In the current energy transition scenario, it is crucial to reach the fundamental understanding of the catalytic process enabling advanced design of new technologies. First-principles-based multiscale modeling has emerged as an effective tool to assist this process by coupling the advanced description of the transport phenomena with detailed atomistically-resolved heterogenous chemistry. However, the benefits of such models are often hindered by prohibitive computational costs which prevent routine adoption in industrial applications. Therefore, it is necessary to develop methodologies able to provide the in-depth description granted by multiscale simulations, at an affordable computational cost. In this context, surrogate models are a potential candidate to overcome such limitations since they can approximate complex systems (e.g. multiscale models of catalytic reactors), by learning the relationships between inputs and outputs from existing data, resulting in computationally inexpensive digital twin of the original system. This work assesses the performances of Artificial Neural Networks (ANN), a popular machine learning (ML) technique, for the generation of surrogate models of detailed multiscale models of catalytic chemical reactors. In particular, a semi-automatic workflow for the generation of surrogate models of chemical reactors has been developed. To this aim, the performances of two different ANN architectures have been considered. First, Feed Forward Neural Networks (FFNN) have been tested followed by Recurrent Neural Networks (RNN). Both networks have shown similar performances in terms of accuracy and efficiency in simple system. Therefore, FFNNs have been considered for the approximation of multiscale system due to the simpler structure and training in this context. Then, two distinct network topologies have been analyzed. Initially, a single FFNN has been used to simultaneously predict all the output variables (i.e. SingleNet). Then, a MultiNet approach has been considered where each of the output variables is predicted by a corresponding network. The latter shows superior performances in terms of accuracy with negligible additional computational burden. This work also attempted to improve the model performances by imposing hard constraints on the network topology to enforce mass conservation by the adoption of a softmax activation in the last layer when the output quantities are mass fractions and coverages. Parametrical investigations of several hyperparameters have been also carried out to identify the optimal settings for the systems hereby considered. The potential of the envisioned approach has been assessed for two different purposes. On the one hand, the generation of surrogate models of macroscopic reactor models (e.g. batch and 1D pseudo-homogeneous packed bed reactors) has been tested. Systems of growing complexity have been considered moving from simple homogenous batch reactors to non-isothermal packed bed reactors. In all the conditions, the FFNN-based surrogate models have demonstrated excellent accuracy and a relevant computational speed-up (up to 18 times). On the other hand, the methodology has been also applied for the approximation of the chemical sub-step in detailed Computational Fluid Dynamics simulation of catalytic reactors. Also here, the FFNN-based surrogates have demonstrated to provide accurate and fast predictions avoiding to solve the expensive chemical sub-step and resulting in a simulation speed-up around 4. As a whole, this work demonstrates the potential of ANN-based surrogate models of chemical reactors and develops a workflow aiming at the efficient generation of digital twins of complex chemical engineering problems.
Nell’attuale scenario di transizione energetica, è cruciale raggiungere una profonda comprensione dei processi catalitici che consenta la progettazione avanzata di nuove tecnologie. La modellazione multiscala basata su un approccio first-principles è emersa come strumento efficace per assistere questo processo, accoppiando la descrizione avanzata dei fenomeni di trasporto ad una chimica eterogenea dettagliata e risolta fino al livello atomico. Tuttavia, i vantaggi ottenibili con tali modelli sono spesso ridotti dai proibitivi costi computazionali che ne impediscono l’adozione di routine nelle applicazioni di rilevanza industriale. Pertanto, è necessario sviluppare metodologie in grado di fornire una descrizione approfondita, garantita dalla modellazione multiscala, ad un costo computazionale sostenibile. In questo contesto, i modelli surrogati sono un potenziale candidato per superare tali limiti, poiché possono approssimare sistemi complessi (e.g. modelli multiscala di reattori catalitici), imparando le relazioni che esistono tra input e output di dati esistenti, generando un digital-twin del sistema originale computazionalmente poco costoso. Questo lavoro di tesi verifica le prestazioni delle reti neurali artificiali (ANN), una popolare tecnica di machine learning (ML), per la generazione di modelli surrogati di modelli multiscala dettagliati di reattori chimici catalitici. A questo scopo sono state considerate le prestazioni di due diverse architetture di reti neurali. Innanzitutto, sono state testate le reti neurali feed-forward (FFNN), seguite dalle reti neurali ricorrenti (RNN). Entrambe le reti hanno mostrato prestazioni simili in termini di accuratezza ed efficienza in sistemi semplici. Le reti feed-forward sono state perciò prese in considerazione per l’approssimazione di sistemi multiscala, a causa della struttura più semplice e della più facile gestione in questo contesto. In seguito, sono state analizzate due distinte topologie di rete feed-forward. Inizialmente, è stata utilizzata una singola FFNN per predire simultaneamente tutte le variabili di output (i.e. SingleNet). Poi, è stato considerato un approccio MultiNet in cui per ciascuna delle variabili di output è