This research contributes to the advancement of surrogate modelling as a powerful technique in the field of computational simulation that offers numerous advantages for solving complex problems efficiently. In particular, this study emphasizes the pivotal role of surrogate modeling in groundwater management. By integrating key factors like climate change and leveraging machine learning, particularly neu-ral networks, the research facilitates more informed decision-making, significantly reducing the computational cost of complex numerical models. The impact of climate change is a central focus and the first study aims to construct surrogate data-driven models for evaluating climate change effects on groundwater resources, also in the future. The study involves a comparison between statistical methods and different types of artificial neural networks (ANNs). The ef-fectiveness of surrogate models was demonstrated in Northern Tuscany (Italy) but can easily extend to any area of interest. The adopted statistical method involves analyzing historical precipitation and temperature data along with groundwater levels recorded in monitoring wells. Initially, the study explores potential correla-tions between meteorological and groundwater indices; if a correlation is identified, a linear regression analysis is employed to establish relationships between them. These established relationships are then used to estimate future groundwater le-vels based on projected precipitation and temperature obtained from an ensemble of Regional Climate Models, under two Representative Concentration Pathways, namely RCP4.5 and RCP8.5. Then, three distinct Artificial Intelligence (AI) models, Nonlinear AutoRegres-sive with eXogenous inputs (NARX), Long-Short Term Memory (LSTM) and Con-volutional Neural Network (CNN) were implemented to evaluate the impact of cli-mate change on groundwater resources for the same case study. Specifically, these models were trained using directly historical precipitation and temperature data as input to provide groundwater levels as output. Following the training phase, the developed AI models were utilized to forecast future groundwater levels using the same precipitation and temperature projections and climate scenarios described above. The results highlighted different outputs among the models used in this work. However, most of them predict a decrease in groundwater levels as a result of future variations in precipitation and temperature. The study also presents the strengths and weaknesses of each model. Notably, the LSTM model emerges as the most promising approach to predict future groundwater levels. Within the same field, an ANN was developed with the capability to simulate groundwater conditions in the Konya closed basin, Turkey, one of the pilot sites investigated as part of the InTheMED project. This model serves as a tool for examining the potential impacts of climate change and agricultural policies on groundwater resources within the region. The final goal of this application, is to provide a user-friendly tool, based on the trained neural network. The inherent simplicity of the surrogate model, with a straightforward interface and results that are simple to understand, plays a crucial role in decision-making processes. Shifting to pollutant transport, an ANN was implemented to solve different direct and inverse problems. The direct problem deals with the evaluation of con-centrations in monitoring wells, while the inverse problem involves the identifica-tion of contaminant sources and their release history. It demonstrated efficiency in addressing both direct and inverse transport problems, offering reliable results with reduced computational burden. The study also addresses the interpretability challenge of ANNs and the so ca-lled “generalization problem” through Physics-Informed Neural Networks (PINNs). By incorporating physics-based constraints, PINNs bridge the gap between data-driven modeling and physics-based interpretations, offering a promising approach for groundwater numerical simulations. In this study, a PINN is developed to si-mulate flow in an unconfined aquifer. Finally, two extra content are presented. First, an ANN is used to solve an inverse problem in the field of sewer systems. Then, an easily interpretable exam-ple of numerical groundwater flow modeling using spreadsheets, from a didactic perspective, is described. In conclusion, this research underscores the importance of surrogate modeling, machine learning, climate change analysis, and physics-informed approaches in ad-vancing groundwater management strategies and beyond, providing valuable tools for decision-makers to address complex groundwater flow problems in changing environmental conditions.
