at operando temperatures critically influence all reaction stages. We demonstrated how the mobility of surface atoms influences diffusion and dehydrogenation processes, leading to an ensemble of reactive pathways rather than a well-defined one. Nonetheless, we found a unified description of the reaction mechanism through charge transfer analysis along these pathways. We then studied the fate and the influence of adsorbed nitrogen (N*) on the surface morphology and activity. For Fe(110), we investigated the nitridation mechanisms during NH3 decomposition in collaboration with experimentalists, who observed the formation and decomposition of Fe nitrides during catalyst activation and NH3 decomposition. Our simulations provide insights into the early stages of nitridation by showing that N* dissolution is favored over recombinative desorption, affecting its storage and release dynamics. Similarly, we studied the influence of a finite N* coverage on NH3 synthesis particularly on the N2 cleavage step for the Fe(111) surface. At high temperatures, this surface presents a surface atom diffusivity, and we found that N* atoms stabilize triangular surface structures, a sign of a frustrated phase transition at the surface. Here, the effects of dynamics are even more relevant, as the reduction in the number of active sites is counterbalanced by an increase in their lifetime. We then studied ammonia decomposition on FeCo alloys, which exhibit a higher catalytic activity, but which have not been studied before due to the high complexity and prohibitive computation cost. Thus, we developed a novel data-efficient protocol for constructing reactive ML potentials, focusing on dataset construction by targeting all relevant, particularly reactive, configurations via local environment uncertainty. This allowed us to characterize the mechanisms of several reactions with only a small fraction of the usual computational effort. Our simulations showed a high similarity between FeCo(110) and Fe(110) catalytic mechanisms across the elementary steps of the reaction. We ascertained a double promotional effect of cobalt, which reduces the reaction's rate-determining step barrier and increases the N* dissolution ones, preventing catalyst nitridation. This microscopic interpretation is confirmed by experimental observations that no bulk nitrides are formed. In conclusion, our findings indicate that it is the delicate balance between different surface processes that explains catalytic activity, such as the formation of iron surface- or bulk-nitrides, and how iron-cobalt prevents it. This underscores the complexity of heterogeneous catalysis and its intrinsic dynamic nature, paving the way for even more realistic simulations.
L’urgente transizione ecologica ha generato un notevole interesse per l'ammoniaca (NH3) come vettore di idrogeno. Per renderlo possibile, è essenziale trovare metodi efficienti per estrarre l’idrogeno dall’NH3. La decomposizione termica di NH3 su catalizzatori a base di ferro (Fe) ha attirato attenzione offrendo un equilibrio ottimale tra efficienza catalitica e sostenibilità. Tuttavia, una comprensione a livello atomico dei meccanismi di reazione, specialmente in condizioni operative, rimane ancora parziale. I modelli tradizionali, basati su rappresentazioni statiche del catalizzatore, non riescono infatti a catturarne la natura dinamica che è cruciale per spiegarne l'attività e la stabilità a lungo termine. In questa tesi presentiamo uno studio computazionale sulla decomposizione dell'NH3 sul Fe e ferro-cobalto (FeCo), utilizzando simulazioni di dinamica molecolare supportate da tecniche machine learning (ML) e di campionamento avanzato. Inizialmente, abbiamo investigato le superfici Fe(110) e Fe(111), che rappresentano i casi limite. Le simulazioni hanno mostrato come, alle temperature operative, la dinamica influenza tutte le fasi della reazione. La mobilità degli atomi superficiali altera la diffusione e gli step di deidrogenazione, portando a molteplici percorsi reattivi. Tuttavia, grazie all’'analisi del trasferimento di carica lungo questi percorsi abbiamo individuato una descrizione unificata del meccanismo di reazione, identificando l’attività catalitica del Fe. Successivamente, abbiamo studiato il destino è l’influenza dell'azoto adsorbito (N*) sulla morfologia supeficiale ed attività. Sulla (110), abbiamo investigato la formazione dei nitruri di Fe, in collaborazione con partner sperimentali, che hanno osservato la formazione e decomposizione di questi durante l'attivazione del catalizzatore e la decomposizione di NH3. Dalle nostre simulazioni è emerso che la dissoluzione di N* è favorita rispetto alla ricombinazione, razionalizzando le evidenze sperimentali. Invece, per la (111), che mostra una maggiore mobilità superficiale, abbiamo riscontrato che gli atomi di N* stabilizzano strutture triangolari, segno di una transizione di fase sulla superficie. Qui, gli effetti della dinamica sono ancora più rilevanti, poiché la riduzione del numero di siti attivi è compensata da un aumento della loro durata, risultando in una barriera di dissociazione dell'N2 sostanzialmente invariata nelle coperture studiate. Abbiamo poi indagato il FeCo, recentemente proposto per la sua maggiore attività rispetto al Fe. A causa della complessità e del costo computazionale, questa lega non è mai stata studiata tramite MD. Pertanto, abbiamo sviluppato un protocollo per costruire potenziali reattivi basati su ML in modo altamente efficiente, concentrandoci sull'individuazione delle sole configurazioni necessarie tramite l'incertezza dell'ambiente locale. Questo approccio ci ha permesso di caratterizzare i meccanismi di diverse reazioni utilizzando solo una frazione (~1/20) dell'usuale costo computazionale. Le simulazioni hanno mostrato una sostanziale somiglianza tra i meccanismi catalitici di FeCo(110) e Fe(110). Abbiamo scoperto che il cobalto ha un doppio effetto promozionale: ridurre la barriera dello step limitante di reazione e penalizzare la dissoluzione di N*, prevenendo quindi la formazione di nitruri. Questa interpretazione microscopica è stata confermata dalle osservazioni sperimentali, secondo cui non si formano nitruri nel bulk. I risultati di questa tesi indicano come sia il delicato equilibrio tra diversi processi superficiali a spiegare l'attività catalitica, come la formazione di nitruri di Fe in superficie o in bulk in determinate condizioni, ed il perchè la lega previene. In conclusione, i nostri risultati sottolineano la complessità della catalisi eterogenea e la sua intrinseca natura dinamica, aprendo la strada a simulazioni ancora più realistiche.
