{"id":{"repo_id":"mtu","oai_identifier":"oai:digitalcommons.mtu.edu:etdr-3000"},"canonical_url":"https://search.ndltd.org/etd/mtu/oai:digitalcommons.mtu.edu:etdr-3000","repository":{"repo_id":"mtu","name":"Michigan Technological University","base_url":"https://digitalcommons.mtu.edu/do/oai/"},"display":{"title":"A MULTI-LOSS WASSERSTEIN GAN FRAMEWORK FOR GEOSTATISTICAL SIMULATION","abstract":"Geostatistical simulation plays a crucial role in modeling geological structures of the subsurface, providing equally plausible realizations for inference and practical applications. It serves as a powerful tool for estimating earth resources by accounting for the stochastic nature of geological systems. Among the various approaches, Multiple-Point Statistics (MPS) algorithms stand out for their ability to capture and reproduce complex spatial relationships and patterns in geological data that traditional geostatistical methods often fail to represent adequately. A fundamental aspect of MPS algorithms is the use of a training image (TI), which encapsulates the spatial distribution and properties of geological features. In mining, TIs represent geological features such as lithological maps, fault networks, or spatial variability in properties like metal grades and density. Similarly, in fields like petroleum engineering and hydrology, TIs model reservoir distributions, flow characteristics, and contaminant transport. The accuracy of these simulations depends heavily on the quality of the TI, but data limitations and selection biases often constrain their effectiveness. This thesis introduces a Multi-loss Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) framework as an innovative tool for geostatistical simulation. The proposed framework leverages the WGAN-GP architecture alongside multi-loss functions, including feature matching loss, pixel intensity loss, structural similarity index loss, and conditioning loss (for conditional simulations). These losses enhance the model's ability to generate realizations that faithfully reproduce the statistical and structural properties of the TI. To address limitations associated with limited TI data and selection bias, image augmentation techniques were integrated into the model, introducing diversity and uncertainty while expanding the training dataset. The WGAN-GP model was evaluated under both conditional and unconditional simulation scenarios. Results demonstrated its capability to preserve structural and spatial heterogeneity. Two key evaluation metrics were employed: (1) ensuring the generated simulations met a structural similarity index threshold, and (2) verifying that the simulations reproduced the spatial properties of the TI through variogram analysis. Statistically, the WGAN-GP framework, augmented with multi-loss functions, not only generated consistent geological realizations but also achieved computational efficiency—a significant advantage given the typically high computational cost of deep learning models. The results from the model demonstrate that the WGAN-GP framework, combined with multi-loss functions, not only generates consistent geological realizations but also offers a significant computational advantage, addressing the high computational demands of most deep learning models. This research emphasizes the importance of generative adversarial networks and their potential applications in geostatistical simulations.","abstract_html":"Geostatistical simulation plays a crucial role in modeling geological structures of the subsurface, providing equally plausible realizations for inference and practical applications. It serves as a powerful tool for estimating earth resources by accounting for the stochastic nature of geological systems. Among the various approaches, Multiple-Point Statistics (MPS) algorithms stand out for their ability to capture and reproduce complex spatial relationships and patterns in geological data that traditional geostatistical methods often fail to represent adequately. A fundamental aspect of MPS algorithms is the use of a training image (TI), which encapsulates the spatial distribution and properties of geological features. In mining, TIs represent geological features such as lithological maps, fault networks, or spatial variability in properties like metal grades and density. Similarly, in fields like petroleum engineering and hydrology, TIs model reservoir distributions, flow characteristics, and contaminant transport. The accuracy of these simulations depends heavily on the quality of the TI, but data limitations and selection biases often constrain their effectiveness. This thesis introduces a Multi-loss Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) framework as an innovative tool for geostatistical simulation. The proposed framework leverages the WGAN-GP architecture alongside multi-loss functions, including feature matching loss, pixel intensity loss, structural similarity index loss, and conditioning loss (for conditional simulations). These losses enhance the model&#x27;s ability to generate realizations that faithfully reproduce the statistical and structural properties of the TI. To address limitations associated with limited TI data and selection bias, image augmentation techniques were integrated into the model, introducing diversity and uncertainty while expanding the training dataset. The WGAN-GP model was evaluated under both conditional and unconditional simulation scenarios. Results demonstrated its capability to preserve structural and spatial heterogeneity. Two key evaluation metrics were employed: (1) ensuring the