betaFeedback

WorksPhysics and Astronomy

Stochastic reconstruction of multiphase composite microstructures using statistics-encoded neural network for poro/micro-mechanical modelling

Abstract

Understanding microstructure-property relationships (MPRs) is essential for optimising the performance of multiphase composites. Image-based poro/micro-mechanical modelling provides a non-invasive approach to exploring MPRs, but the randomness of multiphase composites often necessitates extensive 3D microstructure datasets for statistical reliability. This study introduces a cost-effective machine learning framework to reconstruct numerous virtual 3D microstructures from limited 2D exemplars, circumventing the high costs of volumetric microscopy. Using feedforward neural networks, termed the statistics-encoded neural network (SENN), the framework encodes 2D morphological statistics and infers 3D morphological statistics via a 2D-to-3D integration scheme. Statistically equivalent 3D microstructures are synthesised using Gibbs sampling. Hierarchical characterisation enables seamless capture of features across multiple scales. Validation on three composites demonstrates strong statistical equivalence between reconstructed and reference microstructures, confirmed by morphological descriptors and simulated macroscopic properties (e.g., stiffness, permeability). The SENN-based framework is a high-fidelity tool for efficiently and accurately reconstructing multiphase microstructures.

Introduction

The internal microstructure of composite materials plays a pivotal role in determining their macroscopic properties [1, 2], including mechanical strength, permeability, effective diffusivity, thermal conductivity, and other critical functionalities. Multiphase composites, in particular, exhibit complex microstructural patterns characterised by heterogeneous phase distributions and intricate interfacial geometries [3]. A fundamental understanding of the microstructure-property relationships (MPRs) in the multiphase composites materials is crucial for optimising their performance in various applications [4, 5, 6], such as lightweight and durable materials for aerospace and automotive industries, electrodes or solid-state electrolyte for energy storage, biocompatible implants and scaffold in medical devices, corrosion-resistant structures for marine and constructions applications, and impact-resistant materials for offshore wind turbine blades. However, the inherent randomness and stochastic nature of multiphase composites pose significant challenges to precise microstructural analysis and accurate MPR modelling [7, 8].

One of the most reliable approaches to studying MPRs involves the use of 3D digital microstructures. These 3D datasets allow researchers to analyse the internal organisation of materials and perform image-based numerical simulations to link microstructural features to macroscopic properties [6, 9, 10]. High-resolution 3D images of composite materials are typically obtained using advanced microscopy techniques such as micro-computed tomography (micro-CT) [11] and scanning electron microscopy (SEM) [12]. While these imaging techniques have provided unprecedented insights into microstructural features, they are often associated with high costs, limited accessibility, and technical constraints, especially when dealing with opaque or fragile materials. Furthermore, the demand for large datasets of 3D digital microstructures to conduct reliable statistical analyses [13, 14, 15] further amplifies these limitations, especially for Monte Carlo analysis.

To address this issue, there has been growing interest in using computational methods to synthesise 3D microstructures from limited data [16, 17]. Stochastic reconstruction techniques [18, 19] leverage statistical descriptions of microstructures to generate virtual 3D samples that are statistically equivalent to real materials. These techniques can broadly be classified into two categories: conventional algorithm-based approaches and computational intelligence-based approaches. Typical conventional methods include stochastic optimisation-based reconstruction [20, 21], Gaussian random field transformation [22], process-based reconstruction [23], multiple-point statistics [24], patch-based reconstruction [25], texture synthesis [26], descriptor-based methodologies [27, 28], and packing algorithms [29]. These methods are often mathematically rigorous, interpretable, and allow for systematic variation of preserved microstructural characteristics by adjusting reconstruction targets or parameters. However, conventional methods face significant limitations when handling complex, multiphase microstructures, as they struggle to accurately capture the intricate geometries, correlations, and dependencies among constituent phases. While capable of generating mathematically equivalent microstructures, these methods fail to produce physically consistent ones, which are critical for preserving fine details required in MPR analysis. In addition, these methods are predominantly developed for bi-phase materials and exhibit limited capability in reconstructing multiphase composite microstructures.

Computational intelligence [30] has unlocked exciting opportunities for stochastic reconstruction of random microstructures, driven by rapid advancements in AI algorithms, processors, data storage, and computing power. A key advantage of deep learning algorithms [31] is their ability to operate without manual feature design, extraction, or selection, streamlining the reconstruction process. As a result, a wide range of computational intelligence-based approaches have been developed for reconstructing random materials. Simple machine learning algorithms, such as support vector machines [32], decision trees [33, 34], and feedforward neural networks [17, 35], require manually designed input features to capture specific morphological characteristics. In contrast, deep learning algorithms automatically extract hierarchical features from raw input images using convolutional or other image processing operations. Notable methods include generative adversarial networks [36, 37, 38], convolutional neural networks [39, 40], deep belief networks [41], variational autoencoders [42], transfer learning [43], generative flow networks [44], recurrent neural networks [45], U-Nets [46], and diffusion-based generative models [47]. Despite their advantages, deep learning models suffer from poor interpretability due to their complex architectures, making it difficult to determine what information learned from the real microstructure is embedded into the reconstructed samples. Most existing methods target bi-phase materials, such as porous media. While some techniques [46, 47, 48, 49] have been applied to reconstruct multiphase microstructures by maintaining morphological similarity, they have not demonstrated the ability to preserve various physical properties. Additionally, computational intelligence-based methods often require oversized datasets, extensive computational resources, and face challenges with inefficient convergence, limiting their scalability and applicability for large-scale analyses.

This study presents an innovative Statistics-Encoded Neural Network (SENN)-based framework for stochastic characterisation and reconstruction of 3D multiphase composite microstructures. Unlike conventional methods that rely solely on statistical modelling, the proposed framework leverages machine learning to encode 2D morphological statistics into feedforward neural networks, facilitating the efficient and accurate reconstruction of statistically equivalent 3D microstructures. Once properly trained, SENN models serve as implicit representations of morphological statistics, offering a high level of interpretability. The SENN-based framework addresses key challenges in stochastic reconstruction of multiphase microstructures by:

  • Encoding complex morphological statistics of multiphase microstructures using feedforward neural networks, where only a limited number of 2D exemplars is required;

  • Employing a 2D-to-3D morphology integration scheme to infer 3D morphological statistics from the 2D measurements embedded in SENN models;

  • Incorporating hierarchical characterisation to capture local, regional, and global microstructural features across multiple length scales;

  • Synthesising statistically equivalent 3D microstructures via Gibbs sampling in a multi-level manner, ensuring fidelity to the real material’s statistical properties.

The versatility and robustness of the novel SENN-based framework are demonstrated through its application to three representative multiphase composites: porous silver-based electrodes (3-phase), porous solid oxide fuel cell anodes (3-phase), and composite cement paste (4-phase). These materials are selected due to their distinct microstructural characteristics and relevance to diverse engineering applications. The reconstructed 3D microstructures are evaluated against reference datasets using various morphological descriptors, including volume fraction, surface area density, triple-phase boundary density, and geometrical tortuosity. Furthermore, macroscopic properties such as stiffness, permeability, diffusivity, and thermal conductivity tensors are computed through image-based numerical simulations to validate the practical applicability of the reconstructed 3D microstructures.

The remainder of this paper is organised as follows: § 2 explains the theoretical foundation and statistical characterisation of multiphase microstructures in detail, including the Markov random field assumption and the SENN modelling; § 3 outlines the methodology for 2D-to-3D stochastic reconstruction, focusing on hierarchical characterisation, multi-level reconstruction, and Gibbs sampling techniques; § 4 presents case studies on the three multiphase composites and evaluates the quality of microstructure reconstruction; § 5 highlights the results of image-based poro/micro-mechanical modelling on the reconstructed 3D microstructures; and § 6 discusses the limitations and potential future directions and concludes with a summary of the SENN-based framework’s contributions.

Statistical microstructure characterisation

The internal microstructures of opaque composite materials can be digitised via advanced microscopic imaging techniques, such as SEM and micro-CT, as shown in Figure 1a. Image segmentation is then applied to the digital images to distinguish and separate different constituent phases, as illustrated in Figure 1b and Figure 1c. This segmentation enables subsequent analyses, including quantitative characterisation and computational modelling. Statistically quantitative characterisation of multiphase microstructures is a crucial preliminary step in generating equivalent virtual samples, given the significant disorder and randomness existing in composite microstructures.

Graphical illustration of image segmentation

Figure 1. Graphical illustration of image segmentation: (a) A 3D microscopic image of Titanate ceramic acquired by scanning electron microscopy (SEM) in backscattered electron imaging (BEI) mode; (b) The histogram of voxel intensity values, based on which the segmentation thresholds can be determined; and (c) The digital microstructure with four distinct constituent phases, where the white phase represents metals, the light grey phase is zirconolite, the dark grey phase is hollandite, and the black phase denotes perovskite loveringite or spinel.

Consider a 3D regular grid of size L×W×HL \times W \times H, where each site (i,j,k)(i,j,k) is associated with a random variable Xijk∈QX_{ijk} \in\mathbb{Q}, forming a random field XX, defined as:

X={Xijk:1≤i≤L, 1≤j≤W, 1≤k≤H},(1)X = \{X_{ijk} : 1 \le i \le L,\ 1 \le j \le W,\ 1 \le k \le H\}, \tag*{(1)}

where i∈Z+i \in\mathbb{Z}^{+}, j∈Z+j \in\mathbb{Z}^{+} and k∈Z+k \in\mathbb{Z}^{+} are the indices of rows, columns and layers in the 3D random field. Assigning a voxel value XijkX_{ijk} into each site (i,j,k)(i,j,k) of this 3D grid generates a 3D digital image XX. This 3D digital microstructure XX, in its segmented format (see Figure 1c) can be represented by an indicator function as follows:

Xijk={0,if (i,j,k) is located at the black phase;1,if (i,j,k) is located at the dark grey phase;⋮n−1,if (i,j,k) is located at the white phase;(2)X_{ijk} = \begin{cases} 0, & \text{if } (i,j,k) \text{ is located at the black phase};\\ 1, & \text{if } (i,j,k) \text{ is located at the dark grey phase};\\ \vdots\\ n-1, & \text{if } (i,j,k) \text{ is located at the white phase}; \end{cases} \tag*{(2)}

where nn denotes the number of constituent phases in the multiphase composite.

Markov random field assumption

The full joint probability distribution function (PDF) of voxel variables provides an ideal statistical model for quantitatively characterising the digital microstructure X\mathbf{X}, because the voxel variables XijkX_{ijk} at different locations exhibit statistical correlations. This full joint PDF, denoted by P(X)P(\mathbf{X}) or P(X111,X112,…,Xijk,…,XLWH)P(X_{111}, X_{112}, \ldots, X_{ijk}, \ldots, X_{LWH}), represents the likelihood of a digital microstructure X\mathbf{X} being assigned a specific configuration of all voxel values. However, deriving the full joint PDF P(X)P(\mathbf{X}) without any prior knowledge necessitates enumerating all possible configurations of voxel variables. This task becomes computationally prohibitive due to the extremely high dimensionality (L×W×H)(L \times W \times H) of voxel space.

The Markov assumption [50] is employed in this study to simplify the statistical characterisation of composite microstructures by reducing computational complexity. Under this assumption, a digital microstructure is represented as a Markov random field (MRF), where the voxel value at any site depends only on its neighbouring voxels within a specified range. This dependency can be expressed as:

P(Xijk∣X(−ijk))=P(Xijk∣Nijk),(3)P(X_{ijk} \mid\mathbf{X}^{(-ijk)}) = P(X_{ijk} \mid N_{ijk}), \tag*{(3)}

where NijkN_{ijk} represents the set of adjacent voxels around a central voxel XijkX_{ijk}, and X(−ijk)\mathbf{X}^{(-ijk)} denotes all other voxels in the microstructure X\mathbf{X}, excluding XijkX_{ijk}. This conditional probability distribution function (CPDF) P(Xijk∣Nijk)P(X_{ijk} \mid N_{ijk}) has significantly lower dimensionality compared to the full joint PDF P(X)P(\mathbf{X}). The reduced dimensionality makes it feasible to approximate P(Xijk∣Nijk)P(X_{ijk} \mid N_{ijk}) by evaluating all possible configurations of the neighbouring voxels. Consequently, the full joint PDF P(X)P(\mathbf{X}) can be efficiently derived from these local statistical properties. This approach enables effective modelling of the microstructure’s stochastic behaviour without requiring exhaustive computation.

To model the local statistical characteristics of a digital microstructure, the initial step is to define the neighbourhood geometry surrounding a central voxel XijkX_{ijk}. This neighbourhood, denoted as NijkN_{ijk}, is confined within a spatial local region characterised by its radii (Rxy,Ryz(R_{xy}, R_{yz} and Rzx)R_{zx}), as illustrated in Figure 6 (a 2D counterpart is illustrated in Figure 2b). The data template scans the digital microstructure X\mathbf{X} to gather local morphology patterns, enabling the approximation of the CPDF P(Xijk∣Nijk)P(X_{ijk} \mid N_{ijk}). Each observation of a local morphology pattern, denoted by (Xijk,Nijk)(X_{ijk}, N_{ijk}), is refereed as a data event. The CPDF P(Xijk∣Nijk)P(X_{ijk} \mid N_{ijk}) is estimated from occurrence frequencies of such events, as expressed in the following equations:

Graphical illustration of training data preparation

Figure 2. Graphical illustration of training data preparation: (a) A 2D exemplar of a composite microstructure with three constituent phases (n=3n = 3); (b) Collection of data events using a square data template defined by radius RR to scan the 2D exemplar; (c) Example illustrating a data event (Xij,NijX_{ij}, N_{ij}), where XijX_{ij} is the central pixel, and NijN_{ij} represents its neighbouring pixels; (d) Separation of the central pixel XijX_{ij} from its neighbouring pixels NijN_{ij}; (e) Conversion of the central pixel value XijX_{ij} to a probability vector VijV_{ij}; and (f) Data conversion process for training.