stata considerata una rete dedicata. Quest’ultimo approccio ha mostrato prestazioni superiori in termini di accuratezza con un onere computazionale aggiuntivo trascurabile. Questo lavoro ha anche tentato di migliorare le prestazioni del modello, imponendo vincoli sulla topologia di rete, per imporre la conservazione della massa mediante l’introduzione di una funzione di attivazione softmax nell’ultimo layer nei casi in cui le quantità di output sono frazioni massive e coverage. Sono state inoltre effettuate delle analisi parametriche su diversi iperparametri per identificare le impostazioni ottimali per i sistemi qui considerati. Il potenziale dell’approccio è stato valutato per due diversi scopi. Da un lato, è stata testata la generazione di modelli surrogati di modelli di reattori macroscopici. Sono stati presi in considerazione sistemi di complessità crescente, da semplici reattori batch omogenei a reattori a letto impaccato non isotermici. In tutte le condizioni, i modelli surrogati basati su FFNN hanno dimostrato un’eccellente accuratezza e un notevole speed-up (fino a 18 volte). D’altra parte, la metodologia è stata applicata anche per l’approssimazione dello step chimico in simulazioni dettagliate di fluidodinamica computazionale di reattori catalitici. Anche in questo caso, i surrogati basati su FFNN hanno dimostrato di fornire predizioni accurate e veloci, evitando di risolvere il costoso step chimico, e determinando uno speed-up pari circa a 4. Nel complesso, questo lavoro dimostra il potenziale dei modelli surrogati di reattori chimici, basati su ANN, e sviluppa una metodologia mirata alla generazione efficiente di digital-twin, per affrontare i complessi problemi di ingegneria chimica.
Development and assessment of ANN-based surrogate models of heterogeneous catalytic reactors
Speri, Andrea
2021/2022
Abstract
In the current energy transition scenario, it is crucial to reach the fundamental understanding of the catalytic process enabling advanced design of new technologies. First-principles-based multiscale modeling has emerged as an effective tool to assist this process by coupling the advanced description of the transport phenomena with detailed atomistically-resolved heterogenous chemistry. However, the benefits of such models are often hindered by prohibitive computational costs which prevent routine adoption in industrial applications. Therefore, it is necessary to develop methodologies able to provide the in-depth description granted by multiscale simulations, at an affordable computational cost. In this context, surrogate models are a potential candidate to overcome such limitations since they can approximate complex systems (e.g. multiscale models of catalytic reactors), by learning the relationships between inputs and outputs from existing data, resulting in computationally inexpensive digital twin of the original system. This work assesses the performances of Artificial Neural Networks (ANN), a popular machine learning (ML) technique, for the generation of surrogate models of detailed multiscale models of catalytic chemical reactors. In particular, a semi-automatic workflow for the generation of surrogate models of chemical reactors has been developed. To this aim, the performances of two different ANN architectures have been considered. First, Feed Forward Neural Networks (FFNN) have been tested followed by Recurrent Neural Networks (RNN). Both networks have shown similar performances in terms of accuracy and efficiency in simple system. Therefore, FFNNs have been considered for the approximation of multiscale system due to the simpler structure and training in this context. Then, two distinct network topologies have been analyzed. Initially, a single FFNN has been used to simultaneously predict all the output variables (i.e. SingleNet). Then, a MultiNet approach has been considered where each of the output variables is predicted by a corresponding network. The latter shows superior performances in terms of accuracy with negligible additional computational burden. This work also attempted to improve the model performances by imposing hard constraints on the network topology to enforce mass conservation by the adoption of a softmax activation in the last layer when the output quantities are mass fractions and coverages. Parametrical investigations of several hyperparameters have been also carried out to identify the optimal settings for the systems hereby considered. The potential of the envisioned approach has been assessed for two different purposes. On the one hand, the generation of surrogate models of macroscopic reactor models (e.g. batch and 1D pseudo-homogeneous packed bed reactors) has been tested. Systems of growing complexity have been considered moving from simple homogenous batch reactors to non-isothermal packed bed reactors. In all the conditions, the FFNN-based surrogate models have demonstrated excellent accuracy and a relevant computational speed-up (up to 18 times). On the other hand, the methodology has been also applied for the approximation of the chemical sub-step in detailed Computational Fluid Dynamics simulation of catalytic reactors. Also here, the FFNN-based surrogates have demonstrated to provide accurate and fast predictions avoiding to solve the expensive chemical sub-step and resulting in a simulation speed-up around 4. As a whole, this work demonstrates the potential of ANN-based surrogate models of chemical reactors and develops a workflow aiming at the efficient generation of digital twins of complex chemical engineering problems.| File | Dimensione | Formato | |
|---|---|---|---|
|
2022_12_Speri_Exec_Summ.pdf
non accessibile
Descrizione: Executive Summary
Dimensione
648.43 kB
Formato
Adobe PDF
|
648.43 kB | Adobe PDF | Visualizza/Apri |
|
2022_12_Speri.pdf
non accessibile
Descrizione: Tesi
Dimensione
3.5 MB
Formato
Adobe PDF
|
3.5 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/201389