generated simulations met a structural similarity index threshold, and (2) verifying that the simulations reproduced the spatial properties of the TI through variogram analysis. Statistically, the WGAN-GP framework, augmented with multi-loss functions, not only generated consistent geological realizations but also achieved computational efficiency—a significant advantage given the typically high computational cost of deep learning models. The results from the model demonstrate that the WGAN-GP framework, combined with multi-loss functions, not only generates consistent geological realizations but also offers a significant computational advantage, addressing the high computational demands of most deep learning models. This research emphasizes the importance of generative adversarial networks and their potential applications in geostatistical simulations.","abstract_has_math":false,"creators":["Addai, Charles O."],"institution":null,"degree_name":"Master of Science in Mining Engineering (MS)","degree_level":"Campus Access Master's Thesis","degree_discipline":"Department of Geological and Mining Engineering and Sciences","degree_department":null,"school":null,"contributors":["Snehamoy Chatterjee"],"advisors":[],"committee_chairs":[],"committee_members":["Rouhollah (Radwin) Askari","Sidike Paheding"],"year":null,"date_issued":"","date_published":null,"updated_at":"2026-11-19T08:19:12Z","subjects":["Generative Adversarial Network","Loss function","Geostatistics","Simulation","equiprobable realization","Artificial Intelligence and Robotics","Computational Engineering","Geological Engineering","Mining Engineering","Oil, Gas, and Energy","Programming Languages and Compilers"],"languages":[],"rights":[],"rights_urls":[],"identifier_entries":[{"key":"dc:identifier","label":"Identifier","values":["https://digitalcommons.mtu.edu/etdr/1812"],"render_values":[{"text":"https://digitalcommons.mtu.edu/etdr/1812","href":"https://digitalcommons.mtu.edu/etdr/1812","code":true}]}],"person_identifier_entries":[]},"links":{"outbound_url":"https://doi.org/10.37099/mtu.dc.etdr/1812","outbound_label":"DOI","outbound_source":"dc:identifier","record_url":"https://doi.org/10.37099/mtu.dc.etdr/1812","record_label":"Repository record","record_kind":"doi","record_source":"dc:identifier","download_url":null,"download_label":null,"download_source":null,"access_status":null,"embargo_until":null,"access_note":null},"source_record":{"url":"https://digitalcommons.mtu.edu/do/oai/?verb=GetRecord&metadataPrefix=oai_etdms&identifier=oai%3Adigitalcommons.mtu.edu%3Aetdr-3000","prefix":"oai_etdms"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor","label":"Contributor","values":["Snehamoy Chatterjee"]},{"key":"dc:contributor.committeemember","label":"Committee Member","values":["Rouhollah (Radwin) Askari","Sidike Paheding"]},{"key":"dc:creator","label":"Author","values":["Addai, Charles O."]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date.available","label":"Dc Date Available","values":["2025-12-31T08:00:00Z"]},{"key":"dc:date.created","label":"Dc Date Created","values":["2024-01-01T08:00:00Z"]},{"key":"thesis:degree_discipline","label":"Discipline","values":["Department of Geological and Mining Engineering and Sciences"]},{"key":"thesis:degree_level","label":"Degree Level","values":["Campus Access Master's Thesis"]},{"key":"thesis:degree_name","label":"Degree Name","values":["Master of Science in Mining Engineering (MS)"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["Generative Adversarial Network","Loss function","Geostatistics","Simulation","equiprobable realization","Artificial Intelligence and Robotics","Computational Engineering","Geological Engineering","Mining Engineering","Oil, Gas, and Energy","Programming Languages and Compilers"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier","label":"Identifier","values":["10.37099/mtu.dc.etdr/1812","https://digitalcommons.mtu.edu/etdr/1812"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description.abstract","label":"Abstract","values":["Geostatistical simulation plays a crucial role in modeling geological structures of the subsurface, providing equally plausible realizations for inference and practical applications. It serves as a powerful tool for estimating earth resources by accounting for the stochastic nature of geological systems. Among the various approaches, Multiple-Point Statistics (MPS) algorithms stand out for their ability to capture and reproduce complex spatial relationships and patterns in geological data that traditional geostatistical methods often fail to represent adequately. A fundamental aspect of MPS algorithms is the use of a training image (TI), which encapsulates the spatial distribution and properties of geological features. In mining, TIs represent geological features such as lithological maps, fault networks, or spatial variability in properties like metal grades and density. Similarly, in fields like petroleum engineering and hydrology, TIs model reservoir distributions, flow characteristics, and contaminant transport. The accuracy of these simulations depends heavily on the quality of the TI, but data limitations and selection biases often constrain their effectiveness. This thesis introduces a Multi-loss Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) framework as an innovative tool for geostatistical simulation. The proposed framework leverages the WGAN-GP architecture alongside multi-loss functions, including feature matching loss, pixel intensity loss, structural similarity index loss, and conditioning loss (for conditional simulations). These losses