A 3D data template composed of three 2D data templates on three orthogonal planes

Figure 6. A 3D data template composed of three 2D data templates (see Figure 2b) on three orthogonal planes. The entire neighbourhood data NijkN_{ijk} in 3D space consist of neighbouring voxels Nijk(xy)N_{ijk}^{(xy)}, Nijk(yz)N_{ijk}^{(yz)} and Nijk(zx)N_{ijk}^{(zx)} on xyxy-, yzyz- and zxzx-planes, and the neighbourhood ranges are measured by RxyR_{xy}, RyzR_{yz} and RzxR_{zx}, respectively.

P(Xijk∣Nijk)=P(Xijk,Nijk)P(Nijk)=O(Xijk,Nijk)O(Nijk),(4)P(X_{ijk} \mid N_{ijk}) = \frac{P(X_{ijk}, N_{ijk})}{P(N_{ijk})} = \frac{O(X_{ijk}, N_{ijk})}{O(N_{ijk})}, \tag*{(4)}

where O(Xijk,Nijk)O(X_{ijk}, N_{ijk}) and O(Nijk)O(N_{ijk}) denote the occurrence counts of data events associated with (Xijk,Nijk)(X_{ijk}, N_{ijk}) and (Nijk)(N_{ijk}) respectively. Consequentially, the CPDF P(Xijk∣Nijk)P(X_{ijk} \mid N_{ijk}) quantifies the probability of a central voxel XijkX_{ijk} taking a specific value, given its neighbouring voxels NijkN_{ijk}. By leveraging the localised CPDF P(Xijk∣Nijk)P(X_{ijk} \mid N_{ijk}), the full joint PDF P(X)P(\mathbf{X}) that statistically characterises the random microstructures can be achieved through Gibbs sampling, which will be detailed in § 3.

Statistics-encoded neural network

Confronting with the common scenario where 3D digital microstructures are scarce while 2D examples are plentiful, this work aims to develop a machine learning-based framework to infer statistically equivalent 3D microstructures from the available 2D exemplars for multiphase composites. Achieving this objective necessitates the statistical characterisation of 3D morphological patterns derived from 2D examples as a foundational step. As discussed earlier, the CPDF P(Xijk∣Nijk)P(X_{ijk} \mid N_{ijk}) encapsulates the statistics of local morphology patterns within random composite microstructures. Considering the inherent complexity of microstructural features, this work employs machine learning to approximate the CPDF P(Xijk∣Nijk)P(X_{ijk} \mid N_{ijk}), where morphological statistics are integrated into feed-forward neural networks, refereed to as Statistics-Encoded Neural Network (SENN). These models efficiently integrate statistical data, enabling the accurate inference of CPDFs and facilitating the synthesis of statistically consistent 3D microstructures from 2D exemplars.

As illustrated in Figure 2a-Figure 2d, data events (Xij,Nij)(X_{ij}, N_{ij}) are efficiently collected by scanning the 2D exemplar in a raster scanning order using the predefined data template. These paired observations (Xij,Nij)(X_{ij}, N_{ij}) form an ideal training dataset for fitting a classifier in the supervised learning paradigm. By minimising misclassification likelihood, the class probabilities encoded in the classifier provide an accurate approximation of the CPDF P(Xij∣Nij)P(X_{ij}|N_{ij}). Among various classifiers, the feedforward neural network (FNN) is particularly effective in handling non-linear classification challenges inherent to image modelling. As depicted in Figure 2e and Figure 2f, the central pixel value XijX_{ij} is converted into a probability vector VijV_{ij}, ensuring that the trained FNN model’s output corresponds to class probabilities equivalent to the CPDF P(Xij∣Nij)P(X_{ij}|N_{ij}). This data conversion process is reversible, which is essential for stochastic reconstruction of 3D microstructures through probabilistic sampling. Furthermore, this conversion procedure can be readily extended to composite microstructures with more than three constituent phases (i.e., n≥3n \ge3).

To fit the FNN model for approximating the CPDF P(Xij∣Nij)P(X_{ij}|N_{ij}), the neighbourhood pattern NijN_{ij} is used as the classification feature (input), while the central pixel value XijX_{ij} serves as the classification category (output). It is crucial that NijN_{ij} is flatten into a feature vector, and XijX_{ij} is converted to be a probability vector VijV_{ij} prior to training the FNN model, resulting in the training dataset (Vij,Nij)(V_{ij}, N_{ij}). In a multi-layer FNN model, the input feature vector NijN_{ij} is propagated through the hidden layers, and the final output, representing the estimated probability vector V^ij\widehat{V}_{ij}, is computed through a series of forward-propagation equations. In essence, the fitted FNN model is a vector-valued surrogate that approximates the mapping between NijN_{ij} and VijV_{ij} for the 2D training image XX, which can be mathematically expressed as follows:

FNN(Nij;W,b):Nij∈RdN⟶Vij∈Rn,(5)\mathcal{F}NN(N_{ij};\boldsymbol{W},\boldsymbol{b}):N_{ij}\in\mathbb{R}^{d_N}\longrightarrow V_{ij}\in\mathbb{R}^{n}, \tag*{(5)}

where FNN(⋅)\mathcal{F}NN(\cdot) denotes the approximation function of the FNN model, WW is the weight matrices, and bb is the bias vector; dN=R2−1d_N=R^2-1 denotes the length of the input feature vector NijN_{ij}, and RR is the radius of the 2D data template (see Figure 2b); and nn is the length of the probability vector VijV_{ij}, corresponding to the number of constituent phases within the composite microstructure.

The training process of FNN is to minimise the discrepancy between the target probability vector VijV_{ij} and the predicted probability vector V^ij\widehat{V}_{ij} by iteratively adjusting the weight matrices WW and bias vector bb. This optimisation problem can be mathematically formulated as:

arg⁡min⁡W,b L(Vij,Nij;W,b)=1m∑(Vij−V^ij)2+λ∥W∥F(6)\underset{W,b}{\arg\min}\ \mathcal{L}(V_{ij},N_{ij};W,b)=\frac{1}{m}\sum(V_{ij}-\widehat{V}_{ij})^2+\lambda\lVert W\rVert_F \tag*{(6)}
=1m∑[Vij−FNN(Nij;W,b)]2+λ∥W∥F,(7)=\frac{1}{m}\sum\left[V_{ij}-\mathcal{F}NN(N_{ij};W,b)\right]^2+\lambda\lVert W\rVert_F, \tag*{(7)}

where L(Vij,Nij;W,b)\mathcal{L}(V_{ij},N_{ij};\mathbf{W},\mathbf{b}) is the loss function, mm is the number of training data points, λ\lambda is the weight regulation constant, and the Frobenius norm ∥W∥F\lVert\mathbf{W}\rVert_{\mathrm{F}} the weight matrix, representing weight decay to encourage smoother network responses and reduce overfitting. The parameter settings for training the SENN are given in Appendix § 7.

Once the FNN model is adequately trained, it can predict the response V^ij\widehat{V}_{ij} for any given input feature NijN_{ij}. Since the sigmoid function (i.e., σ(x)=1/(1+e−x)\sigma(x)=1/(1+e^{-x})) is used as the activation function in the final layer, each element in the output probability vector V^ij\widehat{\mathbf{V}}_{ij} will range between 0 and 1. Essentially, the predicted probability vector V^ij\widehat{\mathbf{V}}_{ij} from the FNN model approximates the CPDF P(Xij∣Nij)P(X_{ij}|N_{ij}) that characterises the morphological statistics of the 2D training image X\mathbf{X}. This relationship is expressed as:

P(Xij∣Nij)={P(Xij=0∣Nij)=P(Vij=[0,0,1]∣Nij)≈V^ij(3)P(Xij=1∣Nij)=P(Vij=[0,1,0]∣Nij)≈V^ij(2)P(Xij=2∣Nij)=P(Vij=[1,0,0]∣Nij)≈V^ij(1)=V^ij=FNN(Nij;W,b),(8)P(X_{ij}\mid N_{ij}) = \begin{cases} P(X_{ij}=0\mid N_{ij})=P(V_{ij}=[0,0,1]\mid N_{ij})\approx\widehat{V}_{ij}(3)\\ P(X_{ij}=1\mid N_{ij})=P(V_{ij}=[0,1,0]\mid N_{ij})\approx\widehat{V}_{ij}(2)\\ P(X_{ij}=2\mid N_{ij})=P(V_{ij}=[1,0,0]\mid N_{ij})\approx\widehat{V}_{ij}(1) \end{cases} = \widehat{\mathbf{V}}_{ij}=\mathcal{FNN}(N_{ij};\mathbf{W},\mathbf{b}), \tag*{(8)}

where V^ij(t)\widehat{V}_{ij}(t) denotes the tt-th element of the output probability vector V^ij\widehat{\mathbf{V}}_{ij} (here, 1≤t≤31\le t\le3, as three constituent phases are considered in this illustrative microstructure). Therefore, the fitted FNN model, which maps data events (Xij,Nij)(X_{ij},N_{ij}), served as an implicit representation of the CPDF P(Xij∣Nij)P(X_{ij}|N_{ij}), and it is referred to as the Statistics-Encoded Neural Network (SENN) in this work.

Multiphase morphological statistics

The CPDF P(Xij∣Nij)P(X_{ij}|N_{ij}) characterises the local morphological statistics associated with the primary microstructural features in the multi-phase composite. To accurately approximate this CPDF, the data template (refer to Figure 2b) used for data collection must be sufficiently large to encompass the key morphological features of different constituent phases. As illustrated in Figure 2a, the morphological features associated with these constituent phases often differ in size. This variation requires the use of appropriately sized data templates for accurate statistical characterisation. For example, the black phase, with its broader range of features, demands a larger data template than the grey phase to ensure a precise CPDF estimation.

In this work, the radius RR of a data template is determined using the two-point correlation function (TPCF) S2(d)S_{2}(d) [35, 51], which quantifies the probability of finding two points separated by a distance dd within a particular constituent phase. As shown in the leftmost column of Figure 3, the TPCF curves stabilise when the distance dd between the two points exceeds a certain value ll that is called correlation length. This stabilisation indicates that two points can be considered statistically uncorrelated when d≥ld\ge l. Therefore, the correlation length ll extracted from the TPCF curve well defines the correlation range, providing a critical value for the template size RR. Additionally, the correlation length ll can also be computed by fitting the TPCF S2(d)S_{2}(d) to the following exponential function [52]:

Procedure of fitting statistics-encoded neural networks for three phases

Figure 3. The procedure of fitting statistics-encoded neural networks (SENNs) to characterise the morphological statistics corresponding to different phases in the composite microstructure (here, a 3-phase composite microstructure is used for illustration).

S2(d)=(ϕ−ϕ2)exp⁡(−3dl)+ϕ2,(9)S_{2}(d)=(\phi-\phi^{2})\exp\left(-\frac{3d}{l}\right)+\phi^{2}, \tag*{(9)}

where ϕ\phi denotes volume fraction of the constituent phase, and the template size RR is defined as the smallest integer greater than or equal to the correlation length ll.

As illustrated in Figure 3, the TPCFs corresponding to three constituent phases (n=3n=3) are separately derived from the training image to determine the appropriate sizes of three distinct data templates. Using three sets of data events, each collected with a varying data template, three SENN models can be independently fitted to characterise the morphological statistics pertinent to each constituent phase. The CPDF P(Xij∣Nij)P(X_{ij}|N_{ij}), which represents multiphase morphological statistics, is computed by averaging the probabilities obtained from each SENN model, as follows:

P(Xij∣Nij)=1n∑t=0n−1P(Xij∣Nij(t))≈1n∑t=0n−1FNN(t)(Nij(t);W(t),b(t)),(10)P(X_{ij}\mid N_{ij})=\frac{1}{n}\sum_{t=0}^{n-1}P(X_{ij}\mid N_{ij}^{(t)})\approx\frac{1}{n}\sum_{t=0}^{n-1}\mathcal{FNN}^{(t)}(N_{ij}^{(t)};\mathbf{W}^{(t)},\mathbf{b}^{(t)}), \tag*{(10)}

where Nij(t)N_{ij}^{(t)} denotes the neighbouring pixels (i.e., the feature vector) defined by the data template for the phase tt, FNN(t)(⋅)\mathcal{FNN}^{(t)}(\cdot) represents the approximation function of the FNN model for the phase tt, W(t)\mathbf{W}^{(t)} is the weight matrix, b(t)\mathbf{b}^{(t)} denotes the bias vector, and nn is the number of constituent phases (here, nn is set at 3, as the example uses a composite microstructure with three phases).

