Microstructure reconstruction and structural equation modeling for computational design of nanodielectrics
 2.2k Downloads
 9 Citations
Abstract
Nanodielectric materials, consisting of nanoparticlefilled polymers, have the potential to become the dielectrics of the future. Although computational design approaches have been proposed for optimizing microstructure, they need to be tailored to suit the special features of nanodielectrics such as low volume fraction, local aggregation, and irregularly shaped large clusters. Furthermore, key independent structural features need to be identified as design variables. To represent the microstructure in a physically meaningful way, we implement a descriptorbased characterization and reconstruction algorithm and propose a new decomposition and reassembly strategy to improve the reconstruction accuracy for microstructures with low volume fraction and uneven distribution of aggregates. In addition, a touching cell splitting algorithm is employed to handle irregularly shaped clusters. To identify key nanodielectric material design variables, we propose a Structural Equation Modeling approach to identify significant microstructure descriptors with the least dependency. The method addresses descriptor redundancy in the existing approach and provides insight into the underlying latent factors for categorizing microstructure. Four descriptors, i.e., volume fraction, cluster size, nearest neighbor distance, and cluster roundness, are identified as important based on the microstructure correlation functions (CF) derived from images. The sufficiency of these four key descriptors is validated through confirmation of the reconstructed images and simulated material properties of the epoxynanosilica system. Among the four key descriptors, volume fraction and cluster size are dominant in determining the dielectric constant and dielectric loss.
Keywords
Nanodielectric Material design Descriptor identification Microstructure characterization and reconstruction Structural Equation ModelingBackground
Dielectric materials are widely used in mobile electronics, electrical transmission, and pulsed power applications [1]. There is an increasing demand for new nanodielectric materials, consisting of nanoparticlefilled polymers, for creating future electrical transmission and storage devices. One example is a new capacitor made from nanodielectrics that can store a large amount of energy and discharge it quickly with highenergy density [2]. The design of nanodielectrics is often multiobjective, for example, a tradeoff between dielectric constant and breakdown strength of dielectric materials has been observed [3]. It has been noted that small volume fractions of nanofillers can significantly improve the composites’ dielectric breakdown strength because of their high surface area/internal volume ratio [4]. To achieve design requirements under different application scenarios, a systematic computational design approach is needed to quickly explore the microstructure design space of nanodielectrics. In this work, we are developing characterization, reconstruction, and key microstructure feature identification techniques to support the computational design of nanodielectric systems.
A traditional material design follows a trialanderror process with the focus on exploring the relationships between processing conditions and material properties. This empirical approach to material design is expensive and time consuming. In integrated computational materials engineering (ICME), a threelink (i.e., processingstructureproperty) chain model that enables “microstructuremediated design” has been proposed to facilitate the design of new materials [5, 6]. The microstructure material design problem can be formulated as an optimization problem, in which the desired material properties drive the design of microstructure first and then the corresponding processing conditions. As pointed out by Xu et al, ICME faces three designrelated challenges: design representation, design evaluation, and design optimization [7]. Design representation requires quantitative representation of the design space of heterogeneous microstructures using a small set of design variables. Design evaluation is the process of assessing material properties for a given microstructure morphology, which often involves finite element modeling (FEM) and simulations. Design optimization searches for the optimal microstructure design to achieve the desired material properties. Using the design of nanodielectrics as a focal application, the main focus of this paper is on developing methods to support design representation and identify key microstructural design variables in material design.
A good design representation means an accurate quantitative description of microstructures that is easy to control from the perspective of simulation, design, and processing. In the existing work, methods have been developed to characterize and reconstruct microstructures for different material systems. They can be mainly classified into two categories: one is to use correlation functions (CFs) such as 2point CF, 2point cluster CF, and surface correlation [8, 9, 10, 11]; the other is to use physical descriptors, such as volume fraction, particle size, and minimum distance between particles [12]. CFbased reconstruction often involves optimization procedures to minimize the error between the actual CF and the target ones. This approach has been extended for reconstructing multiphase microstructures, for which each phase has its own CF [13, 14]. Although CFbased approaches are flexible and can be adapted for different microstructures, it is computationally expensive and prohibitive for use as a part of the iterative material design procedure. In addition, CFs are infinitely dimensional. While coefficients of the functions can be treated as design variables, they lack physical meaning. The descriptorbased approach, on the other hand, is much more intuitive and offers low dimensionality of design variables with clear physical meaning. Toward this end, Xu et. al. [12, 15] proposed a descriptorbased approach to fully characterize particlebased microstructures by introducing three categories of descriptors: (1) composition: e.g., volume fraction; (2) dispersion: e.g., nearest center distance, interphase area, cluster number, local volume fraction, and orientation; (3) geometry: e.g., cluster area, equivalent radius, aspect ratio, eccentricity, roundness, compactness, tortuosity, pore size, and rectangularity.