Questa ricerca propone nuovi avanzamenti nella modellazione surrogata che, nella simulazione di problemi complessi, offre vantaggi rilevanti a supporto della mo-dellazione numerica usuale. In particolare, questo studio sottolinea il ruolo cru-ciale della modellazione surrogata nella gestione delle risorse idriche sotterranee. Integrando fattori chiave come il cambiamento climatico e sfruttando l’appren-dimento automatico, in particolare le reti neurali, lo studio concorre a rendere più facile il processo decisionale informato, riducendo significativamente il costo computazionale dei complessi modelli numerici. L’impatto del cambiamento climatico è al centro dell’attenzione e il primo studio mira a costruire modelli surrogati del tipo "data-driven" per valutare gli effetti del cambiamento climatico sulle risorse idriche sotterranee nel futuro. Esso confronta un metodo statistico e diversi tipi di reti neurali artificiali (ANN) per migliorare la comprensione e facilitare le decisioni nella gestione delle acque sotter-ranee. L’efficacia dei modelli surrogati è stata dimostrata in una applicazione nella Toscana settentrionale, ma può facilmente estendersi a qualsiasi area di interesse. Il metodo statistico adottato coinvolge l’analisi di dati storici sulle precipitazioni e sulla temperatura insieme ai livelli freatici registrati nei pozzi di monitoraggio. Ini-zialmente, lo studio esplora correlazioni potenziali tra indici meteorologici e indici delle acque sotterranee. Se viene individuata una valida correlazione tra questi, si costruisce una regressione lineare che stabilisce una relazione tra di essi. Queste relazioni vengono poi utilizzate per stimare futuri livelli di falda sulla base delle proiezioni di precipitazione e temperatura ottenute da un insieme di Modelli Cli-matici Regionali, considerando due Scenari di Emissione Rappresentativi, ovvero RCP4.5 e RCP8.5. Successivamente, sono stati implementati tre distinti modelli di Intelligenza Ar-tificiale (AI), Rete neurale Autoregressiva con Ingressi Eterogenei (NARX), Rete Neurale con Memoria a Lungo e Breve Termine (LSTM) e Rete Neurale Convo-luzionale (CNN), per valutare l’impatto del cambiamento climatico sui livelli di falda per lo stesso caso di studio. In particolare, questi modelli sono stati adde-strati utilizzando direttamente dati storici di precipitazioni e di temperatura come input e per fornire i livelli freatici come output. Dopo la fase di addestramento, i modelli di AI sviluppati sono stati utilizzati per prevedere i livelli delle acque sotterranee utilizzando le stesse proiezioni di precipitazioni e temperatura e gli scenari climatici descritti in precedenza. I risultati hanno evidenziato diversi out-put tra i modelli utilizzati in questo studio. Tuttavia, la maggior parte di essi prevede una diminuzione dei livelli di falda a seguito di future variazioni di pre-cipitazione e temperatura. Lo studio presenta anche i punti di forza e debolezza di ciascun modello. In particolare, il modello LSTM emerge come l’approccio più promettente per prevedere i futuri livelli di falda. Nello stesso campo, è stata sviluppata una rete neurale artificiale con la capa-cità di simulare lo stato dell’acquifero nel bacino di Konya, Turchia, uno dei siti pilota indagati nell’ambito del progetto InTheMED. Questo modello si compor-ta da strumento per esaminare gli impatti potenziali del cambiamento climatico e delle politiche agricole sulle risorse idriche sotterranee. L’obiettivo finale di questa applicazione è fornire uno strumento "user-friendly" basato sulla rete neu-rale addestrata. La semplicità intrinseca del modello surrogato, sviluppato con un’interfaccia chiara e risultati di facile comprensione, svolge un ruolo cruciale nei processi decisionali. Passando al trasporto di inquinanti, è stata implementata una rete neurale artificiale per risolvere diversi problemi diretti e inversi. Il problema diretto ri-guarda la valutazione delle concentrazioni nei pozzi di monitoraggio, mentre il problema inverso comporta l’identificazione delle fonti di contaminazione e la loro storia di rilascio. La tecnica ha dimostrato efficienza nell’affrontare sia problemi diretti che inversi di trasporto, offrendo risultati affidabili con un ridotto onere computazionale. Lo studio affronta anche la sfida dell’interpretabilità fisica delle reti neurali artificiali e del cosiddetto "problema della generalizzazione" attraverso le Reti Neurali Fisicamente Basate (PINN). Integrando vincoli basati sulla fisica, le PINN colmano il divario tra la modellazione basata sui dati e i modelli numerici costruiti sulle equazioni differenziali dedotte dalla fisica, offrendo un approccio promettente per le simulazioni numeriche delle acque sotterranee. In questo studio, una PINN è stata sviluppata per simulare il flusso in un acquifero non confinato. Infine, vengono presentati due contenuti aggiuntivi. Nel primo, una rete neurale artificiale è utilizzata per risolvere un problema inverso nel campo dei sistemi fognari. In secondo, un esempio di modellazione numerica del flusso delle acque sotterranee mediante fogli di calcolo con una ottima prospettiva didattica. In conclusione, questa ricerca sottolinea l’importanza della modellazione sur-rogata, dell’apprendimento automatico, dell’analisi del cambiamento climatico e degli approcci basati sulla fisica per progredire nelle strategie di gestione delle ac-que sotterranee e affrontare sfide complesse, offrendo strumenti preziosi ai decisori.