2D-to-3D conversion of morphological statistics

To address the dimension disparity between 2D morphological statistics P(Xij∣Nij)P\left(X_{ij}\mid N_{ij}\right) and 3D microstructure reconstruction, inferring 3D morphological statistics P(Xijk∣Nijk)P\left(X_{ijk}\mid N_{ijk}\right) from 2D measurements is essential. A specialised morphological integration scheme has been developed here for this purpose, enabling the derivation of 3D CPDF P(Xijk∣Nijk)P\left(X_{ijk}\mid N_{ijk}\right) from 2D morphological statistics, as illustrated in Figure 4. The MRF assumption, defined in Eq. (3), is further simplified to model the phase values of a central voxel XijkX_{ijk} influenced by its neighbouring voxels on three orthogonal planes: xyxy-, yzyz-, and zxzx-planes. This simplification is expressed as:

Computational framework for deriving 3D morphological statistics from 2D morphological statistics

Figure 4. The computational framework of deriving 3D morphological statistics P(Xijk∣Nijk)P(X_{ijk}\mid N_{ijk}) from 2D morphological statistics P(Xij∣Nij(xy))P(X_{ij}\mid N_{ij}^{(xy)}), P(Xjk∣Njk(yz))P(X_{jk}\mid N_{jk}^{(yz)}) and P(Xki∣Nki(zx))P(X_{ki}\mid N_{ki}^{(zx)}) on three orthogonal planes (i.e., xyxy-, yzyz-, and zxzx-planes), where the 2D CPDFs are approximated through fitting the SENN models.

P(Xijk∣X(−ijk))=P(Xijk∣Nijk)=P[Xijk∣(Nijk(xy)∪Nijk(yz)∪Nijk(zx))],(11)P\left(X_{ijk}\mid X^{(-ijk)}\right)=P\left(X_{ijk}\mid N_{ijk}\right)=P\left[X_{ijk}\mid\left(N_{ijk}^{(xy)}\cup N_{ijk}^{(yz)}\cup N_{ijk}^{(zx)}\right)\right], \tag*{(11)}

where Nijk(xy)N_{ijk}^{(xy)}, Nijk(yz)N_{ijk}^{(yz)} and Nijk(zx)N_{ijk}^{(zx)} represent the sets of neighbouring voxels on xyxy-, yzyz- and zxzx-planes, respectively.

For a multiphase composite microstructure, the morphological patterns observed on cross-sectional planes are consistent with those in 3D space. When the range of neighbouring voxels is sufficiently large, a relationship exists between 2D and 3D morphological statistics [17, 18], mathematically expressed as:

P(Xijk∣Nijk)=P(Xijk∣Nijk(xy))=P(Xijk∣Nijk(yz))=P(Xijk∣Nijk(zx)),(12)P\left(X_{ijk}\mid N_{ijk}\right)=P\left(X_{ijk}\mid N_{ijk}^{(xy)}\right)=P\left(X_{ijk}\mid N_{ijk}^{(yz)}\right)=P\left(X_{ijk}\mid N_{ijk}^{(zx)}\right), \tag*{(12)}

where the CPDFs P(Xijk∣Nijk)P\left(X_{ijk}\mid N_{ijk}\right), P(Xijk∣Nijk(xy))P\left(X_{ijk}\mid N_{ijk}^{(xy)}\right), P(Xijk∣Nijk(yz))P\left(X_{ijk}\mid N_{ijk}^{(yz)}\right) and P(Xijk∣Nijk(zx))P\left(X_{ijk}\mid N_{ijk}^{(zx)}\right) statistically describe the primary morphological features in 3D space and on xyxy-, yzyz- and zxzx-planes, respectively. It is important to note that while the neighbourhoods NijkN_{ijk}, Nijk(xy)N_{ijk}^{(xy)}, Nijk(yz)N_{ijk}^{(yz)} and Nijk(zx)N_{ijk}^{(zx)} can differ significantly, the relationships described by this equation remain valid. Using the morphology-consistency assumption from Eq. (12), the 3D CPDF P(Xijk∣Nijk)P\left(X_{ijk}\mid N_{ijk}\right) can derived from the 2D CPDFs on three orthogonal planes by averaging:

P(Xijk∣Nijk)=13[P(Xijk∣Nijk(xy))+P(Xijk∣Nijk(yz))+P(Xijk∣Nijk(zx))],(13)P(X_{ijk}\mid N_{ijk}) = \frac{1}{3}\left[P(X_{ijk}\mid N_{ijk}^{(xy)}) + P(X_{ijk}\mid N_{ijk}^{(yz)}) + P(X_{ijk}\mid N_{ijk}^{(zx)})\right], \tag*{(13)}

This morphology integration scheme treats the morphological components across the three principal planes equally, so the phase value of a voxel is determined by the information from its neighbouring voxels on all three orthogonal planes.

By utilising the 2D exemplars on three orthogonal planes, three SENN models can be separately trained to accurately characterise the 2D morphological statistics, as illustrated in Figure 4. As explained by Eq. (8), these SENN models implicitly represent the 2D CPDFs P(Xij(xy)∣Nij(xy))P(X_{ij}^{(xy)}\mid N_{ij}^{(xy)}), P(Xjk(yz)∣Njk(yz))P(X_{jk}^{(yz)}\mid N_{jk}^{(yz)}) and P(Xki(zx)∣Nki(zx))P(X_{ki}^{(zx)}\mid N_{ki}^{(zx)}) respectively. These SENN models enable the accurate approximation of the 3D CPDF P(Xijk∣Nijk)P(X_{ijk}\mid N_{ijk}) using Eq. (13):

P(Xijk∣Nijk)≈13[FN(xy)(Nij(xy);W(xy),b(xy))+FN(yz)(Njk(yz);W(yz),b(yz))+FN(zx)(Nki(zx);W(zx),b(zx))],(14)P(X_{ijk}\mid N_{ijk}) \approx\frac{1}{3}\left[\mathcal{F}\mathcal{N}^{(xy)}(N_{ij}^{(xy)};\mathbf{W}^{(xy)},\mathbf{b}^{(xy)})+\mathcal{F}\mathcal{N}^{(yz)}(N_{jk}^{(yz)};\mathbf{W}^{(yz)},\mathbf{b}^{(yz)})+\mathcal{F}\mathcal{N}^{(zx)}(N_{ki}^{(zx)};\mathbf{W}^{(zx)},\mathbf{b}^{(zx)})\right], \tag*{(14)}

where FN(xy)(⋅)\mathcal{F}\mathcal{N}^{(xy)}(\cdot), FN(yz)(⋅)\mathcal{F}\mathcal{N}^{(yz)}(\cdot) and FN(zx)(⋅)\mathcal{F}\mathcal{N}^{(zx)}(\cdot) represent the approximation functions of the SENN models for statistically characterising the 2D exemplars on the xyxy-, yzyz-, and zxzx-planes, respectively. The main steps of the SENN-based statistical characterisation for multiphase microstructures are summarised in Algorithm 1.

In summary, for isotropic multiphase microstructures, the morphological characteristics are equivalent in all directions. Consequently, cross-sectional images in a single direction are sufficient to serve as 2D exemplars for training the SENN model. In contrast, anisotropic microstructures exhibit significant variation in morphological patterns across different directions. To capture this anisotropy, 2D slices from three orthogonal planes are required, necessitating the training of three separate SENN models: FN(xy)(⋅)\mathcal{F}\mathcal{N}^{(xy)}(\cdot), FN(yz)(⋅)\mathcal{F}\mathcal{N}^{(yz)}(\cdot) and FN(zx)(⋅)\mathcal{F}\mathcal{N}^{(zx)}(\cdot) should be trained separately. Additionally, incorporating 2D exemplars from diagonal planes into the morphology integration scheme can further enhance the characterisation of 3D microstructures, as illustrated in Figure 5. In this case, the 3D morphological statistics P(Xijk∣Nijk)P(X_{ijk}\mid N_{ijk}) can be estimated as a weighted average of the 2D CPDFs from nine different planes:

Schematic representations of 2D measurements on principal and diagonal planes

Figure 5. The 2D measurements on both principal and diagonal planes can be incorporated into the 3D microstructure characterisation.

P(Xijk∣Nijk)≈19⋅[FNN(xy)(Nij(xy))+FNN(yz)(Njk(yz))+FNN(zx)(Nki(zx))+FNN(xy−45∘)(Nij(xy−45∘))+FNN(yz−45∘)(Njk(yz−45∘))+FNN(zx−45∘)(Nki(zx−45∘))+FNN(xy−135∘)(Nij(xy−135∘))+FNN(yz−135∘)(Njk(yz−135∘))+FNN(zx−135∘)(Nki(zx−135∘))],(15)P\left(X_{ijk}\mid N_{ijk}\right) \approx\frac{1}{9}\cdot\left[\mathcal{FNN}^{(xy)}\left(N_{ij}^{(xy)}\right)+\mathcal{FNN}^{(yz)}\left(N_{jk}^{(yz)}\right)+\mathcal{FNN}^{(zx)}\left(N_{ki}^{(zx)}\right) +\mathcal{FNN}^{(xy-45^\circ)}\left(N_{ij}^{(xy-45^\circ)}\right)+\mathcal{FNN}^{(yz-45^\circ)}\left(N_{jk}^{(yz-45^\circ)}\right)+\mathcal{FNN}^{(zx-45^\circ)}\left(N_{ki}^{(zx-45^\circ)}\right) +\mathcal{FNN}^{(xy-135^\circ)}\left(N_{ij}^{(xy-135^\circ)}\right)+\mathcal{FNN}^{(yz-135^\circ)}\left(N_{jk}^{(yz-135^\circ)}\right)+\mathcal{FNN}^{(zx-135^\circ)}\left(N_{ki}^{(zx-135^\circ)}\right)\right], \tag*{(15)}

where FNN(xy−45∘)(⋅)\mathcal{FNN}^{(xy-45^\circ)}(\cdot), FNN(yz−45∘)(⋅)\mathcal{FNN}^{(yz-45^\circ)}(\cdot), FNN(zx−45∘)(⋅)\mathcal{FNN}^{(zx-45^\circ)}(\cdot), FNN(xy−135∘)(⋅)\mathcal{FNN}^{(xy-135^\circ)}(\cdot), FNN(yz−135∘)(⋅)\mathcal{FNN}^{(yz-135^\circ)}(\cdot) and FNN(zx−135∘)(⋅)\mathcal{FNN}^{(zx-135^\circ)}(\cdot) represent the approximation functions of the SENN models for statistically characterizing the 2D exemplars on the xyxy-45∘^\circ-, yzyz-45∘^\circ-, zxzx-45∘^\circ-, xyxy-135∘^\circ-, yzyz-135∘^\circ-, and zxzx-135∘^\circ-planes, respectively.

Algorithm 1: Statistical characterisation of multiphase composite microstructures by fitting SENN models

Training images: The available 2D exemplars on xyxy-, yzyz-, and zxzx-planes: X(xy)\mathbf{X}^{(xy)}, X(yz)\mathbf{X}^{(yz)} and X(zx)\mathbf{X}^{(zx)}.

(1) Determine the sizes of 2D data templates:

Compute two-point correlation functions (TPCFs) for all constituent phases in the selected 2D exemplars;

Measure the correlation lengths ll from TPCF curves to determine the sizes of 2D data templates for different constituent phases (using the Eq. (9)): R0(xy)R_{0}^{(xy)}, R1(xy)R_{1}^{(xy)} and R2(xy)R_{2}^{(xy)}; R0(yz)R_{0}^{(yz)}, R1(yz)R_{1}^{(yz)} and R2(yz)R_{2}^{(yz)}; R0(zx)R_{0}^{(zx)}, R1(zx)R_{1}^{(zx)} and R2(zx)R_{2}^{(zx)}.

Return: A set of 2D data templates: DT0(xy)DT_{0}^{(xy)}, DT1(xy)DT_{1}^{(xy)} and DT2(xy)DT_{2}^{(xy)}; DT0(yz)DT_{0}^{(yz)}, DT1(yz)DT_{1}^{(yz)} and DT2(yz)DT_{2}^{(yz)}; DT0(zx)DT_{0}^{(zx)}, DT1(zx)DT_{1}^{(zx)} and DT2(zx)DT_{2}^{(zx)}.