Low volume fraction and small number of clusters

Uneven distribution of aggregates (heterogeneity)

Irregularly shaped large clusters that cannot be modeled using simple geometries like a sphere or ellipse (see Fig. 1c)
The volume fraction of the nanodielectric fillers ranges from 0.5 to 3 % over the samples available in our study (collected from several dielectric systems with similar polymer dielectric permittivity). When the filler phase is on the nanoscale, small filler loadings can result in significant property improvement because of the large interfacial area. When the volume fraction is high, aggregation is harder to control and the property enhancements are reduced. As a consequence, the distribution is heterogeneous after processing, which is the reason why, as shown in Fig. 1c, local aggregates (marked by circles) can be observed in these microstructures. In addition, the dispersion (ability to separate primary particles) depends on the particle/particle and particle/polymer attraction [16]. The greater the particle/polymer enthalpic incompatibility, the greater the driving force for agglomeration.
While the descriptorbased method is generally applicable for particlebased nanodielectrics, the original reconstruction algorithm requires the microstructures to be simple and the distribution of filler phase to be even, which is not always satisfied in low volume fraction nanodielectric systems. Therefore, the existing descriptorbased method needs to be tailored to suit the special features of nanodielectrics.
Material informatics [17, 18] is a growing area that exploits information technology and data science to represent, manage, and analyze material data for accelerating new material discovery and design. One of the common challenges associated with material informatics is the high dimensionality of the data and the design space. Recently, efforts have been made in microstructure dimensionality reduction via manifold learning [19] and principal components [20]. However, dimension reduction based only on microstructures does not reflect the influence of microstructure on the properties of interest. To address such limitations, our recent work applied a supervised learning algorithm by using structural information from images and material properties from simulations [21] as supervisory (response) signals. Even though the method can determine the relative importance of descriptors, it introduces subjectivity when determining the final set of design variables. In addition, this learning method is not capable of discriminating redundant features (descriptors); neither it is reliable for cases with a small number of sample images.
In this work, we employ the descriptorbased characterization and reconstruction method to the particular nanodielectric system of interest. To achieve more realistic characterization and more accurate reconstruction, the existing algorithms are modified considering the aforementioned special features of the nanodielectric system. To address the issue of irregularly shaped large clusters, a touching cell splitting algorithm [22] is incorporated to reproduce more realistic structures. To capture the unevenly distributed aggregates, we propose a decomposition and reassembly strategy for reconstruction that preserves local microstructural information. To overcome the limitations of the existing machine learning approach, a Structural Equation Modeling [23] approach is proposed in this work to choose the proper set of independent descriptors using the information learned from images. By introducing latent layers in mapping input and output relations, we are able to identify the relationships and dependencies among descriptors, which allow determination of a small set of key descriptors as design variables. Finally, we illustrate the obtained relationship between key microstructural descriptors and the dielectric properties to support design of a nanoscale silica/epoxy matrix system.
Methods
Descriptorbased characterization and reconstruction
With various microscopic imaging techniques, such as scanning electron microscopy [24] and transmission electron microscopy (TEM) [25], material microstructures can be represented by grayscale digital images. Descriptor characterization is a process for extracting statistical information about the structure descriptors from the images. Due to the heterogeneity of these microstructures, statistical moments are often used as a part of the descriptors. For instance, “cluster area” as a descriptor is best described by a statistical distribution rather than a single value. To reduce dimensionality in representation, the first several orders of statistical moments, such as mean and variance are often used rather than considering the whole distribution.
Once a microstructure is characterized by descriptors, a typical 2D or 3D reconstruction follows a sequential procedure [12]: (1) dispersion reconstruction: center positions of clusters are adjusted to match dispersion descriptors, e.g., the nearest center distances, using optimization algorithms such as simulated annealing (SA), (2) geometry reconstruction: the geometry is randomly generated for each cluster based on the geometry descriptors and geometry profiles are assigned to each cluster, (3) composition adjustment: the edge of clusters are modified to satisfy the composition descriptors such as the volume fraction.
Because the primary particle size is small in this system, noise in the grayscale TEM images is more likely to be recognized as particles in image processing. In this work, we first employ Gaussian filtering [26] to remove the influence of noise. Gaussian filtering utilizes a Gaussian kernel to smooth the image, which will remove isolated pixels. After the filtering process, the existing descriptorbased characterization and reconstruction algorithm as described earlier is employed.
Touching cell splitting method
 1)
Polygon approximation of cluster edges
 2)
Identification of concave points and segmentation of cluster edges
 3)
Separation with ellipse fitting
 1)
concavity(P_{c}) ∈ (a_{1}, a_{2}), and
 2)
Line \( \frac{}{P_{\mathrm{pre}}{P}_{\mathrm{next}}} \) should not cross over the inner region of the aggregates.
