{"id":{"repo_id":"uiuc","oai_identifier":"oai:www.ideals.illinois.edu:2142/124355"},"canonical_url":"https://search.dev.ndltd.org/etd/uiuc/oai:www.ideals.illinois.edu:2142/124355","repository":{"repo_id":"uiuc","name":"University of Illinois - Urbana-Champaign","base_url":"https://www.ideals.illinois.edu/oai-pmh"},"display":{"title":"Machine learning surrogates for multiphysics design","abstract":"Submission original under an indefinite embargo labeled 'Open Access'. The submission was exported from vireo on 2024-09-16 without embargo terms","abstract_html":"Submission original under an indefinite embargo labeled 'Open Access'. The submission was exported from vireo on 2024-09-16 without embargo terms","abstract_has_math":false,"creators":["Parrott, Corey M."],"institution":"University of Illinois at Urbana-Champaign","degree_name":"Ph.D.","degree_level":"Dissertation","degree_discipline":"Aerospace Engineering","degree_department":null,"school":null,"contributors":["James, Kai","Allison, James","Zhang, Xiaojia","Geubelle, Philippe"],"advisors":[],"committee_chairs":[],"committee_members":[],"year":2024,"date_issued":"2024-05","date_published":"2024-05","updated_at":"2026-07-22T22:25:00Z","subjects":["Optimization","Multidisciplinary Design Optimization","Machine Learning","Data-driven Design"],"languages":["en","eng"],"rights":["Copyright 2024 Corey Parrott"],"rights_urls":[],"identifier_entries":[]},"links":{"outbound_url":"https://hdl.handle.net/2142/124355","outbound_label":"Handle","outbound_source":"dc:identifier"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor","label":"Contributor","values":["James, Kai","Allison, James","Zhang, Xiaojia","Geubelle, Philippe"]},{"key":"dc:creator","label":"Author","values":["Parrott, Corey M."]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date","label":"Dc Date","values":["2024-05","2024-04-26"]},{"key":"dc:type","label":"Dc Type","values":["text"]},{"key":"thesis:degree_discipline","label":"Discipline","values":["Aerospace Engineering"]},{"key":"thesis:degree_level","label":"Degree Level","values":["Dissertation"]},{"key":"thesis:degree_name","label":"Degree Name","values":["Ph.D."]},{"key":"thesis:institution_name","label":"Thesis Institution Name","values":["University of Illinois at Urbana-Champaign"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["Optimization","Multidisciplinary Design Optimization","Machine Learning","Data-driven Design"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language","label":"Dc Language","values":["en","eng"]},{"key":"dc:rights","label":"Dc Rights","values":["Copyright 2024 Corey Parrott"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier","label":"Identifier","values":["https://hdl.handle.net/2142/124355"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description","label":"Description","values":["Submission original under an indefinite embargo labeled 'Open Access'. The submission was exported from vireo on 2024-09-16 without embargo terms","The student, Corey Parrott, accepted the attached license on 2024-04-21 at 20:31.","The student, Corey Parrott, submitted this Dissertation for approval on 2024-04-21 at 20:48.","This Dissertation was approved for publication on 2024-04-26 at 10:52.","DSpace SAF Submission Ingestion Package generated from Vireo submission #20525 on 2024-09-16 at 00:35:40","Gradient-based optimization techniques are powerful, providing capabilities for optimal design of any system represented with a differentiable physics model. However, the computational cost of these traditional gradient-based methods is amplified for multidisciplinary design optimization (MDO) problems, most notably when coupling between physics disciplines is accounted for. This approach is characteristically iterative. With the current state of the design variable driving physical response, physics simulations are completed, sensitivities computed, and design variables updated (with sometimes costly algorithms to satisfy optimization constraints) in a cyclical process until convergence criteria are met. To alleviate this, this work investigates new methods and applications of machine learning as a surrogate for MDO, allowing for design synthesis in real-time. Though training a machine learning model is still an optimization problem at its core, a trained model is able to synthesize outputs in real-time, only requiring a single forward pass through the network. Some of the key areas of machine learning advancements lie in computer vision, networking, and natural language processing (NLP). Though many of the architectures proposed for these tasks could be trained and applied as a `black box' tool to MDO, the structure of the data would often be unaccounted for. Researchers using machine learning surrogates are often challenged with models that are highly unreliable. Though many designs will appear similar to those produced through physics-based methods (from a qualitative perspective), there are vulnerabilities in producing outliers when analyzed quantitatively. Without the proper mechanisms in place to address the characteristics of problems in physics, the models are susceptible to producing low performing designs. This work is proposed to address management of data and algorithms when creating surrogate models for problems of this sort. This is presented for a number of case studies. A design enhancement model is proposed, filtering single physics designs into those supporting multiphysics boundary conditions. In this study, an emphasis is placed on the single physics domain used and its correlation to the multiphysics load path. A multi-head self-attention network is proposed to capture global dependancies of data. This is shown to improve connectivity and stability of synthesized designs. Additionally, a network addressing the combined problem of packaging, routing, and physics-based