enhance the model's ability to generate realizations that faithfully reproduce the statistical and structural properties of the TI. To address limitations associated with limited TI data and selection bias, image augmentation techniques were integrated into the model, introducing diversity and uncertainty while expanding the training dataset. The WGAN-GP model was evaluated under both conditional and unconditional simulation scenarios. Results demonstrated its capability to preserve structural and spatial heterogeneity. Two key evaluation metrics were employed: (1) ensuring the generated simulations met a structural similarity index threshold, and (2) verifying that the simulations reproduced the spatial properties of the TI through variogram analysis. Statistically, the WGAN-GP framework, augmented with multi-loss functions, not only generated consistent geological realizations but also achieved computational efficiency—a significant advantage given the typically high computational cost of deep learning models. The results from the model demonstrate that the WGAN-GP framework, combined with multi-loss functions, not only generates consistent geological realizations but also offers a significant computational advantage, addressing the high computational demands of most deep learning models. This research emphasizes the importance of generative adversarial networks and their potential applications in geostatistical simulations."]},{"key":"dc:title","label":"Title","values":["A MULTI-LOSS WASSERSTEIN GAN FRAMEWORK FOR GEOSTATISTICAL SIMULATION"]}]}],"canonical_facts":{"dc:contributor":["Snehamoy Chatterjee"],"dc:contributor.committeemember":["Rouhollah (Radwin) Askari","Sidike Paheding"],"dc:creator":["Addai, Charles O."],"dc:date.available":["2025-12-31T08:00:00Z"],"dc:date.created":["2024-01-01T08:00:00Z"],"dc:description.abstract":["Geostatistical simulation plays a crucial role in modeling geological structures of the subsurface, providing equally plausible realizations for inference and practical applications. It serves as a powerful tool for estimating earth resources by accounting for the stochastic nature of geological systems. Among the various approaches, Multiple-Point Statistics (MPS) algorithms stand out for their ability to capture and reproduce complex spatial relationships and patterns in geological data that traditional geostatistical methods often fail to represent adequately. A fundamental aspect of MPS algorithms is the use of a training image (TI), which encapsulates the spatial distribution and properties of geological features. In mining, TIs represent geological features such as lithological maps, fault networks, or spatial variability in properties like metal grades and density. Similarly, in fields like petroleum engineering and hydrology, TIs model reservoir distributions, flow characteristics, and contaminant transport. The accuracy of these simulations depends heavily on the quality of the TI, but data limitations and selection biases often constrain their effectiveness. This thesis introduces a Multi-loss Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) framework as an innovative tool for geostatistical simulation. The proposed framework leverages the WGAN-GP architecture alongside multi-loss functions, including feature matching loss, pixel intensity loss, structural similarity index loss, and conditioning loss (for conditional simulations). These losses enhance the model's ability to generate realizations that faithfully reproduce the statistical and structural properties of the TI. To address limitations associated with limited TI data and selection bias, image augmentation techniques were integrated into the model, introducing diversity and uncertainty while expanding the training dataset. The WGAN-GP model was evaluated under both conditional and unconditional simulation scenarios. Results demonstrated its capability to preserve structural and spatial heterogeneity. Two key evaluation metrics were employed: (1) ensuring the generated simulations met a structural similarity index threshold, and (2) verifying that the simulations reproduced the spatial properties of the TI through variogram analysis. Statistically, the WGAN-GP framework, augmented with multi-loss functions, not only generated consistent geological realizations but also achieved computational efficiency—a significant advantage given the typically high computational cost of deep learning models. The results from the model demonstrate that the WGAN-GP framework, combined with multi-loss functions, not only generates consistent geological realizations but also offers a significant computational advantage, addressing the high computational demands of most deep learning models. This research emphasizes the importance of generative adversarial networks and their potential applications in geostatistical simulations."],"dc:identifier":["10.37099/mtu.dc.etdr/1812","https://digitalcommons.mtu.edu/etdr/1812"],"dc:subject":["Generative Adversarial Network","Loss function","Geostatistics","Simulation","equiprobable realization","Artificial Intelligence and Robotics","Computational Engineering","Geological Engineering","Mining Engineering","Oil, Gas, and Energy","Programming Languages and Compilers"],"dc:title":["A MULTI-LOSS WASSERSTEIN GAN FRAMEWORK FOR GEOSTATISTICAL SIMULATION"],"thesis:degree_discipline":["Department of Geological and Mining Engineering and Sciences"],"thesis:degree_level":["Campus Access Master's Thesis"],"thesis:degree_name":["Master of Science in Mining Engineering (MS)"]},"updated_at":"2026-11-19T08:19:12Z"}