Reconstruction using decomposition and reassembly
 1)
Divide the original microstructure into multiple equal size subblocks
 2)
Apply the 2D descriptor characterization and reconstruction algorithm to each subblock
 3)
Randomly assemble the subblock reconstructions to obtain the fully reconstructed microstructure
Our proposed method is inspired by the Morisita Index approach [27], which is used to analyze local versus global dispersions. The Morisita Index divides the original image using different sizes of small blocks, and then based on the number of particles in each block, a weighted index is calculated to represent the dispersion status of an image. Here we need to choose an appropriate block size based on a specific problem, but the basic idea is similar: let the subblocks keep the local information, which would be lost if characterization and reconstruction were directly applied to the global region. The size of the subblocks may influence the reconstruction accuracy. In this study, it is found that when the block size is slightly larger than the largest clusters in the microstructure, satisfactory results can be obtained. A decomposition and reassembly strategy was also employed in the evaluation of structureproperty relationships and proved to be effective [28]. One additional advantage of the proposed method is that in some subblocks, there may be no particles. The block is then maintained as pure matrix (void space) in the reconstruction to ease computation as well as to capture the void space feature which has been used in literature to quantitatively characterize material microstructure dispersion [29, 30] in low volume fraction systems.
Identification of key microstructure descriptors
Each of these equations is a regression equation; factor analysis seeks to find the coefficients λ_{ij} (loadings on factors) that best reproduce the observed variables from the factors. If all coefficients are correlations and factors are uncorrelated, then the sum of the squared loadings for variable X_{i}, e.g., \( {\displaystyle \sum_j}{\lambda}_{ij}^2 \), shows the proportion of the variance of X_{i} explained by these factors. This is called the communality, and the larger the communality for each variable, the more successful the factor solution is.
 Step 1:
Determine the number of latent factors (exogenous variables) n
 Step 2:
Conduct factor analysis with the n factors using proper rotation
 Step 3:
Identify exogenous variables that are poor indicators of latent factors (uniqueness > 0.5) [32]
Equation (5) shows the mathematical relationship of the structural model, the relationship between latent variables, while Eqs. (6) and (7) both represent the measurement model, illustrating the relationship between latent variables and corresponding indicators.
In the context of material structureproperty analysis, different categories of microstructure descriptors can be either predefined based on experience [12] or identified as the exogenous latent variables ξ_{i} for microstructure descriptors in the proposed Structural Equation Modeling structure after going through the EFA process. Latent variables are often not directly measured. For instance, dispersion can be a latent variable (exogenous variable ξ_{i}) but there is no explicit mathematical definition, while dispersion descriptors, such as nearest center distances and nearest boundary distances [7], can be viewed as indicators, represented by x’s. Different indicators (microstructure descriptors in this case) provide measurements of certain features of the microstructure. The error term, δ_{i}, in the general model can effectively take into account the errors introduced by the approximations in measurement. Considering measurement errors is another strength of the Structural Equation Modeling approach compared to other methods. Rather than assuming the different categories of descriptors are independent, the correlation between them can also be analyzed by studying the relationship among latent factors in the Structural Equation model.
where a is a fitting parameter that describes the shape of CF.
The structural error ζ_{i} describes unmodeled factors that may influence the endogenous variables η_{i}. If the exogenous variables ξ_{i} in the structured model cannot explain the endogenous variables η_{i} well, then a large structural error ζ may exist in the final model, which indicates that additional descriptors may need to be included.
As a result of applying Structural Equation Modeling analysis, key microstructure descriptors are chosen as material design variables. Ideally, for each identified significant latent factor, we want to pick one descriptor as the best indicator (or microstructural design variable). The choice of multiple descriptors within one latent factor will lead to redundancy as multiple descriptors are often correlated.
Constructing structureproperty relationship for microstructure design
As a simple illustration of microstructure design of nanodielectrics, the design objectives associated with the mechanical properties can be chosen as maximizing the real part ε ' and minimizing the energy loss, tan δ. Previous work has shown that dielectric permittivity in polymer nanocomposites can be analyzed using a Prony Series approach adapted from viscoelasticity studies. This approach incorporates explicit consideration of microstructure dispersion as well as polymer interphase between the nanofillers and matrix into the Finite Element simulation [44]. Based on experimental results of bimodal brush grafted silica nanoparticles in an epoxy matrix, finite element modeling has been used to accurately capture the dielectric permittivity and loss angle measured in experiments by superposition of frequencydependent dielectric relaxation constants. Optimization of nanodielectric materials can be achieved following the framework described in Fig. 8 once sufficient data are collected and multiple simulation property models are built.
Results and discussion
Characterization and reconstruction
As shown in Fig. 9b, the irregular clusters are each split into multiple ellipses, providing a better representation compared to the result in Fig. 9a where single ellipses are used to represent the irregular clusters. As a confirmation of improved accuracy using the splitting algorithm, we compare the interfacial area and the nearest center distance using the two methods. The interfacial area (2D boundary between the filler and matrix) in Fig. 9b is 0.0075 after using the splitting algorithm, which matches better with the real surface fraction 0.0078 of the original image, compared to 0.006 of the result in Fig. 9a without using the splitting algorithm. A similar observation is made for the characterized nearest center distance, which becomes 40 pixels after splitting versus 60 pixels before splitting. Nearest center distance reflects the local clustering behavior, and the “true” value is unknown as the evaluation depends on the way the cluster is characterized. The ellipse splitting algorithm implemented in this work is shown to be effective for splitting touching clusters and offering more accurate characterization. This splitting algorithm is utilized for all microstructure characterizations in the following sections of this paper.

