Abstract:

Structure-property linkages are commonly used to estimate various properties of materials at various regimes or length scales. These linkages help us to optimize the various process and quality control variables during product development. For many examples like Teflon, Velcro, Lithium-Ion batteries, and several others, the time frame of deploying new materials into successful products is certainly large, as much as 10-15 years. Acceleration in the design and deployment of advanced materials for various business verticals has emphasized the development of reliable and effective structure-property linkages. The conventional method of rigorous experimentation and validation is resourceful and cost-intensive. A combination of reduced-order models and certain approximations under ICME has decreased the time frame significantly but incur heavy computational costs. The by-products of both the above methods result in the generation of large amounts of data. Materials informatics is a field that applies the concepts of information engineering like machine learning, statistics, control theory, etc., to materials science and engineering. The concepts from the above field can be applied to the microstructure data to predict corresponding mechanical properties and subsequently develop effective structure-property correlations.

In our work, we propose to develop a complete data-enabled predictive framework to estimate micro(structure) mechanical(property) correlations. Out of the many challenges in the field, in the present study, we focus on quantification and representation of microstructure data. Two-point correlations are employed to represent the generated 2D and 3D microstructure data in terms of a probability-based quantification. The influence of captured microstructural features via two-point correlations is studied. Although a probability-based spatial quantification is achieved, the dimensionality of the data is still huge making it computationally expensive during processing. Intrinsic dimensionality of the data is a reduced dimension where maximum information in the data is retained. An intrinsic dimensionality estimation framework is developed to achieve a reduced dimension of microstructure data. The framework is independent of the dimensionality reduction method used as it uses general global and local information preservation estimates.​

----