250
D. Patel et al.
where data science may be best applicable. The current state-of-the-art materials
engineering practice relies on reliable synthesis and/or manufacturing process with
quality control checks and extensive laboratory data on performance characteristics
as a metric to guide design processes. The material structure is mostly characterized
via nondestructive tests, rather than traditional microscopy. X-ray CT, eddy current
test, and ultrasound imaging are some of these techniques. On the other hand,
most materials models require detailed microstructural information at all length
scales using optical, SEM, and TEM imaging techniques. When required, the
industry employs such destructive methods on coupons from each batch of material
manufactured. In some cases, processing-microstructure-property relationships are
semi-empirically derived (physics-based but fast acting) and used in engineering
design to achieve desired properties. This is especially true for applications that
can benefit significantly from ICME. It is those applications that are best suited for
the use of data science or machine learning (ML) [2]. In this work we focus on such
applications, and we show that the ML approach has potential as a structure-property
linkage model and is materials agnostic. It must be noted that the ML framework can incorporate the physics behind the linkages, despite staying materials
agnostic.
We review the data-driven, material agnostic framework towards possible use
in ICME-based applications. We illustrate the application of the framework across
different materials, without sacrificing the physics behind the linkages. In particular,
the data-driven methods are shown to successfully establish the process-structure or
structure-property (P-S-P) linkages applied to a wide range of materials, including
metallic alloys [3], ceramics [4], composites [5], and polymeric materials [6].
2 Material Agnostic Data-Driven Framework
to Process-Structure-Property Linkages
A key element to data science in materials science is a inherent versatile framework
which is amenable to data acquisition, curation, dissemination, and reuse of highvalue knowledge in a highly efficient compact manner at a desirable cost at the
relevant length scale. The reason the approach is material agnostic is that the
treatment is entirely mathematical in how data is represented and relationships
are established. The physics are captured in the preprocessing stage. With respect
to microstructure, preprocessing involves representation, data reduction, feature
extraction, and/or simulations. With respect to properties, this involves use of
physics-based simulations to generate sufficient data for ML learning. The same
is true for processing-structure relations. While the physics-based simulations carry
the computationally expensive physics, the ML tool derived from them is fast acting
and amenable for inverse solutions.
The approach can be easily coupled with multiscale materials’ modeling effort
where the data-driven (P-S-P) models can be integrated to reduce the computational
cost associated with the physics-based model while improving the efficacy of the
D. Patel et al.
where data science may be best applicable. The current state-of-the-art materials
engineering practice relies on reliable synthesis and/or manufacturing process with
quality control checks and extensive laboratory data on performance characteristics
as a metric to guide design processes. The material structure is mostly characterized
via nondestructive tests, rather than traditional microscopy. X-ray CT, eddy current
test, and ultrasound imaging are some of these techniques. On the other hand,
most materials models require detailed microstructural information at all length
scales using optical, SEM, and TEM imaging techniques. When required, the
industry employs such destructive methods on coupons from each batch of material
manufactured. In some cases, processing-microstructure-property relationships are
semi-empirically derived (physics-based but fast acting) and used in engineering
design to achieve desired properties. This is especially true for applications that
can benefit significantly from ICME. It is those applications that are best suited for
the use of data science or machine learning (ML) [2]. In this work we focus on such
applications, and we show that the ML approach has potential as a structure-property
linkage model and is materials agnostic. It must be noted that the ML framework can incorporate the physics behind the linkages, despite staying materials
agnostic.
We review the data-driven, material agnostic framework towards possible use
in ICME-based applications. We illustrate the application of the framework across
different materials, without sacrificing the physics behind the linkages. In particular,
the data-driven methods are shown to successfully establish the process-structure or
structure-property (P-S-P) linkages applied to a wide range of materials, including
metallic alloys [3], ceramics [4], composites [5], and polymeric materials [6].
2 Material Agnostic Data-Driven Framework
to Process-Structure-Property Linkages
A key element to data science in materials science is a inherent versatile framework
which is amenable to data acquisition, curation, dissemination, and reuse of highvalue knowledge in a highly efficient compact manner at a desirable cost at the
relevant length scale. The reason the approach is material agnostic is that the
treatment is entirely mathematical in how data is represented and relationships
are established. The physics are captured in the preprocessing stage. With respect
to microstructure, preprocessing involves representation, data reduction, feature
extraction, and/or simulations. With respect to properties, this involves use of
physics-based simulations to generate sufficient data for ML learning. The same
is true for processing-structure relations. While the physics-based simulations carry
the computationally expensive physics, the ML tool derived from them is fast acting
and amenable for inverse solutions.
The approach can be easily coupled with multiscale materials’ modeling effort
where the data-driven (P-S-P) models can be integrated to reduce the computational
cost associated with the physics-based model while improving the efficacy of the