Volume fraction (deterministic)

Nearest center distance (mean and variance)

Aspect ratio (mean, variance, normal distribution)

Cluster area (mean, exponential)
Descriptor information of the two sample microstructures (unit of length: nm)
Sample 1  Sample 2  

Volume fraction  0.53 %  0.9 % 
Nearest center distance  1st mean = 124, 1st var = 15,410  1st mean = 79, 1st var = 4151 
Aspect ratio  Mean = 0.7513, var = 0.0234  Mean = 0.9701, var = 0.0315 
Cluster area  Mean = 200  Mean = 10,120 
where S_{2}(r) is the target 2point CF and S_{2} ' (r) is the 2point CF of reconstruction. The average error from six reconstructions is found to be 5 %, which is acceptable for this low volume fraction material system.
By comparing Fig. 12a, b, it is noted that the 2point CF obtained using our proposed decomposition and reassembly strategy matches much more closely than using the original algorithm, especially in the short range from r = 50 to 100. The relative error achieved is 14.8 %, which is a big improvement compared to 48.6 % using the original approach. The results show that the uneven local information is maintained through the proposed strategy of using small blocks. It should be noted that when assembling the small blocks, a totally random sequence is applied in our study. Omitting the relationship between the subblocks is acceptable here for two reasons: (1) The large clusters are sparse in the microstructure images; therefore, it is difficult to come up with a statistical characterization that is representative for the whole image. (2) For such a low volume fraction system, it has been observed that the main differences among 2point CF occur in the short distance range, which has been captured by the information obtained from individual small blocks. 2point CF are usually oscillating with several peaks and valleys. The location of the first deepest drop corresponds to the size of local clusters and the peaks at longer distance relate to certain global periodical patterns. As for this case study, no obvious long distance pattern can be observed, so it is not necessary to consider the higher order blockblock relationships. In real implementation, the best results will be achieved by randomly assembling the small blocks multiple times and choosing the best match for the target CF, to compensate for the lack of blockblock characterization.
Identification of key descriptors from images using the Structural Equation Modeling approach
Descriptor set for statistical learning. Reference are provided for those descriptors without explicit meaning
Descriptor  Definition  Type 

Composition  
VF  Volume fraction  Deterministic 
Dispersion  
ncd  Cluster’s nearest surface distance  Statistical 
nbd  Cluster’s nearest center distance  Statistical 
ornang  Principle axis orientation angle [51]  Statistical 
intph  Surface area of matrix phase  Deterministic 
N  Cluster number  Deterministic 
Loc_VF  Local VF of Voronoi cells [52]  Statistical 
Geometry  
pores  Pore sizes (inscribed circle’s radius) [53]  Statistical 
area  Cluster area  Statistical 
rc  Equivalent radius, \( rc=\sqrt{A/\pi } \)  Statistical 
comp  Compactness [54]  Statistical 
rnds  Roundness [55]  Statistical 
eccen  Eccentricity [55]  Statistical 
els  Statistical  
rectan  Rectangularity [55]  Statistical 
tsst  Tortuosity [55]  Statistical 
The exploratory factor analysis (EFA) approach presented in the “Methods” section and the associated three steps are first followed to identify the proper number of latent factors and group the microstructure descriptors (indicators) based on their associations with the latent factors. When applying the EFA approach, only the data collected on microstructure descriptors from 117 images is used. Methods for choosing the number of latent factors have been well studied and three widely used criteria, K1 method (eigenvalue) [45], Nongraphical Cattell’s Scree Test (optimal coordinates and acceleration factor) [46], and Horn’s Parallel Analysis [47], are employed in Step 1 of EFA to build confidence in the result.