(2025). Machine Learning and Molecular Dynamics in operando simulations of catalytic reactions. (Tesi di dottorato, , 2025).
Machine Learning and Molecular Dynamics in operando simulations of catalytic reactions
PEREGO, SIMONE
2025
Abstract
The urgent need to transition to a green economy has generated significant interest in ammonia (NH3) as a hydrogen carrier. To make this possible, efficient routes for extracting hydrogen via NH3 decomposition must be found. NH3 thermal decomposition on iron-based catalysts has received much interest, as they offer an attractive balance between catalytic efficiency and sustainability. However, an atomic-level understanding of the reaction mechanisms, particularly under high-temperature operando conditions, has not been fully elucidated. Traditional models, often based on static representations of the catalyst, fail to capture the dynamic nature of surface reactions, which is crucial for explaining its activity and long-term stability. In this thesis, we present a comprehensive computational investigation of NH3 decomposition on Fe and iron-cobalt (FeCo) catalysts, addressing these challenges via molecular dynamics simulations powered by machine learning (ML) and enhanced sampling techniques. We first investigated NH3 decomposition on Fe(110) and Fe(111) surfaces, representing the two limiting cases. Our simulations revealed that dynamic effects at operando temperatures critically influence all reaction stages. We demonstrated how the mobility of surface atoms influences diffusion and dehydrogenation processes, leading to an ensemble of reactive pathways rather than a well-defined one. Nonetheless, we found a unified description of the reaction mechanism through charge transfer analysis along these pathways. We then studied the fate and the influence of adsorbed nitrogen (N*) on the surface morphology and activity. For Fe(110), we investigated the nitridation mechanisms during NH3 decomposition in collaboration with experimentalists, who observed the formation and decomposition of Fe nitrides during catalyst activation and NH3 decomposition. Our simulations provide insights into the early stages of nitridation by showing that N* dissolution is favored over recombinative desorption, affecting its storage and release dynamics. Similarly, we studied the influence of a finite N* coverage on NH3 synthesis particularly on the N2 cleavage step for the Fe(111) surface. At high temperatures, this surface presents a surface atom diffusivity, and we found that N* atoms stabilize triangular surface structures, a sign of a frustrated phase transition at the surface. Here, the effects of dynamics are even more relevant, as the reduction in the number of active sites is counterbalanced by an increase in their lifetime. We then studied ammonia decomposition on FeCo alloys, which exhibit a higher catalytic activity, but which have not been studied before due to the high complexity and prohibitive computation cost. Thus, we developed a novel data-efficient protocol for constructing reactive ML potentials, focusing on dataset construction by targeting all relevant, particularly reactive, configurations via local environment uncertainty. This allowed us to characterize the mechanisms of several reactions with only a small fraction of the usual computational effort. Our simulations showed a high similarity between FeCo(110) and Fe(110) catalytic mechanisms across the elementary steps of the reaction. We ascertained a double promotional effect of cobalt, which reduces the reaction's rate-determining step barrier and increases the N* dissolution ones, preventing catalyst nitridation. This microscopic interpretation is confirmed by experimental observations that no bulk nitrides are formed. In conclusion, our findings indicate that it is the delicate balance between different surface processes that explains catalytic activity, such as the formation of iron surface- or bulk-nitrides, and how iron-cobalt prevents it. This underscores the complexity of heterogeneous catalysis and its intrinsic dynamic nature, paving the way for even more realistic simulations.File | Dimensione | Formato | |
---|---|---|---|
phd_unimib_813901.pdf
accesso aperto
Descrizione: PhD_thesis_Perego_813901
Tipologia di allegato:
Doctoral thesis
Dimensione
36.37 MB
Formato
Adobe PDF
|
36.37 MB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.