(2) Training data preparation:

for R0(xy)+1≤i≤H(xy)−R0(xy)R_{0}^{(xy)} + 1 \le i \le H^{(xy)} - R_{0}^{(xy)} do

  for R0(xy)+1≤j≤W(xy)−R0(xy)R_{0}^{(xy)} + 1 \le j \le W^{(xy)} - R_{0}^{(xy)} do

    Collect the data event (Xij(xy:0),Nij(xy:0))(X_{ij}^{(xy: 0)}, \mathbf{N}_{ij}^{(xy: 0)}) from X(xy)\mathbf{X}^{(xy)} using the 2D data template DT0(xy)DT_{0}^{(xy)};

    Convert the central pixel value Xij(xy:0)X_{ij}^{(xy: 0)} into a probability vector Vij(xy:0)\mathbf{V}_{ij}^{(xy: 0)};

    Move the 2D data template DT0(xy)DT_{0}^{(xy)} to the next position in the raster scan order: j=j+1j = j + 1;

  end

  Move the 2D data template DT0(xy)DT_{0}^{(xy)} to the next line: i=i+1i = i + 1;

end

Store the training dataset (Vij(xy:0),Nij(xy:0))(\mathbf{V}_{ij}^{(xy: 0)}, \mathbf{N}_{ij}^{(xy: 0)}) for the constituent phase 0 in 2D exemplar X(xy)\mathbf{X}^{(xy)};

Repeat the above procedures to collect training data for different constituent phases in X(xy)\mathbf{X}^{(xy)}, X(yz)\mathbf{X}^{(yz)} and X(zx)\mathbf{X}^{(zx)};

Return: The training datasets: (Vij(xy:0),Nij(xy:0))(\mathbf{V}_{ij}^{(xy: 0)}, \mathbf{N}_{ij}^{(xy: 0)}), (Vij(xy:1),Nij(xy:1))(\mathbf{V}_{ij}^{(xy: 1)}, \mathbf{N}_{ij}^{(xy: 1)}) and (Vij(xy:2),Nij(xy:2))(\mathbf{V}_{ij}^{(xy: 2)}, \mathbf{N}_{ij}^{(xy: 2)}); (Vjk(yz:0),Njk(yz:0))(\mathbf{V}_{jk}^{(yz: 0)}, \mathbf{N}_{jk}^{(yz: 0)}), (Vjk(yz:1),Njk(yz:1))(\mathbf{V}_{jk}^{(yz: 1)}, \mathbf{N}_{jk}^{(yz: 1)}) and (Vjk(yz:2),Njk(yz:2))(\mathbf{V}_{jk}^{(yz: 2)}, \mathbf{N}_{jk}^{(yz: 2)}); (Vki(zx:0),Nki(zx:0))(\mathbf{V}_{ki}^{(zx: 0)}, \mathbf{N}_{ki}^{(zx: 0)}), (Vki(zx:1),Nki(zx:1))(\mathbf{V}_{ki}^{(zx: 1)}, \mathbf{N}_{ki}^{(zx: 1)}) and (Vki(zx:2),Nki(zx:2))(\mathbf{V}_{ki}^{(zx: 2)}, \mathbf{N}_{ki}^{(zx: 2)}).

(3) Training SENN models for statistical characterisation:

Use the training dataset (Vij(xy:0),Nij(xy:0))(\mathbf{V}_{ij}^{(xy: 0)}, \mathbf{N}_{ij}^{(xy: 0)}) to fit a SENN model FNN(xy:0)\mathcal{FNN}^{(xy: 0)} for approximating P(Xij(xy:0)∣Nij(xy:0))P(X_{ij}^{(xy: 0)} \mid\mathbf{N}_{ij}^{(xy: 0)});

Repeat the above procedure to fit SENN models using the prepared training datasets for approximating the following CPDFs: P(Xij(xy:1)∣Nij(xy:1))P(X_{ij}^{(xy: 1)} \mid\mathbf{N}_{ij}^{(xy: 1)}) and P(Xij(xy:2)∣Nij(xy:2))P(X_{ij}^{(xy: 2)} \mid\mathbf{N}_{ij}^{(xy: 2)}); P(Xjk(yz:0)∣Njk(yz:0))P(X_{jk}^{(yz: 0)} \mid\mathbf{N}_{jk}^{(yz: 0)}), P(Xjk(yz:1)∣Njk(yz:1))P(X_{jk}^{(yz: 1)} \mid\mathbf{N}_{jk}^{(yz: 1)}) and P(Xjk(yz:2)∣Njk(yz:2))P(X_{jk}^{(yz: 2)} \mid\mathbf{N}_{jk}^{(yz: 2)}); P(Xki(zx:0)∣Nki(zx:0))P(X_{ki}^{(zx: 0)} \mid\mathbf{N}_{ki}^{(zx: 0)}), P(Xki(zx:1)∣Nki(zx:1))P(X_{ki}^{(zx: 1)} \mid\mathbf{N}_{ki}^{(zx: 1)}) and P(Xki(zx:2)∣Nki(zx:2))P(X_{ki}^{(zx: 2)} \mid\mathbf{N}_{ki}^{(zx: 2)});

Derive the following 2D CPDFs using Eq. (10) for statistical characterisation of the 2D exemplars X(xy)\mathbf{X}^{(xy)}, X(yz)\mathbf{X}^{(yz)} and X(zx)\mathbf{X}^{(zx)}, respectively: P(Xij(xy)∣Nij(xy))P(X_{ij}^{(xy)} \mid\mathbf{N}_{ij}^{(xy)}), P(Xjk(yz)∣Njk(yz))P(X_{jk}^{(yz)} \mid\mathbf{N}_{jk}^{(yz)}) and P(Xki(zx)∣Nki(zx))P(X_{ki}^{(zx)} \mid\mathbf{N}_{ki}^{(zx)});

Infer the 3D CPDF P(Xijk∣Nijk)P(X_{ijk} \mid\mathbf{N}_{ijk}) from the 2D morphological statistics on three orthogonal planes using Eq. (14).

Return: A set of SENN models representing the 2D morphological statistics: FNN(xy:0)\mathcal{FNN}^{(xy: 0)}, FNN(xy:1)\mathcal{FNN}^{(xy: 1)} and FNN(xy:2)\mathcal{FNN}^{(xy: 2)}; FNN(yz:0)\mathcal{FNN}^{(yz: 0)}, FNN(yz:1)\mathcal{FNN}^{(yz: 1)} and FNN(yz:2)\mathcal{FNN}^{(yz: 2)}; FNN(zx:0)\mathcal{FNN}^{(zx: 0)}, FNN(zx:1)\mathcal{FNN}^{(zx: 1)} and FNN(zx:2)\mathcal{FNN}^{(zx: 2)}.

(Note: H(xy)H^{(xy)} and W(xy)W^{(xy)} represent the height and width of the 2D exemplar X(xy)\mathbf{X}^{(xy)}, respectively.)

Stochastic microstructure reconstruction

In the preceding section, a statistical characterisation approach has been developed to derive 3D morphological statistics P(Xijk∣Nijk)P(X_{ijk}|N_{ijk}) from 2D measurements, where 2D morphological statistics are approximated by training SENN models on representative 2D exemplars. The focus now shifts to leveraging this machine learning-based characterisation to synthesise 3D microstructure samples that are statistically equivalent to the 2D exemplars. In this section, a stochastic 2D-to-3D reconstruction approach is introduced for generating 3D multiphase microstructures, which utilises the 2D morphological statistics embedded in the pretrained SENN models. A Gibbs sampler is employed to iteratively generate voxel values, progressively updating the 3D realisation to converge toward the target probability distribution.

Voxel generation via probability sampling

Generating observations of multivariate random variables through direct sampling can be challenging without complete knowledge of the full joint PDF. However, Gibbs sampling [53, 54] provides an effective solution by leveraging conditional PDFs to produce a sequence of observations that approximate the target multivariate probability distribution. The 3D CPDF P(Xijk∣Nijk)P(X_{ijk}|N_{ijk}), as defined in Eq. (14), characterises the local morphology of multiphase microstructures. This CPDF, efficiently derived from pretrained SENN models, supplies the conditional statistics necessary to construct a 3D Markov random field (MRF). The Gibbs sampler iteratively updates this 3D MRF using the 3D CPDF, guiding it toward the underlying full joint PDF. After sufficient iterations, the 3D MRF converges to a stationary state that closely approximates the target full joint PDF, regardless of the initial conditions. This equilibrium state of the 3D MRF represents a reconstructed 3D microstructure sample that is statistically equivalent to the 2D exemplars.

To achieve the objective of 2D-to-3D microstructure reconstruction, a specially designed 3D data template is employed to generate voxels using Gibbs sampling, as illustrated in Figure 6. This 3D data template integrates three 2D data templates on orthogonal planes, which were previously utilised for statistically characterising the 2D exemplars (as shown in Figure 2). The geometric consistency between the 3D and 2D data templates ensures seamless application of the derived 3D CPDF P(Xijk∣Nijk)P(X_{ijk}|N_{ijk}), as defined in Eq. (14), during the microstructure reconstruction process.

Gibbs sampling [53, 54] iteratively generates and updates voxel values in the initial 3D MRF Y0Y_0, using the CPDF P(Xijk∣Nijk)P(X_{ijk}|N_{ijk}). This process ultimately converges to the full joint PDF P(Y)P(Y). The 3D MRF is characterised by three key factors: the initial probability P(Y0)P(Y_0), the transition probability P(Xijk∣Nijk)P(X_{ijk}|N_{ijk}), and the stationary probability P(Y)P(Y). The iterative process of Gibbs sampling is expressed as:

P(Y0)∏P(Yijk∣Nijk)⟶P(Y),(16)P(Y_0)\prod P(Y_{ijk}|N_{ijk}) \longrightarrow P(Y), \tag*{(16)}

where YY represents the equilibrium state of the 3D MRF, which can be regarded as a reconstructed 3D microstructure sample statistically equivalent to the 2D exemplars. This approach effectively leverages plane-wise neighbourhood information to iteratively reconstruct 3D microstructures, bridging the gap between 2D statistical characterisation and 3D microstructure reconstruction.

In this methodology, a 3D Gaussian random field (white noise) is initially generated and then segmented into a discrete field with nn distinct voxel values, as defined by the indicator function in Eq. (2). The obtained discrete random field serves as the initial assumption Y0Y_0 for stochastic microstructure reconstruction. Subsequently, the predefined 3D data template (as illustrated in Figure 6) is then employed to scan Y0Y_0 voxel by voxel in the raster scanning order (i.e., left-to-right and top-to-bottom). The collected neighbouring voxels NijkN_{ijk} in each scanning step are decomposed into feature vectors Nijk(xy)N_{ijk}^{(xy)}, Nijk(yz)N_{ijk}^{(yz)} and Nijk(zx)N_{ijk}^{(zx)} on xyxy-, yzyz- and zxzx-planes, respectively. These feature vectors are input into the pretrained SENN models FNN(xy)\mathcal{F}NN^{(xy)}, FNN(yz)\mathcal{F}NN^{(yz)} and FNN(zx)\mathcal{F}NN^{(zx)} to rapidly approximate the 3D CPDF P(Yijk∣Nijk)P(Y_{ijk}|N_{ijk}), as elucidated by Eq. (14). Given this probability P(Yijk∣Nijk)P(Y_{ijk}|N_{ijk}), Gibbs sampler can assign a new phase value to the current (central) voxel YijkY_{ijk} through probability sampling. This newly generated voxel is then integrated into the feature vectors to update subsequent adjacent voxels within the predefined neighbourhood ranges. By iterating these steps, the morphological patterns of the 3D random field gradually converge toward those of the 2D exemplars. Once equilibrium is achieved, a 3D microstructure sample statistically equivalent to the 2D exemplars is reconstructed, as illustrated in Figure 7c.

Graphical illustration of probability sampling of voxel values

Figure 7. Graphical illustration of probability sampling of voxel values: (a) A 3D random field is used as the initial assumption Y0Y_0; (b) The predefined 3D data template is utilised to collect neighbouring voxels NijkN_{ijk} and generate/update the value of the central voxel YijkY_{ijk} via Gibbs sampling; and (c) A reconstructed 3D microstructure sample (i.e., a realization of 3D MRF) that is statistically equivalent to the 2D examples.

Algorithm 2: Stochastic reconstruction of 3D multiphase microstructures using SENN-based 2D characterisation

Data: A 3D lattice grid of size I×J×KI \times J \times K.

Initialization: Assign a discrete random field to the 3D lattice grid and use it as the initial assumption Y0Y_0.

SENN models: The pretrained SENN models statistically characterising the 2D exemplars: FNN(xy)\mathcal{F}NN^{(xy)}, FNN(yz)\mathcal{F}NN^{(yz)} and FNN(zx)\mathcal{F}NN^{(zx)}.

Data templates: The 3D data template DTDT with the predefined size.

Add a periodic or reflective boundary to Y0Y_0;

while the convergence criterion is not satisfied do

    for each voxel variable YijkY_{ijk} in Y0Y_0 do

        Extract the neighbouring data NijkN_{ijk} from Y0Y_0 using the 3D data template DTDT;

        Rearrange and split NijkN_{ijk} into three feature vectors Nijk(xy)N_{ijk}^{(xy)}, Nijk(yz)N_{ijk}^{(yz)} and Nijk(zx)N_{ijk}^{(zx)} on xyxy-, yzyz- and zxzx-planes, respectively;

        Input the above feature vectors into corresponding SENN models FNN(xy)\mathcal{F}NN^{(xy)}, FNN(yz)\mathcal{F}NN^{(yz)} and FNN(zx)\mathcal{F}NN^{(zx)};

        Derive the 3D CPDF P(Yijk∣Nijk)P(Y_{ijk}|N_{ijk}) from 2D morphological statistics through Eq. (14), given by:

P(Xijk∣Nijk)≈13[FNN(xy)(Nij(xy);W(xy),b(xy))+FNN(yz)(Njk(yz);W(yz),b(yz))+FNN(zx)(Nki(zx);W(zx),b(zx))];P(X_{ijk}|N_{ijk}) \approx\frac{1}{3}\left[\mathcal{F}NN^{(xy)}(N_{ij}^{(xy)};W^{(xy)},b^{(xy)})+\mathcal{F}NN^{(yz)}(N_{jk}^{(yz)};W^{(yz)},b^{(yz)})+\mathcal{F}NN^{(zx)}(N_{ki}^{(zx)};W^{(zx)},b^{(zx)})\right];

        Generate a new voxel value Yijk(new)Y_{ijk}^{(\mathrm{new})} through probability sampling from the 3D CPDF P(Yijk∣Nijk)P(Y_{ijk}|N_{ijk});

        Update the voxel value YijkY_{ijk} using the newly generated Yijk(new)Y_{ijk}^{(\mathrm{new})};

    end

    Reset the periodic or reflective boundary condition for YnewY_{\mathrm{new}}.

end

Remove the periodic or reflective boundary condition for YnewY_{\mathrm{new}}.