Results of EFA (loading, uniqueness, and complexity). The meaning of descriptors can be found in Table 2. Indices 1 and 2 indicate the mean and variance of the corresponding descriptor, respectively, e.g., area1 stands for the mean of clusters area
F1, cluster size  F2, distribution  F3, composition  F4, geometry1  F5, geometry2  Uniqueness^{**}  Complexity  

area1^{*}  0.99  −0.03  −0.08  −0.04  −0.09  0.0066  1 
area2  0.87  −0.02  0.19  −0.03  0.25  0.2016  1.3 
comp1  0.05  −0.06  0.13  −0.01  0.92  0.1244  1.1 
comp2  0.06  0.15  0.13  0.23  −0.28  0.7996  3.2 
eccen1  0.08  −0.03  0.08  0.19  −0.43  0.7363  1.6 
eccen2  0.09  0.24  0.12  −0.12  0.15  0.9222  3.3 
els1  0.27  −0.01  0.13  0.24  −0.32  0.6794  3.2 
els2  0.12  0.23  0.06  0.39  0.15  0.7368  2.3 
locvf1  0.37  −0.02  0.57  −0.1  −0.22  0.397  2.1 
locvf2  0.36  0.09  0.23  0.02  −0.14  0.7351  2.2 
nbd1  0.11  0.69  −0.38  −0.01  0.23  0.0913  1.9 
nbd2  −0.17  1.02  0.08  −0.03  −0.17  0.0544  1.1 
ncd1  0.25  0.65  −0.38  −0.01  0.16  0.0949  2.1 
ncd2  −0.13  1.01  0.08  −0.04  −0.18  0.0496  1.1 
ornang1  −0.04  −0.09  −0.03  0.16  0.04  0.9673  2 
ornang2  −0.03  −0.02  0.07  0.15  0.28  0.8921  1.7 
pores1  0.92  −0.09  −0.35  −0.1  −0.02  0.2054  1.3 
pores2  0.92  0.01  −0.04  0  0.06  0.1733  1 
rc1  0.88  −0.01  −0.22  −0.06  −0.28  0.0704  1.4 
rc2  0.97  −0.02  0.05  −0.01  0.11  0.0905  1 
rctan1  −0.1  −0.03  −0.09  1  −0.04  0.0075  1 
rctan2  −0.1  −0.05  −0.1  1  −0.04  0.015  1 
rnds1  0.03  −0.06  −0.02  −0.1  0.75  0.4154  1.1 
rnds2  0.21  0.01  0.06  0.21  0.25  0.8562  3.1 
ttst1  0.05  −0.08  0.15  0.51  0.44  0.5319  2.2 
ttst2  0.02  0.03  −0.09  0.85  −0.21  0.1797  1.2 
intph0  −0.07  −0.02  0.94  −0.09  0.01  0.0992  1 
n  −0.37  0.03  0.91  −0.08  0.21  0.0726  1.5 
vf  0.49  −0.03  0.74  −0.08  0.15  0.0977  1.8 
In Step 3 of the EFA procedure, based on the rule that at least half of the variance of an independent variable should be explained by a latent factor (uniqueness ≤0.5), we can identify poor factor indicators (underlined) and withdraw them from our data set. Going through steps 1 to 3 in EFA, we reduced the number of characterization parameters from 29 to 19 and associate them to five latent factors. Examining the results, we can relate each factor to a physical interpretation: Factor 1 represents the size of clusters, with 6 descriptors as indicators: cluster area (area1, area2), pore size (pores1, pores2), and equivalent radius (rc1, rc2). Factor 2 represents the distribution status, with the nearest neighbor distance such as nearest boundary distance (nbd) and nearest center distance (ncd) as the indicator. Factor 3 describes the composition information, including volume fraction (vf), interfacial fraction (intph), number of clusters (n), and local volume fraction (locvf1). Volume fraction clearly describes the composition, and the other three describe higher order composition information. Factors 4 and 5 represent two geometric characteristics: Factor 4 is associated with rectangularity (rctan1, rctan2), and tortuosity (ttst2); Factor 5 is associated with compactness (comp1) and roundness (rnds1).
The obtained Structural Equation Model is verified based on the physical meaning of factor coefficients. For example, Factor 1 represents the size of clusters and it is meaningful that it has a negative effect (−0.542) on the surface correlation and a strong positive effect (0.893) on the lineal path correlation. Intuitively, a larger cluster size leads to smaller surface area and larger lineal path. Based on the Structural Equation Model analysis, a potential set of material design variables are chosen by including four descriptor variables, each underlines Factor 1, Factor 2, Factor 3, and Factor 5, respectively. Based on the loadings of each descriptor with respect to each latent factor, the most significant set is identified to be area1 (0.995, average cluster area) for Factor 1, ncd1 (0.959, average nearest center distance) for Factor 2, vf (0.943, volume fraction) for Factor 3, and rnds1 (0.898, average cluster roundness) for Factor 5.
