京都大学 / 新制・課程博士 / 博士(医学) / 甲第23778号 / 医博第4824号 / 新制||医||1057(附属図書館) / 京都大学大学院医学研究科医学専攻 / (主査)教授 黒田 知宏, 教授 溝脇 尚志, 教授 波多野 悦朗 / 学位規則第4条第1項該当 / Doctor of Medical Science / Kyoto University / DFAM

京都大学 / 新制・課程博士 / 博士(情報学) / 甲第25427号 / 情博第865号 / 新制||情||145(附属図書館) / 京都大学大学院情報学研究科社会情報学専攻 / (主査)教授 伊藤 孝行, 教授 黒田 知宏, 教授 岡部 寿男, 吉川 正俊(京都大学 名誉教授) / 学位規則第4条第1項該当 / Doctor of Informatics / Kyoto University / DFAM

<p dir="ltr">There is a growing need to reconstruct high quality tomographic images from sparse view measurements to accommodate time and space constraints as well as patient well-being in medical CT. Analytical methods perform poorly with sub-Nyquist acquisition rates. In extreme cases with 4 or fewer views, effective reconstruction approaches must be able to incorporate side information to constrain the solution space of an otherwise under-determined problem. This thesis presents two sparse view tomography problems that are solved using techniques that exploit. knowledge of the structural and physical properties of the scanned objects.</p><p dir="ltr"><br></p><p dir="ltr">First, we reconstruct four view CT datasets obtained from an in-situ imaging system used to observe Kolsky bar impact experiments. Test subjects are typically 3D-printed out ofhomogeneous materials into shapes with circular cross sections. Two advanced prior modelsare formulated to incorporate these assumptions in a modular fashion into the iterativeradiographic inversion framework. The first is a Multi-Slice Fusion and the latter is TotalVariation regularization that operates in cylindrical coordinates.</p><p dir="ltr"><br></p><p dir="ltr">In the second problem, artificial neural networks (NN) are used to directly invert a temporal sequence of four radiographic images of discontinuities propagating through an imploding steel shell. The NN is fed the radiographic features that are robust to scatter and is trained using density simulations synthesized as solutions to hydrodynamic equations of state. The proposed reconstruction pipeline learns and enforces physics-based assumptions of hydrodynamics and shock physics to constrain the final reconstruction to a space ofphysically admissible solutions.</p>

<p dir="ltr">For additive manufacturing to be a viable method to build metal parts for industries such as nuclear, the manufactured parts must be of higher quality and have lower variation in said quality than what can be achieved today. This high variation in quality bars the techniques from being used in high safety tolerance fields, such as nuclear. If this obstacle could be overcome, the benefits of additive manufacturing would be in lower cost for complex parts, as well as the ability to design and test parts in a very short timeframe, as only the CAD model needs to be created to manufacture the part. In this study, work to achieve this lower variation of quality was approached in two ways. The first was in the development of surrogate models, utilizing machine learning, to predict the end quality of additively manufactured parts. This was done by using experimental data for the mechanical properties of built parts as outputs to be predicted, and in-situ signals captured during the manufacturing process as inputs to the model. To capture the in-situ signals, cameras were used for thermal and optical imaging, leveraging the natural layer-by-layer manufacturing method used in AM techniques. The final models were created using support vector machine and gaussian process regression machine learning algorithms, giving high correlations between the insitu signals and mechanical properties of relative density, elongation to fracture, uniform elongation, and the work hardening exponent. The second approach to this study was in the development of a reduced order model for a computer simulation of an AM build. For project, a ROM was built inside the MOOSE framework, and was developed for an AM model designed by the MOOSE team, using proper orthogonal decomposition to project the problem onto a lower dimensional subspace, using POD to design the reduced basis subspace. The ROM was able to achieve a reduction to 1% the original dimensionality of the problem, while only allowing 2-5% relative error associated with the projection.</p>