Return: A reconstructed 3D microstructure sample YnewY_{\mathrm{new}} that is statistically equivalent to the 2D exemplars.

(A constructed Markov random field YnewY_{\mathrm{new}} that closely approximates the target full joint PDF P(Y)P(Y).)

The primary workflow of this SENN-based framework for 2D-to-3D reconstruction of multiphase microstructures are summarised in Algorithm 2. During the reconstruction process, the border voxels lack sufficient neighbourhood information to form feature vectors for the pretrained SENN models, thus preventing their voxel values from being updated. To address this issue, a periodic or reflective boundary condition can be incorporated into the intermediate constructed MRF, transforming the original "border voxels" into inner voxels that can be updated during the construction process. The thickness of the periodic or reflective boundary should align with the dimension of the 3D data template used for microstructure reconstruction. Additionally, it is essential to reset the periodic or reflective boundary after each round of Gibbs sampling until stability is achieved. Further details on periodic or reflective boundaries can be found in relevant references [26, 35].

Parallel voxel generation

Generating voxels sequentially can make the microstructure reconstruction process highly time-consuming, as representative elementary volumes (REVs) of digital microstructures often contain millions of voxels. Leveraging the MRF assumption defined in Eq. (11), distant voxels can be treated as independent. This allows for simultaneous generation of voxels beyond each other’s neighbourhood ranges using independent Gibbs samplers. As shown in Figure 8, the initial 3D random field can be divided into multiple subdomains, enabling parallel voxel generation and updating within these subdomains. This parallel approach significantly can accelerate the microstructure reconstruction process, especially as the number of subdomains increases. However, achieving large-scale parallelism, where thousands of voxels are generated concurrently, requires substantial data storage capacity. Consequently, implementing a parallel voxel generation strategy necessitates balancing computational speed with memory requirements to ensure efficiency.

Initial 3D random field divided into multiple subdomains

Figure 8. The initial 3D random field is divided into multiple subdomains, where voxels located beyond each other’s neighbourhood ranges can be parallelly generated using independent Gibbs samplers.

Multi-level reconstruction

As detailed in § 2, the data template must be sufficiently large to capture the primary microstructural characteristics of multiphase composites. While effective for random microstructures with short-range correlations [18, 35], capturing long-range features in multiphase microstructures poses challenges. Simply enlarging the data template can drastically increase memory requirements for storing training data, complicating the SENN model training process for CPDF approximation. Furthermore, prioritising global characteristics may inadvertently dilute critical local features.

To address the limitations of single-level microstructure characterisation, a hierarchical characterisation approach has been developed to statistically measure the local, regional, and global features of multiphase microstructures, enabling the derivation of 3D morphological statistics across multiple length scales. The methodology underlying this hierarchical characterisation is graphically illustrated in Figure 9, with the procedure summarised as follows:

Graphical illustration of hierarchical characterisation and multi-level reconstruction for multiphase composite microstructures

Figure 9. Graphical illustration of hierarchical characterisation and multi-level reconstruction for multiphase composite microstructures.

  • Gaussian image pyramids: The original high-resolution 2D exemplars are transformed into Gaussian image pyramids [55], comprising 2D images at three resolution levels: X(low)\mathbf{X}^{(\mathrm{low})}, X(middle)\mathbf{X}^{(\mathrm{middle})}, and X(high)\mathbf{X}^{(\mathrm{high})}. These pyramids enable the stepwise characterisation of global, regional, and local microstructural features, allowing large-scale metrics to be effectively captured using relatively small data templates.

  • Low-level characterisation: The 2D low-level exemplars, denoted by X(low)\mathbf{X}^{(\mathrm{low})}, are statistically characterised by training SENN models to approximate 2D CPDFs on three orthogonal planes, as detailed in § 2. Based on Eqs. (13) and (14), the 3D morphological statistics P(Xijk(low)∣Nijk(low))P\left(\mathbf{X}_{ijk}^{(\mathrm{low})}\mid\mathbf{N}_{ijk}^{(\mathrm{low})}\right) at the low level can be inferred from the 2D CPDFs on three orthogonal planes, given by:

P(Xijk(low)∣Nijk(low))=P(Xijk(low)∣(Nijk(low:xy)∪Nijk(low:yz)∪Nijk(low:zx)))=13[P(Xijk(low)∣Nijk(low:xy))+P(Xijk(low)∣Nijk(low:yz))+P(Xijk(low)∣Nijk(low:zx))]≈13[FNNxy(low)(Nijk(low:xy))+FNNyz(low)(Nijk(low:yz))+FNNzx(low)(Nijk(low:zx))],(17)\begin{aligned} P\left(\mathbf{X}_{ijk}^{(\mathrm{low})}\mid\mathbf{N}_{ijk}^{(\mathrm{low})}\right) &=P\left(\mathbf{X}_{ijk}^{(\mathrm{low})}\mid\left(\mathbf{N}_{ijk}^{(\mathrm{low}:xy)}\cup\mathbf{N}_{ijk}^{(\mathrm{low}:yz)}\cup\mathbf{N}_{ijk}^{(\mathrm{low}:zx)}\right)\right)\\ &=\frac{1}{3}\left[P\left(\mathbf{X}_{ijk}^{(\mathrm{low})}\mid\mathbf{N}_{ijk}^{(\mathrm{low}:xy)}\right)+P\left(\mathbf{X}_{ijk}^{(\mathrm{low})}\mid\mathbf{N}_{ijk}^{(\mathrm{low}:yz)}\right)+P\left(\mathbf{X}_{ijk}^{(\mathrm{low})}\mid\mathbf{N}_{ijk}^{(\mathrm{low}:zx)}\right)\right]\\ &\approx\frac{1}{3}\left[\mathcal{FNN}_{xy}^{(\mathrm{low})}\left(\mathbf{N}_{ijk}^{(\mathrm{low}:xy)}\right)+\mathcal{FNN}_{yz}^{(\mathrm{low})}\left(\mathbf{N}_{ijk}^{(\mathrm{low}:yz)}\right)+\mathcal{FNN}_{zx}^{(\mathrm{low})}\left(\mathbf{N}_{ijk}^{(\mathrm{low}:zx)}\right)\right], \tag*{(17)} \end{aligned}

where Nijk(low:xy)\mathbf{N}_{ijk}^{(\mathrm{low}:xy)}, Nijk(low:yz)\mathbf{N}_{ijk}^{(\mathrm{low}:yz)} and Nijk(low:zx)\mathbf{N}_{ijk}^{(\mathrm{low}:zx)} represent the neighbouring voxels on xyxy-, yzyz- and zxzx-planes, respectively; and meanwhile, FNNxy(low)(⋅)\mathcal{FNN}_{xy}^{(\mathrm{low})}(\cdot), FNNyz(low)(⋅)\mathcal{FNN}_{yz}^{(\mathrm{low})}(\cdot) and FNNzx(low)(⋅)\mathcal{FNN}_{zx}^{(\mathrm{low})}(\cdot) denote the approximation functions of the SENN models used to statistically characterise the low-level 2D exemplars on the respective planes.

  • Middle-level characterisation: The 2D middle-level exemplars X(middle)\mathbf{X}^{(\mathrm{middle})} are utilised alongside the 2D low-level exemplars X(low)\mathbf{X}^{(\mathrm{low})}, to train another set of SENN models for the statistical characterisation at the middle level. The input feature

vector for SENN comprises voxel values from NijkmiddleN_{ijk}^{\mathrm{middle}}, XijklowX_{ijk}^{\mathrm{low}}, and NijklowN_{ijk}^{\mathrm{low}}, while the output is the probability vector VijkmiddleV_{ijk}^{\mathrm{middle}} corresponding to XijkmiddleX_{ijk}^{\mathrm{middle}}. This operation not only captures statistical information from XmiddleX^{\mathrm{middle}} but also learns correlations between XmiddleX^{\mathrm{middle}} and XlowX^{\mathrm{low}}, such as feature consistency and morphology continuity across adjacent levels. The 3D morphological statistics, P[Xijkmiddle∣(Nijkmiddle∪Xijklow∪Nijklow)]P\left[X_{ijk}^{\mathrm{middle}}\mid\left(N_{ijk}^{\mathrm{middle}}\cup X_{ijk}^{\mathrm{low}}\cup N_{ijk}^{\mathrm{low}}\right)\right], at the middle level can be derived from the 2D CPDFs on the three orthogonal planes as follows:

\begin{aligned} P\left[X_{ijk}^{\mathrm{middle}}\mid\left(N_{ijk}^{\mathrm{middle}}\cup X_{ijk}^{\mathrm{low}}\cup N_{ijk}^{\mathrm{low}}\right)\right] &=\frac{1}{3}\cdot\left\{P\left[X_{ijk}^{\mathrm{middle}}\mid\left(N_{ijk}^{\mathrm{middle}:xy}\cup X_{ijk}^{\mathrm{low}:xy}\cup N_{ijk}^{\mathrm{low}:xy}\right)\right]\\ &\quad+P\left[X_{ijk}^{\mathrm{middle}}\mid\left(N_{ijk}^{\mathrm{middle}:yz}\cup X_{ijk}^{\mathrm{low}:yz}\cup N_{ijk}^{\mathrm{low}:yz}\right)\right]\\ &\quad+P\left[X_{ijk}^{\mathrm{middle}}\mid\left(N_{ijk}^{\mathrm{middle}:zx}\cup X_{ijk}^{\mathrm{low}:zx}\cup N_{ijk}^{\mathrm{low}:zx}\right)\right]\right\}\\ &\approx\frac{1}{3}\cdot\left[\mathcal{F}\mathcal{N}_{xy}^{\mathrm{middle}}\left(N_{ijk}^{\mathrm{middle}:xy}\cup X_{ijk}^{\mathrm{low}}\cup N_{ijk}^{\mathrm{low}:xy}\right)\right.\\ &\quad+\mathcal{F}\mathcal{N}_{yz}^{\mathrm{middle}}\left(N_{ijk}^{\mathrm{middle}:yz}\cup X_{ijk}^{\mathrm{low}}\cup N_{ijk}^{\mathrm{low}:yz}\right)\\ &\quad\left.+\mathcal{F}\mathcal{N}_{zx}^{\mathrm{middle}}\left(N_{ijk}^{\mathrm{middle}:zx}\cup X_{ijk}^{\mathrm{low}}\cup N_{ijk}^{\mathrm{low}:zx}\right)\right], \tag*{(18)} \end{aligned}

where Nijkmiddle:xyN_{ijk}^{\mathrm{middle}:xy}, Nijkmiddle:yzN_{ijk}^{\mathrm{middle}:yz} and Nijkmiddle:zxN_{ijk}^{\mathrm{middle}:zx} represent the neighbouring voxels on xyxy-, yzyz- and zxzx-planes of the middle-level image XmiddleX^{\mathrm{middle}}, respectively; The 2D CPDFs P[Xijkmiddle∣(Nijkmiddle:xy∪Xijklow∪Nijklow:xy)]P\left[X_{ijk}^{\mathrm{middle}}\mid\left(N_{ijk}^{\mathrm{middle}:xy}\cup X_{ijk}^{\mathrm{low}}\cup N_{ijk}^{\mathrm{low}:xy}\right)\right], P[Xijkmiddle∣(Nijkmiddle:yz∪Xijklow∪Nijklow:yz)]P\left[X_{ijk}^{\mathrm{middle}}\mid\left(N_{ijk}^{\mathrm{middle}:yz}\cup X_{ijk}^{\mathrm{low}}\cup N_{ijk}^{\mathrm{low}:yz}\right)\right] and P[Xijkmiddle∣(Nijkmiddle:zx∪Xijklow∪Nijklow:zx)]P\left[X_{ijk}^{\mathrm{middle}}\mid\left(N_{ijk}^{\mathrm{middle}:zx}\cup X_{ijk}^{\mathrm{low}}\cup N_{ijk}^{\mathrm{low}:zx}\right)\right] can be obtained from the pretrained SENN models; FNxymiddle(⋅)\mathcal{F}\mathcal{N}_{xy}^{\mathrm{middle}}(\cdot), FNyzmiddle(⋅)\mathcal{F}\mathcal{N}_{yz}^{\mathrm{middle}}(\cdot) and FNzxmiddle(⋅)\mathcal{F}\mathcal{N}_{zx}^{\mathrm{middle}}(\cdot) denote the approximation functions of the SENN models to statistically characterise the middle-level 2D exemplars on xyxy-, yzyz- and zxzx-planes, respectively.