Correlation comparison of chosen descriptors. Proposed and prior ranking algorithm, high correlation among area1 (average cluster area), pores2 (variance of pore size), and rc2 (variance of equivalent radius)
Structural Equation Modelingbased analysis  vf  area1  ncd1  rnds1 
Vf—volume fraction  
area1—average cluster area  0.46  
ncd1—average nearest center distance  −0.32  0.33  
rnds1—average cluster roundness  0.08  −0.23  −0.10  
Prior ranking algorithm [21]  vf  area1  pores2  rc2 
Vf—volume fraction  
area1—average cluster area  0.56  
pores2—variance of pore size  0.43  0.90  
rc2—variance of equivalent radius  0.54  0.93  0.91 
Validation using reconstruction and property simulations
Comparison of dielectric properties of the original and reconstructed images. Eight samples are chosen; properties are averaged over five reconstructions for each sample
Dielectric constant ε '  Loss angle tan δ  

Original  Reconstruction (average)  Error (%)  Original  Reconstruction (average)  Error (%)  
1  4.215  4.204  0.254  0.0167  0.0161  3.50 
2  4.084  4.084  0.016  0.0173  0.0164  4.74 
3  3.926  3.941  0.384  0.0190  0.0185  2.78 
4  3.897  3.903  0.168  0.0192  0.0190  0.92 
5  4.109  4.173  1.546  0.0167  0.0158  5.36 
6  3.910  3.920  0.253  0.0191  0.0188  1.59 
7  3.962  3.971  0.212  0.0185  0.0181  2.14 
8  3.981  4.009  0.709  0.0181  0.0175  3.24 
Average  0.443  Average  3.50 
Results of linear model fitting
Model  CF (six parameters)  Key descriptors (four)  Standardized coefficients of descriptors  

Prediction error  Rsquared  Prediction error  Rsquared  Average cluster area  Volume fraction  Average nearest distance  Average roundness  
ε '  5.53e−04  0.95  8.36e−04  0.93  −0.419^{a}  1.060^{a}  0.0008  0.096^{a} 
tan δ  1.12e−07  0.91  7.5e−08  0.88  0.614^{a}  −1.034^{a}  0.0043  −0.048 
Conclusions
In this paper, new characterization, reconstruction, and key microstructure feature identification techniques are developed to support the computational design of nanodielectric systems. For design representation, a descriptorbased characterization and reconstruction method is employed and tailored for a low volume fraction nanodielectric system with uneven local aggregations and irregularly shaped clusters. To handle special microstructures with large local aggregates, we propose a new decomposition and reassembly strategy, based on which the reconstruction accuracy is greatly improved. We also incorporate a touching cell splitting algorithm into the descriptorbased method to deal with irregularly shaped clusters to achieve more realistic characterization. To simplify the material design process and minimize the redundancy among design variables, a new Structural Equation Modelbased method is developed to identify key descriptors. To keep the results independent from the material properties of interest, the analysis presented in this paper is based on information from the microstructure images (CF). According to the fitted Structural Equation Model, in which descriptors are classified into five groups based on the identified latent factors, we find volume fraction, cluster size, nearest center distance, and cluster roundness to be a sufficient set of descriptors to represent the structural features of the very low volume fraction nanodielectric system in this research. The relationship between the microstructure and properties are explored based on the epoxysilica system and a close to linear relationship is observed between dielectric permittivity and the identified key descriptors.
In a future work, more image data of nanodielectric material systems with a wider range of volume fraction will be collected and simulation models for predicting all important dielectric properties will be included. In addition, the design problem will be extended to include both permittivity and breakdown strength as objectives. To make the tradeoff between the two, the Pareto frontier will be used to first identify a set of nondominated (best achievable) optimal solutions. In addition, the Structural Equation Modelbased method can also be applied to find the relations between descriptors and certain properties, which can then be directly used as predictive models in material design. Processing conditions will be taken into consideration to establish the mapping relations across the chain of processingstructureproperty to ensure the manufacturability of new nanodielectric materials.
Availability of supporting data
Data presented in this work will be made available upon request.
Notes
Acknowledgements
The support from NSF for this collaborative research: CMMI1334929 (Northwestern University) and CMMI1333977 (RPI), is greatly appreciated.
References
 1.Nalwa HS (1999) Handbook of low and high dielectric constant materials and their applications, twovolume set., Academic Press, Waltham, Massachusetts, USAGoogle Scholar
 2.Barber P, Balasubramanian S, Anguchamy Y, Gong S, Wibowo A, Gao H, Zur Loye HC (2009) Polymer composite and nanocomposite dielectric materials for pulse power energy storage. Materials 2(4), 1697–1733CrossRefGoogle Scholar
