b'Develop High FidelityNew machine learning models can predict the effective thermal Computation Modelsconductivity of closed-cell porous media, crack-containing microstructures, and additively manufactured materials.to Calculate the Effective Material PropertiesT he team developed machine learning metamodels that can predict the effective thermal conductivity of closed-cell porous media and crack-of Porous Cells containing microstructures, such as tristructural isotropic compacts, and additively manufactured materials. This was accomplished by exploring the design space and perturbing all related features to identify the effect of pore size, crack size, orientation, geometric configuration, as well as tristructural isotropic particle distribution on effective thermal conductivity.To achieve this goal, physics-based modeling through finite element analysis was PROJECT NUMBER:performed, followed by a systematic machine learning pipeline covering data sampling, 20A44-177 preprocessing, cleaning, sensitivity studies, surrogate modeling, then model validation. The model evolved from encompassing a single cell with a single spherical pore in TOTAL APPROVED AMOUNT:a two-dimensional setup to multiple pores organized in several configurations to $1,448,000 over 3 years elliptical pores, to facilitate the control of the aspect ratios, and hence model cracks PRINCIPAL INVESTIGATOR:as well. Next, tristructural isotropic particles were introduced. Then the pores and Mohammad Abdo tristructural isotropic particles were randomized following distributions fitted from the x-ray computed tomography scans of the available samples. CO-INVESTIGATORS:Boopathy Kombaiah, INL Several metamodels were constructed including linear regressors, support Isabella Von Rooyen, INL vector regressors, random forests, polynomial regressors, and feed-forward Yu-Lin Shin, University of New Mexico neural networks. These metamodels were validated and showed a coefficient of determination score that exceeded 0.95, reflecting the goodness of fit. However, COLLABORATORS: these models still require the user to perform some data preprocessing to compute Center of Advanced Energy Studies features of the geometric distribution. Hence, to render this more efficient, an The Ohio State Universityensemble model was built to concatenate the numeric data reflecting local conductivities, porosity, and particle packing fractions with a convolutional neural network to capture the geometric distribution, shapes, and orientations of pores and tristructural isotropic particles. This final model was in turn validated.Finally, the model was converted to a three-dimensional finite element model. At each evolution, metamodels were built to assess the accuracy and compare to the literature if available. The research resulted in two journal articles for the sensitivity studies, one for the experimental analysis, and one for the machine learning work. The general finding was that the influential attributes were the local conductivities of the composite material and tristructural isotropic particles, the pore and crack densities, and the particle packing factor. Minor effects were observed and attributed to orientations and configurations. Insensitivity to orientation suggested the isotropic nature of the heat flux traveling through each medium. 30'