{"id":{"repo_id":"uiuc","oai_identifier":"oai:www.ideals.illinois.edu:2142/125551"},"canonical_url":"https://search.dev.ndltd.org/etd/uiuc/oai:www.ideals.illinois.edu:2142/125551","repository":{"repo_id":"uiuc","name":"University of Illinois - Urbana-Champaign","base_url":"https://www.ideals.illinois.edu/oai-pmh"},"display":{"title":"Towards data-driven inverse design for materials and structures","abstract":"Submission original under an indefinite embargo labeled 'Open Access'. The submission was exported from vireo on 2025-02-04 without embargo terms","abstract_html":"Submission original under an indefinite embargo labeled 'Open Access'. The submission was exported from vireo on 2025-02-04 without embargo terms","abstract_has_math":false,"creators":["Jasperson, Ben A."],"institution":"University of Illinois at Urbana-Champaign","degree_name":"Ph.D.","degree_level":"Dissertation","degree_discipline":"Theoretical & Applied Mechans","degree_department":null,"school":null,"contributors":["Johnson, Harley T","Ertekin, Elif","Admal, Nikhil","Zhang, Xiaojia Shelly","Shao, Chenhui"],"advisors":[],"committee_chairs":[],"committee_members":[],"year":2024,"date_issued":"2024-08","date_published":"2024-08","updated_at":"2026-07-22T22:25:02Z","subjects":["Interatomic Potential","Molecular Simulation","Machine Learning","Neural Networks","Topology Optimization","Inverse Design","Electromagnetics","Computational Methods","Atomistic Models","Mechanical Properties","Grain Boundary Energy"],"languages":["en","eng"],"rights":["Copyright 2024 Benjamin A. Jasperson"],"rights_urls":[],"identifier_entries":[]},"links":{"outbound_url":"https://hdl.handle.net/2142/125551","outbound_label":"Handle","outbound_source":"dc:identifier"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor","label":"Contributor","values":["Johnson, Harley T","Ertekin, Elif","Admal, Nikhil","Zhang, Xiaojia Shelly","Shao, Chenhui"]},{"key":"dc:creator","label":"Author","values":["Jasperson, Ben A."]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date","label":"Dc Date","values":["2024-08","2024-07-01"]},{"key":"dc:type","label":"Dc Type","values":["text","Thesis"]},{"key":"thesis:degree_discipline","label":"Discipline","values":["Theoretical & Applied Mechans"]},{"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":["Interatomic Potential","Molecular Simulation","Machine Learning","Neural Networks","Topology Optimization","Inverse Design","Electromagnetics","Computational Methods","Atomistic Models","Mechanical Properties","Grain Boundary Energy"]}]},{"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 Benjamin A. Jasperson"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier","label":"Identifier","values":["https://hdl.handle.net/2142/125551"]}]},{"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 2025-02-04 without embargo terms","The student, Ben Jasperson, accepted the attached license on 2024-07-01 at 09:10.","The student, Ben Jasperson, submitted this Dissertation for approval on 2024-07-01 at 09:22.","This Dissertation was approved for publication on 2024-07-01 at 11:26.","DSpace SAF Submission Ingestion Package generated from Vireo submission #20904 on 2025-02-04 at 21:04:01","Optimization problems are ubiquitous in engineering, with applications ranging from parameter fitting to inverse design. Iterative methods are common, relying on the evaluation of previous steps to predict the next one. This evaluation process typically requires function or solver calls that are often the most computationally demanding steps in the method. These computational costs, coupled with the complexities of real-world applications, can quickly make these problems intractable. As a result, there has been an interest in using trained surrogate models to replace the solver calls. In addition to the potential computational cost savings, the use of surrogate models allows for the exploration of novel optimization algorithms. We present a data-driven, neural network (NN)-based optimization architecture that allows us to solve challenging inverse design problems of interest. The method uses a pre-trained surrogate model, or forward model (FM), to predict device performance. It is trained with readily available data, taking as input the device design or model parameters and returning the necessary performance metrics. The type of FM used is problem specific and can take many forms, including multiple linear regression, support vector regression, and NN-based approaches. The FM can be trained using previously generated data if available, minimizing the required computational resources. The FM is coupled with an optimization neural network (ONN) that takes as input a fixed descriptor and, when trained, produces an optimized set of design parameters to minimize loss when compared to target metrics of interest. This work applies the proposed NN-based algorithm to two unique applications that challenge traditional optimization approaches. First, topology optimization of an optical device fabricated using a phase change material, vanadium dioxide, is performed. The use of a surrogate model allows us to solve a challenging multi-physics problem with single-physics, multi-domain simulations. The training data is generated using the finite-element method in COMSOL. An optimized design front (Pareto curve) is efficiently generated for a range of target values, highlighting the cost effectiveness of the approach. We then shift to optimizing interatomic potentials (IPs) for atomistic simulations that predict properties of interest. To this end, we explore developing surrogate models to predict separately either grain boundary (GB) energy or plastic flow strength. In this instance, the FM replaces the atomistic simulations that are necessary to predict GB energy in symmetric-tilt grain boundaries and plastic flow strength. Inputs to the surrogate model are canonical properties, which we define as fundamental properties that can be calculated using low-cost, few-atom simulations (e.g. lattice constants and elastic constants). We leverage observed correlations between canonical properties and predicted GB energy analytical model coefficients or strength to create the necessary FM, thereby substituting for many-atom atomistic simulations. We expect that these surrogate models can be used as a part of an overall cost function during IP parameter fitting to better predict the properties of interest. As a final step towards connecting the work, we present initial results in developing an FM that predicts canonical properties based on IP parameter values. We extract model parameters from OpenKIM and identify correlations between them and the canonical properties. We then create the surrogate model and couple it with the optimization algorithm to determine the desired IP parameters. This information can be leveraged, along with DFT training data and additional atomistic simulations, to further optimize the parameters. This surrogate model could be coupled with either the GB energy or strength surrogate models to improve IM parameters, with the goal of improving IP prediction accuracy."]