 3.McPherson JW, Kim J, Shanware A, Mogul H, Rodriguez J (2003) Trends in the ultimate breakdown strength of high dielectricconstant materials. Electron Devices, IEEE Transactions on, 50(8):1771–1778CrossRefGoogle Scholar
 4.Ding HZ, Varlow BR (2004) Effect of nanofillers on electrical treeing in epoxy resin subjected to AC voltage. In: Electrical Insulation and Dielectric Phenomena, 2004. CEIDP'04. 2004 Annual Report Conference on. IEEE, pp 332–335Google Scholar
 5.Olson GB (1997) Computational design of hierarchically structured materials. Science 277(5330):1237–1242CrossRefGoogle Scholar
 6.McDowell DL, Olson GB (2009) Concurrent design of hierarchical materials and structures. In: Scientific Modeling and Simulations, Springer Netherlands, pp 207–240Google Scholar
 7.Xu H, Dikin DA, Burkhart C, Chen W (2014) Descriptorbased methodology for statistical characterization and 3D reconstruction of microstructural materials. Comput Mat Sci 85:206–216CrossRefGoogle Scholar
 8.Torquato S, Stell G (1982) Microstructure of twophase random media. I. The npoint probability functions. J Chem Phys 77(4):2071–2077CrossRefGoogle Scholar
 9.Torquato S, Stell G (1983) Microstructure of twophase random media. II. The Mayer–Montroll and Kirkwood–Salsburg hierarchies. J Chem Phys 78(6):3262–3272CrossRefGoogle Scholar
 10.Lu B, Torquato S (1992) Linealpath function for random heterogeneous materials. Phys Rev A 45(2):922CrossRefGoogle Scholar
 11.Torquato S, Beasley J, Chiew Y (1988) Twopoint cluster function for continuum percolation. J Chem Phys 88(10):6540–6547CrossRefGoogle Scholar
 12.Xu H, Li Y, Brinson C, Chen W (2014) A descriptorbased design methodology for developing heterogeneous microstructural materials system. J Mech Des 136(5):051007CrossRefGoogle Scholar
 13.Jiao Y, Chawla N (2014) Three dimensional modeling of complex heterogeneous materials via statistical microstructural descriptors. Integ Mat Manufac Innov 3(1):1–19CrossRefGoogle Scholar
 14.Chen D, He X, Teng Q, Xu Z, Li Z (2014) Reconstruction of multiphase microstructure based on statistical descriptors. Physica A: Statistical Mechanics and its Applications 415:240–250CrossRefGoogle Scholar
 15.Xu H, Li Y, Brinson C, Chen W (2014) A descriptorbased design methodology for developing heterogeneous microstructural materials system. J Mechanical Design 136(5):051007CrossRefGoogle Scholar
 16.Xu H, Li Y, Brinson C, Chen W (2013) Descriptorbased methodology for designing heterogeneous microstructural materials system. In: ASME 2013 International Design Engineering Technical Conferences and Computers and Information in Engineering Conference. American Society of Mechanical Engineers. pp V03AT03A049V03AT03A049Google Scholar
 17.Ferris KF, Peurrung LM, Marder JM (2007) Materials informatics: fast track to new materials. Advan Mater Processes 165(1):50–51, 165(PNNLSA52427)Google Scholar
 18.Wei Q, Peng X, Liu X, Xie W (2006) Materials informatics and study on its further development. Chinese Sci Bulletin 51(4):498–504CrossRefGoogle Scholar
 19.Sundararaghavan V, Zabaras N (2005) Classification and reconstruction of threedimensional microstructures using support vector machines. Comput Mater Sci 32(2):223–239CrossRefGoogle Scholar
 20.Basanta D, Miodownik MA, Holm EA, Bentley PJ (2005) Using genetic algorithms to evolve threedimensional microstructures from twodimensional micrographs. Metall Mater Trans A 36(7):1643–1652CrossRefGoogle Scholar
 21.Xu H, Liu R, Choudhary A, Chen W (2015) A machine learningbased design representation method for designing heterogeneous microstructures. J Mechanic Design 137(5):051403CrossRefGoogle Scholar
 22.Bai X, Sun C, Zhou F (2008) Touching cells splitting by using concave points and ellipse fitting., pp 271–278Google Scholar
 23.Vinzi VE, Trinchera L, Amato S (2010) PLS path modeling: from foundations to recent developments and open issues for model assessment and improvement. In: Handbook of partial least squares, Springer Berlin Heidelberg, pp 47–82Google Scholar
 24.Goldstein J, Newbury DE, Echlin P, Joy DC, Romig Jr AD, Lyman CE, Lifshin E (2012) Scanning electron microscopy and Xray microanalysis: a text for biologists, materials scientists, and geologists. Springer Science & Business Media, New York, Philadelphia, USAGoogle Scholar
 25.Williams DB, Carter CB (1996) The transmission electron microscope. Springer, US, p 317CrossRefGoogle Scholar
 26.Forsyth DA, Ponce J (2002) Computer vision: a modern approach., Prentice Hall Professional Technical Reference, New Jersey, USAGoogle Scholar
 27.Morisita M (1962) I σIndex, a measure of dispersion of individuals. Res Popul Ecol 4(1):1–7CrossRefGoogle Scholar
 28.Xu H, Greene MS, Deng H, Dikin D, Brinson C, Liu W K, Chen W (2013) Stochastic reassembly strategy for managing information complexity in heterogeneous materials analysis and design. J Mech Des 135(10):101010CrossRefGoogle Scholar
 29.Khare HS, Burris DL (2010) A quantitative method for measuring nanocomposite dispersion. Polymer 51(3):719–729CrossRefGoogle Scholar
 30.Luo ZP, Koo JH (2007) Quantifying the dispersion of mixture microstructures. J Microsc 225(2):118–125CrossRefGoogle Scholar
