256
D. Patel et al.
effective in avoiding over-fitting of the data to the model. Data splitting is another
validation method in which each ensemble dataset is generally split into calibration
and test subsets. Data splitting was shown to be an effective technique, where
a collection of new validation data is avoided. In this step, a model selection
is accomplished iteratively based on the optimization of error parameters. Error
metrics therefore play an important role in the model selection process. Popular
choices have included various combinations and variants of the mean of absolute
error (MAE), the standard deviation of error (SDE), the coefficient of correlation
(R), and the explained variance (R2) [10–12, 19].
Once a physics guided data-driven model is obtained, a new data point, not in
the calibration dataset, is tested/validated. If the errors are not satisfactory, it is
important to identify the step contributing to the unreliable model to allow suitable
modification for the next iteration. For instance, one might select a different learning
algorithm or select/identify new features via different data reduction techniques.
The modular nature of the framework allows exploration of a number of machine
learning models in a highly computationally efficient manner to best capture the
phenomena being studied.
3 Application of the Material Agnostic Framework
to Different Material Systems
In this section, the data-driven, process- structure, and structure-property linkages
applied to materials are reviewed. Microstructure plays an important role in the
formulation of P-S-P linkages and requires a higher dimensional representation
compared to other input/output (i.e., process parameters, properties) variables. As
microstructure quantification and representation form the bedrock to the material
agnostic models, the section is focused on various quantification techniques to P-S
and S-P linkages for diverse material systems.
3.1 Composites
Recent advancements in the development of composite materials systems sparked
the use of such material system in turbine engines replacing heavy metals parts.
Arguably, performance of composites, mainly ceramic matrix composites (CMCs),
is largely affected by its internal microstructure. At the scale of the microstructure,
the characteristic response of CMCs depends on the heterogeneous distribution of
local constituents (i.e., fiber, coating, crack, voids, and matrix itself) as seen in
Fig. 4. Thus, the damage response is sensitive to the local microstructure within
a CMC. The material agnostic, data-driven framework is ideal for exploring the
microstructural variabilities and their effects on performance.
D. Patel et al.
effective in avoiding over-fitting of the data to the model. Data splitting is another
validation method in which each ensemble dataset is generally split into calibration
and test subsets. Data splitting was shown to be an effective technique, where
a collection of new validation data is avoided. In this step, a model selection
is accomplished iteratively based on the optimization of error parameters. Error
metrics therefore play an important role in the model selection process. Popular
choices have included various combinations and variants of the mean of absolute
error (MAE), the standard deviation of error (SDE), the coefficient of correlation
(R), and the explained variance (R2) [10–12, 19].
Once a physics guided data-driven model is obtained, a new data point, not in
the calibration dataset, is tested/validated. If the errors are not satisfactory, it is
important to identify the step contributing to the unreliable model to allow suitable
modification for the next iteration. For instance, one might select a different learning
algorithm or select/identify new features via different data reduction techniques.
The modular nature of the framework allows exploration of a number of machine
learning models in a highly computationally efficient manner to best capture the
phenomena being studied.
3 Application of the Material Agnostic Framework
to Different Material Systems
In this section, the data-driven, process- structure, and structure-property linkages
applied to materials are reviewed. Microstructure plays an important role in the
formulation of P-S-P linkages and requires a higher dimensional representation
compared to other input/output (i.e., process parameters, properties) variables. As
microstructure quantification and representation form the bedrock to the material
agnostic models, the section is focused on various quantification techniques to P-S
and S-P linkages for diverse material systems.
3.1 Composites
Recent advancements in the development of composite materials systems sparked
the use of such material system in turbine engines replacing heavy metals parts.
Arguably, performance of composites, mainly ceramic matrix composites (CMCs),
is largely affected by its internal microstructure. At the scale of the microstructure,
the characteristic response of CMCs depends on the heterogeneous distribution of
local constituents (i.e., fiber, coating, crack, voids, and matrix itself) as seen in
Fig. 4. Thus, the damage response is sensitive to the local microstructure within
a CMC. The material agnostic, data-driven framework is ideal for exploring the
microstructural variabilities and their effects on performance.