  • High-level characterisation: The 2D high-level exemplars XhighX^{\mathrm{high}} are combined with the 2D middle-level exemplars

XmiddleX^{\mathrm{middle}} to train a new set of SENN models. The input feature vector for SENN contain the voxel values of NijkhighN_{ijk}^{\mathrm{high}}, XijkmiddleX_{ijk}^{\mathrm{middle}} and NijkmiddleN_{ijk}^{\mathrm{middle}}. This characterisation manner is not only to extract statistical informatics from XhighX^{\mathrm{high}}, but also to learn the correlations between XhighX^{\mathrm{high}} and XmiddleX^{\mathrm{middle}}, thereby preserving the feature consistency and the morphology continuity between adjacent levels. The 3D morphological statistics P[Xijkhigh|(Nijkhigh∪Xijkmiddle∪Nijkmiddle)]P\left[X_{ijk}^{\mathrm{high}}\middle|\left(N_{ijk}^{\mathrm{high}}\cup X_{ijk}^{\mathrm{middle}}\cup N_{ijk}^{\mathrm{middle}}\right)\right] at the high level can be approximated from the 2D CPDFs on three orthogonal planes, as follows:

P[Xijkhigh|(Nijkhigh∪Xijkmiddle∪Nijkmiddle)]=13⋅{P[Xijkhigh|(Nijkhigh:xy∪Xijkmiddle∪Nijkmiddle:xy)]+P[Xijkhigh|(Nijkhigh:yz∪Xijkmiddle∪Nijkmiddle:yz)]+P[Xijkhigh|(Nijkhigh:zx∪Xijkmiddle∪Nijkmiddle:zx)]}≈13[FNNxyhigh(Nijkhigh:xy∪Xijkmiddle∪Nijkmiddle:xy)+FNNyzhigh(Nijkhigh:yz∪Xijkmiddle∪Nijkmiddle:yz)+FNNzxhigh(Nijkhigh:zx∪Xijkmiddle∪Nijkmiddle:zx)],(19)\begin{aligned} P\left[X_{ijk}^{\mathrm{high}}\middle|\left(N_{ijk}^{\mathrm{high}}\cup X_{ijk}^{\mathrm{middle}}\cup N_{ijk}^{\mathrm{middle}}\right)\right] &=\frac{1}{3}\cdot\left\{P\left[X_{ijk}^{\mathrm{high}}\middle|\left(N_{ijk}^{\mathrm{high:xy}}\cup X_{ijk}^{\mathrm{middle}}\cup N_{ijk}^{\mathrm{middle:xy}}\right)\right]\right.\\ &\quad+P\left[X_{ijk}^{\mathrm{high}}\middle|\left(N_{ijk}^{\mathrm{high:yz}}\cup X_{ijk}^{\mathrm{middle}}\cup N_{ijk}^{\mathrm{middle:yz}}\right)\right]\\ &\quad\left.+P\left[X_{ijk}^{\mathrm{high}}\middle|\left(N_{ijk}^{\mathrm{high:zx}}\cup X_{ijk}^{\mathrm{middle}}\cup N_{ijk}^{\mathrm{middle:zx}}\right)\right]\right\}\\ &\approx\frac{1}{3}\left[\mathcal{F}NN_{xy}^{\mathrm{high}}\left(N_{ijk}^{\mathrm{high:xy}}\cup X_{ijk}^{\mathrm{middle}}\cup N_{ijk}^{\mathrm{middle:xy}}\right)\right.\\ &\quad+\mathcal{F}NN_{yz}^{\mathrm{high}}\left(N_{ijk}^{\mathrm{high:yz}}\cup X_{ijk}^{\mathrm{middle}}\cup N_{ijk}^{\mathrm{middle:yz}}\right)\\ &\quad\left.+\mathcal{F}NN_{zx}^{\mathrm{high}}\left(N_{ijk}^{\mathrm{high:zx}}\cup X_{ijk}^{\mathrm{middle}}\cup N_{ijk}^{\mathrm{middle:zx}}\right)\right], \tag*{(19)} \end{aligned}

where Nijkhigh:xyN_{ijk}^{\mathrm{high:xy}}, Nijkhigh:yzN_{ijk}^{\mathrm{high:yz}} and Nijkhigh:zxN_{ijk}^{\mathrm{high:zx}} represent the neighbouring voxels on xyxy-, yzyz- and zxzx-planes of the high-level images XhighX^{\mathrm{high}}, respectively; The 2D CPDFs P[Xijkhigh|(Nijkhigh:xy∪Xijkmiddle∪Nijkmiddle:xy)]P\left[X_{ijk}^{\mathrm{high}}\middle|\left(N_{ijk}^{\mathrm{high:xy}}\cup X_{ijk}^{\mathrm{middle}}\cup N_{ijk}^{\mathrm{middle:xy}}\right)\right], P[Xijkhigh|(Nijkhigh:yz∪Xijkmiddle∪Nijkmiddle:yz)]P\left[X_{ijk}^{\mathrm{high}}\middle|\left(N_{ijk}^{\mathrm{high:yz}}\cup X_{ijk}^{\mathrm{middle}}\cup N_{ijk}^{\mathrm{middle:yz}}\right)\right] and P[Xijkhigh|(Nijkhigh:zx∪Xijkmiddle∪Nijkmiddle:zx)]P\left[X_{ijk}^{\mathrm{high}}\middle|\left(N_{ijk}^{\mathrm{high:zx}}\cup X_{ijk}^{\mathrm{middle}}\cup N_{ijk}^{\mathrm{middle:zx}}\right)\right] can be approximated from the pretrained SENN models; and FNNxyhigh(⋅)\mathcal{F}NN_{xy}^{\mathrm{high}}(\cdot), FNNyzhigh(⋅)\mathcal{F}NN_{yz}^{\mathrm{high}}(\cdot) and FNNzxhigh(⋅)\mathcal{F}NN_{zx}^{\mathrm{high}}(\cdot) denote the approximation functions of the SENN models for statistically characterizing the high-level 2D exemplars on xyxy-, yzyz- and zxzx-planes, respectively.

In order to fully utilise the hierarchical SENN-based characterisation results, a multi-level reconstruction approach has thus been specially developed to generate statistically equivalent 3D microstructure samples for multiphase composites, as graphically illustrated in Figure 9. During microstructure reconstruction process, the pretrained SENN models serve as the dictionaries of conditional probabilities, for generating/updating voxels through Gibbs sampling. In the low-level reconstruction step, long-distance features present in the 2D exemplars can be efficiently created and preserved in the 3D reconstructed microstructure sample. Conditional on the low-level reconstruction result, the middle-level reconstruction step generates and incorporates regional microstructural characteristics while ensuring morphological consistency and continuity with the low-level reconstruction. As for high-level construction, local microstructural details are seamlessly integrated into the 3D reconstructed sample to make it morphologically similar to real microstructures. This innovative multi-level reconstruction approach reproduces global, regional, and local morphological patterns separately while ensuring seamless fusion across different length scales. By leveraging the interrelation between the 3D CPDFs at each level, it achieves a cohesive, statistically accurate representation of multiphase microstructures through Gibbs sampling.

Convergence criterion

As detailed in Algorithm 2, Gibbs sampling iterates until the reconstructed microstructure sample reaches a steady state. Convergence is achieved when the difference between consecutive iterations falls below a specified tolerance. The convergence criterion is given by:

εi=∥Yi−Yi−1∥2∥Yi−1∥2≤η,(20)\varepsilon_i=\frac{\left\|\mathbf{Y}_i-\mathbf{Y}_{i-1}\right\|_2}{\left\|\mathbf{Y}_{i-1}\right\|_2}\leq\eta, \tag*{(20)}

Here, εi\varepsilon_i is the relative L2L_2-norm error, quantifying the difference between the current Yi\mathbf{Y}_i and previous Yi−1\mathbf{Y}_{i-1} samples. The iteration index is represented by ii, and η≪1\eta\ll1 is a user-defined tolerance threshold. When εi\varepsilon_i falls below η\eta, the reconstruction process is considered stable.

Examples and results

To evaluate the effectiveness and robustness of the proposed SENN-based framework for multiphase microstructure reconstruction, it is applied to generate virtual microstructure samples for three representative multiphase materials with distinct morphologies. SENN models are trained to approximate morphological statistics required for microstructural characterisation, using training data gathered from available 2D exemplars. Based on the SENN-based characterisation, a group of thirty 3D microstructure samples are stochastically synthesised, preserving the morphological patterns observed in the 2D training data. Finally, morphological descriptors [6, 18, 51] are extracted from both the reconstructed and reference 3D microstructures to quantitatively assess their morphological similarity and statistical equivalence.

Example 1: Porous silver-based electrode

Porous silver-based electrodes [56, 57] are widely employed in electrochemical applications such as solid oxide fuel cells, oxygen sensors, and batteries, due to their superior electrical conductivity, catalytic performance, and chemical stability. In this study, a porous silver-based electrode is chosen as a representative multiphase composite material to demonstrate the effectiveness of the proposed SENN-based reconstruction approach. The electrode consists of three constituent phases: silver, polytetrafluoroethylene (PTFE), and pore space, as depicted in Figure 10.

Visual comparison of reference and reconstructed 2D/3D microstructures of a porous silver-based electrode

Figure 10. Visual comparison between the reference (top row) and the reconstructed (bottom row) 2D/3D microstructures of a porous silver-based electrode, highlighting its three distinct constituent phases: black regions represent silver, grey regions represent PTFE, and white regions represent pore space.

As illustrated in the upper row of Figure 10, representative 2D exemplars on three orthogonal planes are selected to statistically reconstruct virtual 3D microstructure samples. From these original 2D exemplars, Gaussian image pyramids with three levels are constructed for hierarchical characterisation of the porous microstructure, as detailed in § 3.3. Three sets of SENN models are then trained to approximate the 2D morphological statistics on the orthogonal planes, enabling the derivation of 3D CPDFs across multiple length scales. These 3D CPDFs include P(Xijk(low)∣Nijk(low))P\left(X_{ijk}^{(\mathrm{low})}\mid N_{ijk}^{(\mathrm{low})}\right), P[Xijk(middle)∣(Nijk(middle)∪Xijk(low)∪Nijk(low))]P\left[X_{ijk}^{(\mathrm{middle})}\mid\left(N_{ijk}^{(\mathrm{middle})}\cup X_{ijk}^{(\mathrm{low})}\cup N_{ijk}^{(\mathrm{low})}\right)\right] and P[Xijk(high)∣(Nijk(high)∪Xijk(middle)∪Nijk(middle))]P\left[X_{ijk}^{(\mathrm{high})}\mid\left(N_{ijk}^{(\mathrm{high})}\cup X_{ijk}^{(\mathrm{middle})}\cup N_{ijk}^{(\mathrm{middle})}\right)\right]. This hierarchical characterisation aims to capture the long-range connectivity of pore networks. Besides, parameter settings for training data perpetration and SENN modelling are summarised in Table A1.

Multiphase compositeLevelSize of data templatesLearning rateEpoch numberLayer 1Layer 2Layer 3Layer 4
Silver-based electrodeLowData collection on three planes<br>xy-plane: R0=10R_0=10, R1=8R_1=8, R2=12R_2=12<br>yz-plane: R0=16R_0=16, R1=15R_1=15, R2=15R_2=15<br>zx-plane: R0=12R_0=12, R1=10R_1=10, R2=9R_2=90.000012001286416–
Silver-based electrodeMiddleData collection on three planes<br>xy-plane: R0=12R_0=12, R1=15R_1=15, R2=13R_2=13<br>yz-plane: R0=10R_0=10, R1=12R_1=12, R2=10R_2=10<br>zx-plane: R0=10R_0=10, R1=10R_1=10, R2=8R_2=80.000012001286432–
Silver-based electrodeHighData collection on three planes<br>xy-plane: R0=6R_0=6, R1=16R_1=16, R2=22R_2=22<br>yz-plane: R0=8R_0=8, R1=22R_1=22, R2=21R_2=21<br>zx-plane: R0=8R_0=8, R1=16R_1=16, R2=14R_2=140.000012501286432–
Porous SOFC anodeLowData collection on three planes<br>xy-plane: R0=20R_0=20, R1=18R_1=18, R2=22R_2=22<br>yz-plane: R0=16R_0=16, R1=22R_1=22, R2=22R_2=22<br>zx-plane: R0=18R_0=18, R1=16R_1=16, R2=12R_2=120.000014001286416–
Porous SOFC anodeMiddleData collection on nine planes<br>xy-plane: R0=25R_0=25, R1=14R_1=14, R2=23R_2=23<br>yz-plane: R0=14R_0=14, R1=22R_1=22, R2=20R_2=20<br>zx-plane: R0=20R_0=20, R1=10R_1=10, R2=19R_2=19<br>Other planes: R0=15R_0=15, R1=18R_1=18, R2=20R_2=200.000012001286432–
Porous SOFC anodeHighData collection on nine planes<br>xy-plane: R0=23R_0=23, R1=11R_1=11, R2=21R_2=21<br>yz-plane: R0=10R_0=10, R1=20R_1=20, R2=18R_2=18<br>zx-plane: R0=18R_0=18, R1=8R_1=8, R2=16R_2=16<br>Other planes: R0=14R_0=14, R1=20R_1=20, R2=16R_2=160.000012501286432–
Composite cement pasteLowData collection on three planes<br>xy-plane: R0=20R_0=20, R1=R2=R3=10R_1=R_2=R_3=10<br>yz-plane: R0=20R_0=20, R1=R2=R3=10R_1=R_2=R_3=10<br>zx-plane: R0=20R_0=20, R1=R2=R3=10R_1=R_2=R_3=100.00005250128643216
Composite cement pasteHighData collection on three planes<br>xy-plane: R0=15R_0=15, R1=R2=R3=25R_1=R_2=R_3=25<br>yz-plane: R0=15R_0=15, R1=R2=R3=25R_1=R_2=R_3=25<br>zx-plane: R0=15R_0=15, R1=R2=R3=25R_1=R_2=R_3=250.00005250128643216

Table A1. Parameter settings for multi-level SENN-based characterisation of multiphase microstructures

Using the SENN-approximated morphological statistics, a group of thirty 3D microstructure samples are synthesised, with representative results shown in the lower row of Figure 10. Visual inspection reveals that the reconstructed 3D microstructure is morphologically similar to the 2D exemplars and nearly indistinguishable from the 3D reference microstructure. To quantitatively evaluate the reconstruction quality, a series of morphological descriptors are extracted from both the reconstructed and reference 3D microstructures for comparison. These descriptors include volume fraction, surface area density [18], triple-phase boundary density [58], mean chord length [51], geometrical tortuosity [59], and total fraction of percolation cells (TFOPC) [6], as presented in Figure 11.