},{"key":"dc:format","label":"Dc Format","values":["application/pdf"]},{"key":"dc:title","label":"Title","values":["Towards data-driven inverse design for materials and structures"]}]}],"canonical_facts":{"dc:contributor":["Johnson, Harley T","Ertekin, Elif","Admal, Nikhil","Zhang, Xiaojia Shelly","Shao, Chenhui"],"dc:creator":["Jasperson, Ben A."],"dc:date":["2024-08","2024-07-01"],"dc:description":["Submission original under an indefinite embargo labeled 'Open Access'. The submission was exported from vireo on 2025-02-04 without embargo terms","The student, Ben Jasperson, accepted the attached license on 2024-07-01 at 09:10.","The student, Ben Jasperson, submitted this Dissertation for approval on 2024-07-01 at 09:22.","This Dissertation was approved for publication on 2024-07-01 at 11:26.","DSpace SAF Submission Ingestion Package generated from Vireo submission #20904 on 2025-02-04 at 21:04:01","Optimization problems are ubiquitous in engineering, with applications ranging from parameter fitting to inverse design. Iterative methods are common, relying on the evaluation of previous steps to predict the next one. This evaluation process typically requires function or solver calls that are often the most computationally demanding steps in the method. These computational costs, coupled with the complexities of real-world applications, can quickly make these problems intractable. As a result, there has been an interest in using trained surrogate models to replace the solver calls. In addition to the potential computational cost savings, the use of surrogate models allows for the exploration of novel optimization algorithms. We present a data-driven, neural network (NN)-based optimization architecture that allows us to solve challenging inverse design problems of interest. The method uses a pre-trained surrogate model, or forward model (FM), to predict device performance. It is trained with readily available data, taking as input the device design or model parameters and returning the necessary performance metrics. The type of FM used is problem specific and can take many forms, including multiple linear regression, support vector regression, and NN-based approaches. The FM can be trained using previously generated data if available, minimizing the required computational resources. The FM is coupled with an optimization neural network (ONN) that takes as input a fixed descriptor and, when trained, produces an optimized set of design parameters to minimize loss when compared to target metrics of interest. This work applies the proposed NN-based algorithm to two unique applications that challenge traditional optimization approaches. First, topology optimization of an optical device fabricated using a phase change material, vanadium dioxide, is performed. The use of a surrogate model allows us to solve a challenging multi-physics problem with single-physics, multi-domain simulations. The training data is generated using the finite-element method in COMSOL. An optimized design front (Pareto curve) is efficiently generated for a range of target values, highlighting the cost effectiveness of the approach. We then shift to optimizing interatomic potentials (IPs) for atomistic simulations that predict properties of interest. To this end, we explore developing surrogate models to predict separately either grain boundary (GB) energy or plastic flow strength. In this instance, the FM replaces the atomistic simulations that are necessary to predict GB energy in symmetric-tilt grain boundaries and plastic flow strength. Inputs to the surrogate model are canonical properties, which we define as fundamental properties that can be calculated using low-cost, few-atom simulations (e.g. lattice constants and elastic constants). We leverage observed correlations between canonical properties and predicted GB energy analytical model coefficients or strength to create the necessary FM, thereby substituting for many-atom atomistic simulations. We expect that these surrogate models can be used as a part of an overall cost function during IP parameter fitting to better predict the properties of interest. As a final step towards connecting the work, we present initial results in developing an FM that predicts canonical properties based on IP parameter values. We extract model parameters from OpenKIM and identify correlations between them and the canonical properties. We then create the surrogate model and couple it with the optimization algorithm to determine the desired IP parameters. This information can be leveraged, along with DFT training data and additional atomistic simulations, to further optimize the parameters. This surrogate model could be coupled with either the GB energy or strength surrogate models to improve IM parameters, with the goal of improving IP prediction accuracy."],"dc:format":["application/pdf"],"dc:identifier":["https://hdl.handle.net/2142/125551"],"dc:language":["en","eng"],"dc:rights":["Copyright 2024 Benjamin A. Jasperson"],"dc:subject":["Interatomic Potential","Molecular Simulation","Machine Learning","Neural Networks","Topology Optimization","Inverse Design","Electromagnetics","Computational Methods","Atomistic Models","Mechanical Properties","Grain Boundary Energy"],"dc:title":["Towards data-driven inverse design for materials and structures"],"dc:type":["text","Thesis"],"thesis:degree_discipline":["Theoretical & Applied Mechans"],"thesis:degree_level":["Dissertation"],"thesis:degree_name":["Ph.D."],"thesis:institution_name":["University of Illinois at Urbana-Champaign"]},"updated_at":"2026-07-22T22:25:02Z"}