Surrogate models, physics-informed neural networks and climate change
Daniele, Secci
2024
Abstract
This research contributes to the advancement of surrogate modelling as a powerful technique in the field of computational simulation that offers numerous advantages for solving complex problems efficiently. In particular, this study emphasizes the pivotal role of surrogate modeling in groundwater management. By integrating key factors like climate change and leveraging machine learning, particularly neu-ral networks, the research facilitates more informed decision-making, significantly reducing the computational cost of complex numerical models. The impact of climate change is a central focus and the first study aims to construct surrogate data-driven models for evaluating climate change effects on groundwater resources, also in the future. The study involves a comparison between statistical methods and different types of artificial neural networks (ANNs). The ef-fectiveness of surrogate models was demonstrated in Northern Tuscany (Italy) but can easily extend to any area of interest. The adopted statistical method involves analyzing historical precipitation and temperature data along with groundwater levels recorded in monitoring wells. Initially, the study explores potential correla-tions between meteorological and groundwater indices; if a correlation is identified, a linear regression analysis is employed to establish relationships between them. These established relationships are then used to estimate future groundwater le-vels based on projected precipitation and temperature obtained from an ensemble of Regional Climate Models, under two Representative Concentration Pathways, namely RCP4.5 and RCP8.5. Then, three distinct Artificial Intelligence (AI) models, Nonlinear AutoRegres-sive with eXogenous inputs (NARX), Long-Short Term Memory (LSTM) and Con-volutional Neural Network (CNN) were implemented to evaluate the impact of cli-mate change on groundwater resources for the same case study. Specifically, these models were trained using directly historical precipitation and temperature data as input to provide groundwater levels as output. Following the training phase, the developed AI models were utilized to forecast future groundwater levels using the same precipitation and temperature projections and climate scenarios described above. The results highlighted different outputs among the models used in this work. However, most of them predict a decrease in groundwater levels as a result of future variations in precipitation and temperature. The study also presents the strengths and weaknesses of each model. Notably, the LSTM model emerges as the most promising approach to predict future groundwater levels. Within the same field, an ANN was developed with the capability to simulate groundwater conditions in the Konya closed basin, Turkey, one of the pilot sites investigated as part of the InTheMED project. This model serves as a tool for examining the potential impacts of climate change and agricultural policies on groundwater resources within the region. The final goal of this application, is to provide a user-friendly tool, based on the trained neural network. The inherent simplicity of the surrogate model, with a straightforward interface and results that are simple to understand, plays a crucial role in decision-making processes. Shifting to pollutant transport, an ANN was implemented to solve different direct and inverse problems. The direct problem deals with the evaluation of con-centrations in monitoring wells, while the inverse problem involves the identifica-tion of contaminant sources and their release history. It demonstrated efficiency in addressing both direct and inverse transport problems, offering reliable results with reduced computational burden. The study also addresses the interpretability challenge of ANNs and the so ca-lled “generalization problem” through Physics-Informed Neural Networks (PINNs). By incorporating physics-based constraints, PINNs bridge the gap between data-driven modeling and physics-based interpretations, offering a promising approach for groundwater numerical simulations. In this study, a PINN is developed to si-mulate flow in an unconfined aquifer. Finally, two extra content are presented. First, an ANN is used to solve an inverse problem in the field of sewer systems. Then, an easily interpretable exam-ple of numerical groundwater flow modeling using spreadsheets, from a didactic perspective, is described. In conclusion, this research underscores the importance of surrogate modeling, machine learning, climate change analysis, and physics-informed approaches in ad-vancing groundwater management strategies and beyond, providing valuable tools for decision-makers to address complex groundwater flow problems in changing environmental conditions.File | Dimensione | Formato | |
---|---|---|---|
Secci_PhD_Thesis_UNIPR_pdfA_v2.pdf
accesso aperto
Dimensione
22.35 MB
Formato
Adobe PDF
|
22.35 MB | Adobe PDF | Visualizza/Apri |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/196741
URN:NBN:IT:UNIPR-196741