Six-panel quantitative comparison of reference and reconstructed 3D microstructures: (a) Volume fraction; (b) Surface area density; (c) Triple phase boundary density; (d) Mean chord length; (e) Geometrical tortuosity; (f) Total fraction of percolation cells

Figure 11. Quantitative comparison between the reference and reconstructed 3D microstructures by evaluating morphological descriptors.

The first four descriptors (Figure 11a–Figure 11d) are calculated for all three phases of the porous electrode, while the final two descriptors (Figure 11e–Figure 11f) are specific to the pore space to quantify its long-distance connectivity. Descriptor values or curves derived from the thirty reconstructed microstructures exhibit slight variations around those obtained from the reference 3D microstructure. This variation arises from probabilistic voxel sampling during the statistical reconstruction process. Despite these variations, the mean values or curves of the descriptors extracted from the reconstructed microstructures closely align with the corresponding reference values or curves. This consistency, combined with the slight variance, demonstrates the statistical equivalence between the reference and reconstructed 3D microstructures. In addition, the excellent agreement observed for geometrical tortuosity and TFOPC further confirms that the developed hierarchical reconstruction approach effectively captures the long-range connectivity of multiphase porous media.

Example 2: Porous SOFC anode

Solid oxide fuel cells (SOFCs) [60, 61] are advanced electrochemical devices that efficiently convert the chemical energy of hydrogen into electricity. They consist of a dense electrolyte sandwiched between an anode and a cathode. Porous nickel/yttria-stabilized zirconia (Ni-YSZ) cermet [61, 62] is widely used as an anode material in SOFCs. In this study, it serves as a representative multiphase composite to evaluate the performance of the proposed SENN-based reconstruction approach. As shown in Figure 12, this Ni-YSZ-based SOFC anode comprises three constituent phases: YSZ (black), Ni (grey), and pore space (black).

Visual comparison between the reference and reconstructed 2D/3D microstructures of a porous SOFC anode

Figure 12. Visual comparison between the reference (top row) and the reconstructed (bottom row) 2D/3D microstructures of a porous SOFC anode with three distinct constituent phases: black regions represent YSZ, grey regions represent Ni, and white regions represent pore space.

As shown in the upper row of Figure 12, the morphological differences between the 2D exemplars along different directions reveal the anisotropy of the porous microstructure in the SOFC anode. Representative 2D exemplars on nine different planes are selected as training images to construct the SENN-based characterisation. To capture the long-range connectivity of the pore networks (white phase), the SOFC anode is statistically characterised at three hierarchical length scales, as described in § 3.3. Three-level Gaussian image pyramids are constructed from the original high-resolution 2D exemplars. Using training data derived from these Gaussian pyramids, a set of SENN models is fitted to infer 3D morphological statistics across different length scales, represented as P(Xijklow∣Nijklow)P(X_{ijk}^{\mathrm{low}} \mid N_{ijk}^{\mathrm{low}}), P[Xijkmiddle∣(Nijkmiddle∪Xijklow∪Nijklow)]P[X_{ijk}^{\mathrm{middle}} \mid(N_{ijk}^{\mathrm{middle}} \cup X_{ijk}^{\mathrm{low}} \cup N_{ijk}^{\mathrm{low}})] and P[Xijkhigh∣(Nijkhigh∪Xijkmiddle∪Nijkmiddle)]P[X_{ijk}^{\mathrm{high}} \mid(N_{ijk}^{\mathrm{high}} \cup X_{ijk}^{\mathrm{middle}} \cup N_{ijk}^{\mathrm{middle}})]. Besides, the parameter settings for training data preparation and SENN modelling are summarised in Table A1.

Leveraging the three-level morphological statistics embedded in the pretrained SENN models, a group of thirty 3D microstructure samples are stochastically synthesised. A representative sample is shown in the bottom row of Figure 12. Visual comparison indicates that the reconstructed 3D microstructure closely resembles the reference sample, particularly in reproducing anisotropic morphology patterns. To quantitatively assess the morphological similarity, four descriptors are extracted from both the reference and reconstructed 3D microstructures. These descriptors include volume fraction, surface area density [62], triple-phase boundary density [61], and the two-point correlation function (TPCF) [18]. These metrics are essential for evaluating the macroscopic behaviour of porous SOFC anodes, such as elastic properties and effective diffusivity [6, 61].

The four morphological descriptors are computed for all three constituent phases of the porous SOFC anode, with results shown in Figure 13. Overall, the descriptor values or curves extracted from the reconstructed 3D microstructure samples exhibit minor fluctuations around those derived from the reference 3D microstructure. This variation is expected due to the inherent randomness of the statistical reconstruction process. Despite these fluctuations, the mean values or curves of the descriptors from the reconstructed microstructures closely match their corresponding reference values or curves. This observed consistency, coupled with the minor variations, confirms the statistical equivalence between the reconstructed and reference 3D microstructures. Moreover, the excellent agreements in geometrical tortuosities along different directions (as shown in Figure 13c) highlight the effectiveness of the proposed multi-level reconstruction method in accurately capturing the anisotropy and long-range connectivity of multiphase porous media.

Quantitative comparison between the reference and reconstructed 3D microstructures

Figure 13. Quantitative comparison between the reference and reconstructed 3D microstructures by evaluating morphological descriptors.

Example 3: Composite cement paste

Composite cement paste [63, 64], a critical component in construction materials, consists of a cement matrix interspersed with various additives or inclusions, such as fly ash, slag, and silica fume. The heterogeneous microstructure of composite cement paste significantly influences its macroscopic properties, including strength, permeability, shrinkage resistance, and thermal conductivity. In this work, composite cement paste is used as a representative example to test the capability of the proposed SENN-based reconstruction approach. As shown in Figure 14 (upper row), it consists of four constituent phases, including calcium silicate hydrate (C-S-H), anhydrous grains, portlandite crystals (calcium hydroxide) and water-filled pores. This four-phase microstructure provides an excellent platform for evaluating the method’s ability to capture complex morphologies and statistical characteristics.

Visual comparison between the reference and the reconstructed 3D microstructure samples

Figure 14. Visual comparison between the reference (top row) and the reconstructed (bottom row) 3D microstructure samples for a composite cement paste with four distinct constituent phases: the black region represents calcium silicate hydrate (C-S-H), the dark grey region represents anhydrous grains, the light grey region represents portlandite (calcium hydroxide) crystals, and the white region represents water-filled pores.

As illustrated in the upper row of Figure 14, representative 2D exemplars from three principal planes are selected for statistical characterisation and subsequent 3D microstructure reconstruction. Two-level Gaussian image pyramids are constructed from the original 2D exemplars, serving as the basis for collecting training data to fit a set of SENN models. These SENN models accurately approximate 2D morphological statistics on different planes, enabling the inference of CPDFs across two length scales, represented as P(Xijk(low)∣Nijk(low))P\left(X_{ijk}^{(\mathrm{low})}\mid N_{ijk}^{(\mathrm{low})}\right) and P[Xijk(high)∣Nijk(high)∪Xijk(middle)∪Nijk(middle)]P\left[X_{ijk}^{(\mathrm{high})}\mid N_{ijk}^{(\mathrm{high})}\cup X_{ijk}^{(\mathrm{middle})}\cup N_{ijk}^{(\mathrm{middle})}\right], respectively. Further details on hierarchical characterisation are provided in § 3.3, while the parameter settings for SENN model training are summarised in Table A1.

Stochastic reconstruction is performed using the 3D CPDFs derived from the pretrained SENN models, producing a group of thirty 3D microstructure samples. A representative reconstruction result is shown in the lower row of Figure 14. Visual inspection reveals that the reconstructed 3D microstructure closely resembles both the 2D exemplars and the reference 3D microstructure, demonstrating the effectiveness of the proposed approach in accurately reproducing morphological patterns. To quantitatively assess the reconstruction quality, a set of morphological descriptors is extracted from the reconstructed and reference 3D microstructures. These descriptors include volume fraction, surface area density, mean chord length, and the two-point cluster correlation function (TPCCF) [6, 51]. These metrics are crucial for evaluating the macroscopic behaviour of composite cement paste, such as mechanical strength [63] and thermal conductivity [64].

The first three descriptors are evaluated for all four constituent phases of the composite cement paste, while the TPCCF is specifically calculated for the three secondary phases to measure cluster sizes. The descriptor values or curves derived from the thirty reconstructed 3D microstructures show slight variations compared to those obtained from the reference 3D microstructure, primarily due to the stochastic nature of voxel generation during the reconstruction process. Nonetheless, the mean values or curves of the reconstructed descriptors closely align with their reference counterparts. This close agreement, alongside the minor variations, validates the statistical equivalence between the reconstructed and reference 3D microstructures. Additionally, the accurate reproduction of mean chord length and TPCCF confirms the capability of the proposed multi-level reconstruction method to effectively capture and preserve the grain and pore size characteristics of multiphase porous media.

Quantitative comparison between the reference and reconstructed 3D microstructures by evaluating morphological descriptors

Figure 15. Quantitative comparison between the reference and reconstructed 3D microstructures by evaluating morphological descriptors.

Image-based poro/micro-mechanical modelling

High-resolution 3D digital microstructures of multiphase composites are essential for image-based poro/micro-mechanical modelling, enabling advanced investigations such as materials design and the exploration of MPRs [65, 66, 67]. These models facilitate accurate simulations of various physical phenomena, but generating precise meshes from 3D image data poses significant challenges due to the intricate geometries and complex topologies of multiphase composites [68, 69]. To overcome these challenges, this study employs advanced image-based meshing techniques [69] to generate high-quality volumetric meshes from digital microstructures. These meshes serve as a foundation for numerical simulations using homogenization methods, enabling the computation of effective material properties. The evaluated physical properties include stiffness, permeability, effective diffusivity, and effective thermal conductivity tensors. These quantitative metrics provide a robust means of assessing the physical equivalence between the reference and reconstructed 3D microstructures. Such evaluations play a critical role in verifying the effectiveness and practicality of the proposed SENN-based reconstruction method. Additionally, the physical properties of the constituent phases in the studied multiphase composites are summarised in Table 1, offering essential inputs for numerical simulations and further analyses.

Multiphase compositeConstituent phaseElastic modulus EEPossion’s ratio ν\nuThermal conductivity χ\chi
Silver-based electrodeSilver76.00 GPa0.380–
Silver-based electrodePTFE0.50 GPa0.460–
Silver-based electrodePore0.000.000–
Porous SOFC anodeYSE215.00 GPa0.317–
Porous SOFC anodeNi200.00 GPa0.300–
Porous SOFC anodePore0.000.000–
Composite cement pastCalcium silicate hydrate25.0 GPa0.2000.40 W/m·K
Composite cement pastAnhydrous grain138.0 GPa0.3001.40 W/m·K
Composite cement pastPortlandite crystal35.4 GPa0.3001.15 W/m·K
Composite cement pastWater-filled pore0.0001 GPa0.4990.61 W/m·K

Table 1. The physical properties of constituent phases in multiphase composite materials

Example 1: Mechanical and transport properties

The performance and durability of porous silver-based electrodes in electrochemical applications rely heavily on their mechanical and transport properties, such as stiffness and permeability. These properties are shaped by the internal microstructure, phase distribution, porosity, and interactions among the silver matrix, PTFE, and pore spaces. Gaining a detailed understanding of these factors is essential for optimising devices like SOFCs and batteries. To analyse these characteristics, image-based meshing is conducted on both the reference and reconstructed 3D microstructures for subsequent poro/micro-mechanical modelling. As shown in Figure 16a and d, unstructured tetrahedral meshes are generated to accurately represent the intricate interfaces and shared nodes among the multiple phases in this heterogeneous material.

Image-based poro/micro-mechanical modelling of porous silver-based electrodes

Figure 16. Image-based poro/micro-mechanical modelling on the reference (top raw) and the reconstructed (bottom raw) 3D microstructures of the porous silver-based electrode: (a) and (d) Mesh generation on multiple microstructures: red areas denote silver, cyan areas denote PTFE, and green areas represents pore space; (b) and (e) Von Mises stress fields in the solid phases under axial compression; (c) and (f) Pressure fields of viscous fluid flow within pore space.