peformance is proposed. With such a large number of constraints, it is found that infusing the physics-based model into the network can become necessary for these cases. Furthermore, this work pursues improvements to methods in multimodal synthesis, focusing on design diversity. Through selective feature extraction with skeletonization algorithms, diversity is selectively imposed on structural features of a controlled size. This is shown to yield diversity in elements that more likely affect convergence to separate local minima, and in some cases, result in synthesis of designs that outperform those produced through physics-based models. The proposed frameworks present mechanisms to better capture the characteristics of the physics-based data, improving surrogate performance for such tasks. These case studies present novel contributions to the literature, with capabilities for a wide range of problems in the field of design optimization."]},{"key":"dc:format","label":"Dc Format","values":["application/pdf"]},{"key":"dc:title","label":"Title","values":["Machine learning surrogates for multiphysics design"]}]}],"canonical_facts":{"dc:contributor":["James, Kai","Allison, James","Zhang, Xiaojia","Geubelle, Philippe"],"dc:creator":["Parrott, Corey M."],"dc:date":["2024-05","2024-04-26"],"dc:description":["Submission original under an indefinite embargo labeled 'Open Access'. The submission was exported from vireo on 2024-09-16 without embargo terms","The student, Corey Parrott, accepted the attached license on 2024-04-21 at 20:31.","The student, Corey Parrott, submitted this Dissertation for approval on 2024-04-21 at 20:48.","This Dissertation was approved for publication on 2024-04-26 at 10:52.","DSpace SAF Submission Ingestion Package generated from Vireo submission #20525 on 2024-09-16 at 00:35:40","Gradient-based optimization techniques are powerful, providing capabilities for optimal design of any system represented with a differentiable physics model. However, the computational cost of these traditional gradient-based methods is amplified for multidisciplinary design optimization (MDO) problems, most notably when coupling between physics disciplines is accounted for. This approach is characteristically iterative. With the current state of the design variable driving physical response, physics simulations are completed, sensitivities computed, and design variables updated (with sometimes costly algorithms to satisfy optimization constraints) in a cyclical process until convergence criteria are met. To alleviate this, this work investigates new methods and applications of machine learning as a surrogate for MDO, allowing for design synthesis in real-time. Though training a machine learning model is still an optimization problem at its core, a trained model is able to synthesize outputs in real-time, only requiring a single forward pass through the network. Some of the key areas of machine learning advancements lie in computer vision, networking, and natural language processing (NLP). Though many of the architectures proposed for these tasks could be trained and applied as a `black box' tool to MDO, the structure of the data would often be unaccounted for. Researchers using machine learning surrogates are often challenged with models that are highly unreliable. Though many designs will appear similar to those produced through physics-based methods (from a qualitative perspective), there are vulnerabilities in producing outliers when analyzed quantitatively. Without the proper mechanisms in place to address the characteristics of problems in physics, the models are susceptible to producing low performing designs. This work is proposed to address management of data and algorithms when creating surrogate models for problems of this sort. This is presented for a number of case studies. A design enhancement model is proposed, filtering single physics designs into those supporting multiphysics boundary conditions. In this study, an emphasis is placed on the single physics domain used and its correlation to the multiphysics load path. A multi-head self-attention network is proposed to capture global dependancies of data. This is shown to improve connectivity and stability of synthesized designs. Additionally, a network addressing the combined problem of packaging, routing, and physics-based peformance is proposed. With such a large number of constraints, it is found that infusing the physics-based model into the network can become necessary for these cases. Furthermore, this work pursues improvements to methods in multimodal synthesis, focusing on design diversity. Through selective feature extraction with skeletonization algorithms, diversity is selectively imposed on structural features of a controlled size. This is shown to yield diversity in elements that more likely affect convergence to separate local minima, and in some cases, result in synthesis of designs that outperform those produced through physics-based models. The proposed frameworks present mechanisms to better capture the characteristics of the physics-based data, improving surrogate performance for such tasks. These case studies present novel contributions to the literature, with capabilities for a wide range of problems in the field of design optimization."],"dc:format":["application/pdf"],"dc:identifier":["https://hdl.handle.net/2142/124355"],"dc:language":["en","eng"],"dc:rights":["Copyright 2024 Corey Parrott"],"dc:subject":["Optimization","Multidisciplinary Design Optimization","Machine Learning","Data-driven Design"],"dc:title":["Machine learning surrogates for multiphysics design"],"dc:type":["text"],"thesis:degree_discipline":["Aerospace Engineering"],"thesis:degree_level":["Dissertation"],"thesis:degree_name":["Ph.D."],"thesis:institution_name":["University of Illinois at Urbana-Champaign"]},"updated_at":"2026-07-22T22:25:00Z"}