Surrogate models assisted bmachine learning for coastal Aquifer Management

Κοψιαύτης, Γεώργιος; Kopsiaftis, Georgios

dc.contributor.author	Κοψιαύτης, Γεώργιος	el
dc.contributor.author	Kopsiaftis, Georgios	en
dc.date.accessioned	2024-05-28T08:42:50Z
dc.identifier.uri	https://dspace.lib.ntua.gr/xmlui/handle/123456789/59506
dc.identifier.uri	http://dx.doi.org/10.26240/heal.ntua.27202
dc.rights	Default License
dc.subject	Υφαλμύριση	el
dc.subject	Μηχανική μάθηση	el
dc.subject	Μετα-μοντέλα	el
dc.subject	Βελτιστοποίηση αντλήσεων	el
dc.subject	Μοντέλα πολλαπλής πιστότητας	el
dc.subject	Seawater intrusion	en
dc.subject	Machine learning	en
dc.subject	Surrogate models	en
dc.subject	Pumping optimization	en
dc.subject	Multi-fidelity models	en
dc.title	Surrogate models assisted bmachine learning for coastal Aquifer Management	en
dc.title	Μετα-μοντέλα υποβοηθούμενα από τη Μηχανική Μάθηση για τη διαχείριση παράκτιων υδροφορέων	el
heal.type	doctoralThesis
heal.classification	Γεωεπιστήμες & Επιστήμες Περιβάλλοντος	el
heal.classification	Επιστήμες Ηλεκτρονικών Υπολογιστών και Πληροφορικής	el
heal.dateAvailable	2025-05-27T21:00:00Z
heal.access	embargo
heal.recordProvider	ntua	el
heal.publicationDate	2024-01-05
heal.abstract	The main objective of the present thesis is to investigate new methods that could be used as surrogates models for the complex physical phenomena that occur in coastal aquifers to mitigate the computational challenges associated with applications that require a large number of iterations. To this end, several single- and multi-fidelity models are examined that utilize physical-based models, data-driven models, or their combination. Initially, an extended comparison is performed between the high-fidelity variable density model and the sharp-interface model proposed by Strack (1976) for a wide range of pumping scenarios. In addition, in a similar context we examine two widely-used modifications of the Starck original model. The comparison results indicate that there is a discrepancy between the variable density and the sharp-interface approximations with respect to the extent of seawater intrusion. It should also be noted that the estimation differences between the models strongly depend on the pumping rates and that they vary within the aquifer. To improve the accuracy of the sharp-interface models and eliminate this difference, we propose a multi-fidelity approach that is based on the use of two machine learning methods -specifically, the Random Forest algorithm and the Gaussian Process Regression algorithm. The multi-fidelity approach includes the original Strack model and the machine learning-based correction factor. A single multi-fidelity model is developed for each point of interest, for example, the well locations. Both machine learning algorithms significantly improved the estimation of seawater intrusion. However, the Gaussian Process Regression algorithm outperformed Random Forests in all evaluation metrics. In addition, the proposed method is incorporated into a pumping optimization framework applied in a coastal test aquifer. The calculated optimal pumping rates were comparable with the results of the variable density model. In a subsequent part of the thesis, we investigate the ability of an ML base single-fidelity method to directly simulate the distribution of hydraulic head and solute concentration in coastal aquifers, excluding the use of a second intermediate model of different fidelity. In particular, four ML methods are examined including: (i) Gaussian Process Regression (GPR), (ii) Random Forests (RF), (iii) Support Vector Machines (SVM), and (iv) Linear Regression (LR). To increase the generalization ability of the proposed models, we applied different initial hydraulic head and concentration initial conditions. The initial conditions were included as an additional feature in the input parameters of the machine learning algorithms. The single-fidelity model was also incorporated into a pumping optimization framework to calculate the constraint values. The proposed surrogate-based optimization delivered optimal pumping rates that approximate the global optimum calculated with the variable density model. In the last part of the thesis, we further increase the complexity of the coastal aquifer model by applying time-varying recharge and pumping rates, using a monthly time step. The long-short memory algorithm was utilized to estimate the evolution of the examined aquifer parameters. The proposed method is applied to a real aquifer located on Kalymnos Island. A detailed model is constructed for the aquifer based on all the available geological, geomorphological and land use data. The time-series of pumping rates is derived from the data on water consumption that is currently accessible. The time and spatial variability of the recharge rate is calculated using a distributed recharge model. The proposed deep learning algorithm proved to be significantly efficient in capturing the dynamic behavior of both the hydraulic head and the seawater intrusion front.	en
heal.advisorName	Μαντόγλου, Αριστοτέλης	el
heal.committeeMemberName	Καρατζάς, Γεώργιος	el
heal.committeeMemberName	Νάνου-Γιάνναρου, Αικατερίνη	el
heal.committeeMemberName	Τσιχριντζής, Βασίλειος	el
heal.committeeMemberName	Ναλμπάντης, Ιωάννης	el
heal.committeeMemberName	Δουλάμης, Νικόλαος	el
heal.committeeMemberName	Βουλόδημος, Αθανάσιος	el
heal.academicPublisher	Εθνικό Μετσόβιο Πολυτεχνείο. Σχολή Αγρονόμων και Τοπογράφων Μηχανικών. Τομέας Έργων Υποδομής και Αγροτικής Ανάπτυξης. Εργαστήριο Εγγειοβελτιωτικών Έργων και Διαχείρισης Υδατικών Πόρων	el
heal.academicPublisherID	ntua
heal.numberOfPages	208 σ.	el
heal.fullTextAvailability	false