 31.Loehlin JC (1998) Latent variable models: an introduction to factor, path, and structural analysis., Lawrence Erlbaum Associates Publishers, New Jersey, USAGoogle Scholar
 32.Thompson B (2004) Exploratory and confirmatory factor analysis: understanding concepts and applications., American Psychological Association, Washington, DC, USACrossRefGoogle Scholar
 33.Tenenhaus M, Vinzi VE, Chatelin YM, Lauro C (2005) PLS path modeling. Computational statistics & data analysis 48(1):159–205.CrossRefGoogle Scholar
 34.Hair JF Jr, Hult GTM, Ringle C, Sarstedt M (2013) A primer on partial least squares structural equation modeling (PLSSEM). Sage Publications. California, USAGoogle Scholar
 35.Hair JF, Sarstedt M, Ringle CM, Mena JA (2012) An assessment of the use of partial least squares structural equation modeling in marketing research. J Acad Marketing Sci 40(3):414–433CrossRefGoogle Scholar
 36.Chin WW (1998) The partial least squares approach to structural equation modeling. Modern Methods Bus Res 295(2):295–336Google Scholar
 37.Debye P, Anderson HR Jr, Brumberger H (1957) Scattering by an inhomogeneous solid. II. The correlation function and its application. J Appl Phys 28(6):679–683CrossRefGoogle Scholar
 38.Lee SY (1990) Covariance structure analysis. Structural Equation Modeling: A Bayesian Approach., pp 31–66Google Scholar
 39.Wold S, Martens H, Wold H (1983) The multivariate calibration problem in chemistry solved by the PLS method, in Matrix pencils, Springer, pp 286–293Google Scholar
 40.Fornell C, Bookstein FL (1982) Two structural equation models: LISREL and PLS applied to consumer exitvoice theory. J Marketing research 19:440–452CrossRefGoogle Scholar
 41.Kock N (2013) WarpPLS 4.0 user manual. ScriptWarp Systems, LaredoGoogle Scholar
 42.Jin R, Chen W, Sudjianto A (2005) An efficient algorithm for constructing optimal design of computer experiments. J Stat Plan Inference 134(1):268–287CrossRefGoogle Scholar
 43.Stein ML (1999) Interpolation of spatial data: some theory for Kriging., SpringVerlag, New YorkGoogle Scholar
 44.Huang Y, Krentz TM, Nelson JK, Schadler LS, Li Y, Zhao H, Breneman CM (2014) Prediction of interface dielectric relaxations in bimodal brush functionalized epoxy nanodielectrics by finite element analysis method. In Electrical Insulation and Dielectric Phenomena (CEIDP), 2014 IEEE Conference on. IEEE pp 748751Google Scholar
 45.Kaiser HF (1960) The application of electronic computers to factor analysis, Educational and psychological measurementCrossRefGoogle Scholar
 46.Raîche G et al. (2013) Nongraphical solutions for Cattell’s scree test. Methodology 9(1):23CrossRefGoogle Scholar
 47.Horn JL (1965) A rationale and test for the number of factors in factor analysis. Psychometrika 30(2):179–185CrossRefGoogle Scholar
 48.Norris M, Lecavalier L (2010) Evaluating the use of exploratory factor analysis in developmental disability psychological research. J Autism Dev Disord 40(1):8–20CrossRefGoogle Scholar
 49.Hofmann RJ (1978) Complexity and simplicity as objective indices descriptive of factor solutions. Multivariate Behav Res 13(2):247–250CrossRefGoogle Scholar
 50.Roy M, Nelson JK, MacCrone RK, Schadler LS, Reed CW, Keefe R (2005) Polymer nanocomposite dielectricsthe role of the interface. Dielectrics and Electrical Insulation, IEEE Transactions on, 12(4):629643.CrossRefGoogle Scholar
 51.Ganesh VV, Chawla N (2005) Effect of particle orientation anisotropy on the tensile behavior of metal matrix composites: experiments and micro structurebased simulation. Mater Sci Eng aStructural Materials Properties Microstructure Processing 391(12):342–353CrossRefGoogle Scholar
 52.Thomas M, Boyard N, Perez L, Jarny Y, Delaunay D (2008) Representative volume element of anisotropic unidirectional carbon–epoxy composite with highfibre volume fraction. Composites Sci Technol 68(15):3184–3192CrossRefGoogle Scholar
 53.Kenney B, Valdmanis M, Baker C, Pharoah JG, Karan K (2009) Computation of TPB length, surface area and pore size from numerical reconstruction of composite solid oxide fuel cell electrodes. J Power Sources 189(2):1051–1059CrossRefGoogle Scholar
 54.Morozov IA, Lauke B, Heinrich G (2011) A novel method of quantitative characterization of filled rubber structures by AFM. KgkKautschuk Gummi Kunststoffe 64(12):24–27Google Scholar
 55.Prakash CP, Mytri VD, Hiremath PS (2010) Classification of Cast Iron Based on Graphite Grain Morphology using Neural Network Approach. Second International Conference on Digital Image Processing, International Society for Optics and Photonics, pp 75462S–75462SGoogle Scholar
 56.Klaysom C, Moon SH, Ladewig BP, Lu GM, Wang L (2011) The effects of aspect ratio of inorganic fillers on the structure and property of composite ionexchange membranes. J Colloid Interface Sci 363(2):431–439CrossRefGoogle Scholar
 57.Jean A, Jeulin D, Forest S, Cantournet S, N'GUYEN F (2011) A multiscale microstructure model of carbon black distribution in rubber.J Microsc 241(3):243–260CrossRefGoogle Scholar
Copyright information
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.