To compute stiffness tensors, the finite element method (FEM) is conducted on image-based meshes of multi-phase electrode microstructures to simulate axial and shear deformations. Figure 16b and e depict stress fields in the reference and reconstructed 3D microstructures under axial compression, providing insights into the mechanical behaviour of the electrode and validating the fidelity of the reconstruction process. Using Voigt notation, the stiffness tensor [70] connects stress to strain through the generalized Hooke’s law:

[σ11σ22σ33σ12σ23σ31]=[e11e12e13000e21e22e23000e31e32e33000000e44000000e55000000e66][ε11ε22ε33ε12ε23ε31](21)\begin{bmatrix} \sigma_{11}\\ \sigma_{22}\\ \sigma_{33}\\ \sigma_{12}\\ \sigma_{23}\\ \sigma_{31} \end{bmatrix} = \begin{bmatrix} e_{11} & e_{12} & e_{13} & 0 & 0 & 0\\ e_{21} & e_{22} & e_{23} & 0 & 0 & 0\\ e_{31} & e_{32} & e_{33} & 0 & 0 & 0\\ 0 & 0 & 0 & e_{44} & 0 & 0\\ 0 & 0 & 0 & 0 & e_{55} & 0\\ 0 & 0 & 0 & 0 & 0 & e_{66} \end{bmatrix} \begin{bmatrix} \varepsilon_{11}\\ \varepsilon_{22}\\ \varepsilon_{33}\\ \varepsilon_{12}\\ \varepsilon_{23}\\ \varepsilon_{31} \tag*{(21)} \end{bmatrix}

where σij\sigma_{ij} denotes stress components, εij\varepsilon_{ij} represents strain components, and eije_{ij} corresponds to stiffness components.

Volumetric stiffness tensors, as visualized in Figure 17a and b, are computed for both reference and reconstructed microstructures. The similarity in elasticity surfaces for these microstructures indicates strong agreement in their mechanical properties. Further, the volumetric stiffness tensors of thirty reconstructed microstructures are projected onto three principal planes, as shown in Figure 17c-Figure 17e. The stiffness tensors of reconstructed samples exhibit minor fluctuations around the reference one, with their mean values closely matching the reference. This observed agreement, combined with minor variability, confirms the statistical equivalence of the reference and reconstructed microstructures in terms of elasticity.

Comparison of stiffness tensors between the reference and reconstructed 3D microstructures of the silver-based electrode

Figure 17. (a) Volumetric representation of stiffness tensor (reference); (b) Volumetric representation of stiffness tensor (reconstruction); (c) Stiffness tensor on xyxy-plane; (d) Stiffness tensor on yzyz-plane; (e) Stiffness tensor on zxzx-plane. Comparison of stiffness tensors between the reference and reconstructed 3D microstructures of the silver-based electrode.

To compute the permeability tensor, the Navier-Stokes equations are numerically solved on the image-based meshes of the porous microstructures via FEM. This involves applying a pressure gradient along specific directions and simulating fluid flow through the pore space. Figure 16c and Figure 16f depict the simulated flow pressure fields within the pore spaces of the reference and reconstructed 3D microstructures, revealing the intricate flow paths. The permeability tensor is derived from Darcy’s law [71], which establishes the relationship between the volumetric flow rate and the pressure gradient as follows:

[qxqyqz]=−1μ[K11K12K13K21K22K23K31K32K33][∂P∂x∂P∂y∂P∂z](22)\begin{bmatrix} q_x \\ q_y \\ q_z \end{bmatrix} = -\frac{1}{\mu} \begin{bmatrix} K_{11} & K_{12} & K_{13} \\ K_{21} & K_{22} & K_{23} \\ K_{31} & K_{32} & K_{33} \end{bmatrix} \begin{bmatrix} \frac{\partial P}{\partial x} \\ \frac{\partial P}{\partial y} \\ \frac{\partial P}{\partial z} \tag*{(22)} \end{bmatrix}

where qxq_x, qyq_y and qzq_z denote volumetric flow rates along xx-, yy- and zz-directions, respectively; μ\mu is the dynamic viscosity; PP represents pressure ; and KijK_{ij} represents the permeability tensor components, quantifying the flow response in the ii-direction to a pressure gradient in the jj-direction.

Volumetric representations of the permeability tensors, as depicted in Figure 18a and Figure 18b, exhibit strong alignment between the reference and reconstructed 3D microstructures. Additionally, permeability values derived from thirty reconstructed samples show fluctuations around the reference value, with the mean closely matching the reference permeability tensor. This alignment is further evidenced by the projections of the tensor onto principal planes in Figure 18c-Figure 18e. This consistency demonstrates that the reconstructed 3D microstructure samples effectively replicate the transport efficiency of the reference microstructure, validating the capability of the proposed reconstruction method in accurately preserving fluid transport properties.

Comparison of permeability tensors between the reference and reconstructed 3D microstructures of the silver-based electrode

Figure 18. (a) Volumetric representation of permeability tensor (reference); (b) Volumetric representation of permeability tensor (reconstruction); (c) Permeability tensor on xyxy-plane; (d) Permeability tensor on yzyz-plane; (e) Permeability tensor on zxzx-plane. Comparison of permeability tensors between the reference and reconstructed 3D microstructures of the silver-based electrode.

Example 2: Mechanical and diffusional properties

The mechanical and diffusional properties of porous SOFC anodes are essential for their functionality and longevity in electrochemical systems. Mechanical properties, such as stiffness, maintain structural integrity under operational stresses, while effective diffusivity governs the efficiency of gas transport through the porous network. These properties are largely determined by the microstructure, including phase distribution, porosity, and pore connectivity. To evaluate these features, image-based meshing of both reference and reconstructed 3D microstructures is conducted for poro/micro-mechanical modelling. As depicted in Figure 19a and Figure 19d, unstructured tetrahedral meshes are generated for 3D digital microstructures, and these volumetric meshes precisely capture the complex interfaces and shared nodes between phases within this heterogeneous material.

Image-based poro/micro-mechanical modelling on the reference and reconstructed 3D microstructures

Figure 19. Image-based poro/micro-mechanical modelling on the reference (top raw) and the reconstructed (bottom raw) 3D microstructures of the porous SOFC anode: (a) Image-based meshing; (b) Stress field in solid phases; (c) Concentration field in pore space; (d) Image-based meshing; (e) Stress field in solid phases; (f) Concentration field in pore space. (a) and (d) Mesh generation on multiphase microstructures: red regions denote YSZ, cyan regions denote Ni, and green regions represent pore space; (b) and (e) Von Mises stress fields in the solid phases under axial compression; (c) and (f) Concentration fields of molecular diffusion within pore space.

FEM is applied to the RVEs of the porous SOFC anode to derive stress-strain relationships, enabling the extraction of the stiffness tensor. Figure 19b and Figure 19e display the stress fields in the reference and reconstructed 3D microstructures under axial compression, providing insights into the mechanical behaviour of the material. Volumetric stiffness tensors, as shown in Figure 20a and Figure 20b, are computed for both the reference and a representative 3D reconstructed microstructure, revealing similar elasticity surfaces in terms of shape and size, thereby demonstrating consistency between the two microstructures. Additionally, stiffness tensors from thirty reconstructed samples are computed and projected onto three principal planes, as depicted in Figure 20c-Figure 20e. The results reveal minor fluctuations around the reference tensor, with the mean closely matching the reference stiffness tensor. This consistency and the slight variations confirm the statistical equivalence of the reconstructed microstructures to the reference in terms of elasticity.

Comparison of stiffness tensors between the reference and reconstructed 3D microstructures

Figure 20. Comparison of stiffness tensors between the reference and reconstructed 3D microstructures of the porous SOFC anode. (a) Volumetric representation of stiffness tensor (reference); (b) Volumetric representation of stiffness tensor (reconstruction); (c) Stiffness tensor on xy-plane; (d) Stiffness tensor on yz-plane; (e) Stiffness tensor on zx-plane.

To compute the effective diffusivity tensor, FEM simulations are conducted on the image-based meshes of the porous SOFC anode. These simulations solve steady-state diffusion equations under prescribed concentration gradients along specific directions within the pore network. Figure 19c and Figure 19f illustrate the concentration fields for the reference and reconstructed 3D microstructures, highlighting the pathways of diffusive transport. The effective diffusivity tensor [72] links the diffusion flux to concentration gradient via Fick’s law, expressed as:

[JxJyJz]=[λ11λ12λ13λ21λ22λ23λ31λ32λ33][∂C∂x∂C∂y∂C∂z](23)\begin{bmatrix} J_x \\ J_y \\ J_z \end{bmatrix} = \begin{bmatrix} \lambda_{11} & \lambda_{12} & \lambda_{13} \\ \lambda_{21} & \lambda_{22} & \lambda_{23} \\ \lambda_{31} & \lambda_{32} & \lambda_{33} \end{bmatrix} \begin{bmatrix} \frac{\partial C}{\partial x} \\ \frac{\partial C}{\partial y} \\ \frac{\partial C}{\partial z} \tag*{(23)} \end{bmatrix}

where JxJ_x, JyJ_y and JzJ_z denote the diffusion flux components along x-, y- and z- directions, respectively; CC represents the concentration; and λij\lambda_{ij} is the element of the effective diffusivity tensor.

Figures 21a and b depict volumetric representations of the effective diffusivity tensors for both the reference microstructure and a representative reconstructed sample. The similarity in shape and size between the two diffusivity surfaces demonstrates strong consistency between the reference and reconstructed microstructures. Additionally, volumetric diffusivity tensors for thirty reconstructed microstructures are computed and projected onto three principal planes, as shown in Figure 21c–Figure 21e. The results indicate that the diffusivity tensors of the reconstructed samples exhibit minor fluctuations around the reference tensor, with the mean closely matching the reference value. These findings confirm the statistical equivalence of the reference and reconstructed microstructures in terms of effective diffusivity, supported by the observed consistency and minor variations.

Comparison of effective diffusivity tensors between the reference and reconstructed 3D microstructures of the porous SOFC anode

Figure 21. Comparison of effective diffusivity tensors between the reference and reconstructed 3D microstructures of the porous SOFC anode.

Example 3: Mechanical and thermal-conduction properties

The mechanical and thermal-conduction properties of composite cement paste play a pivotal role in its structural integrity and thermal performance. Mechanical properties, such as stiffness and strength, enable the material to withstand external loads, while effective thermal conductivity determines its heat transfer capability and impacts its susceptibility to thermal cracking. These properties are governed by the microstructure, including phase composition, porosity, and the connectivity of the solid matrix and pore network. To evaluate these properties, image-based meshing is applied to both reference and reconstructed 3D microstructures, facilitating micro-mechanical and thermal simulations. As depicted in Figure 22a and Figure 22d, the unstructured tetrahedral meshes are designed to precisely capture the intricate interfaces and interactions between phases, accurately representing the complex microstructures of this heterogeneous material.

Image not extracted

Figure 22. (a) Image-based meshing (b) Stress field (c) Heat flux field (d) Image-based meshing (e) Stress field (f) Heat flux field

Figure 22: (a) and (d) Mesh generation on multiphase microstructures, where red regions correspond to calcium silicate hydrate, cyan regions represent anhydrous grains, green regions indicate portlandite (calcium hydroxide) crystals, and yellow regions denote water-filled pores; (b) and (e) Von Mises stress fields under axial compression; (c) and (f) Heat flux distribution at steady state.

FEM is applied to the RVEs of composite cement paste to extract stress-strain relationships, enabling the derivation of the stiffness tensor. Figure 22b and Figure 22e illustrate the Von Mises stress fields under axial compression for the reference and reconstructed 3D microstructures, offering insights into their mechanical behaviour. Volumetric stiffness tensors are calculated for both the reference and a representative reconstructed 3D microstructure, as illustrated in Figure 23a and Figure 23b. The resulting elasticity surfaces exhibit comparable shape and size, underscoring the alignment between the reference and reconstructed microstructures. Furthermore, stiffness tensors of 30 reconstructed microstructure samples are analysed and projected onto three principal planes, as shown in Figure 23c-Figure 23e. The projections indicate minor deviations from the reference tensor, with the average values closely aligning with the reference stiffness tensor. These findings highlight the reconstructed microstructures’ ability to statistically replicate the elastic properties of the reference, supported by their consistency and minor variations.

Image not extracted

Figure 23. (a) Volumetric representation of stiffness tensor (reference) (b) Volumetric representation of stiffness tensor (reconstruction) (c) Stiffness tensor on xy-plane (d) Stiffness tensor on yz-plane (e) Stiffness tensor on zx-plane. Comparison of stiffness tensors between the reference and reconstructed 3D microstructures of the composite cement past.

To compute the effective thermal conductivity tensors, FEM simulations are performed on image-based meshes of the composite cement paste. These simulations solve steady-state heat conduction equations under imposed temperature gradients in specific directions within the heterogeneous microstructures. Figure 22c and Figure 22f depict the heat flux fields for the reference and reconstructed thirty microstructures, illustrating the pathways of heat transfer through the solid phases and water-filled pores. The effective thermal conductivity tensor relates the heat flux to the temperature gradient through Fourier’s law, expressed as:

[HxHyHz]=[χ11χ12χ13χ21χ22χ23χ31χ32χ33][∂T∂x∂T∂y∂T∂z](24)\begin{bmatrix} H_x \\ H_y \\ H_z \end{bmatrix} = \begin{bmatrix} \chi_{11} & \chi_{12} & \chi_{13} \\ \chi_{21} & \chi_{22} & \chi_{23} \\ \chi_{31} & \chi_{32} & \chi_{33} \end{bmatrix} \begin{bmatrix} \frac{\partial T}{\partial x} \\ \frac{\partial T}{\partial y} \\ \frac{\partial T}{\partial z} \tag*{(24)} \end{bmatrix}

where HxH_x, HyH_y and HzH_z represent the heat flux components along xx-, yy- and zz- directions, respectively; TT represents the temperature; and χij\chi_{ij} is the element of the effective thermal conductivity tensor.

Figure 24a and Figure 24b illustrate the 3D volumetric distributions of the effective thermal conductivity tensors for both the reference and a selected reconstructed microstructure. The comparable shape and size of the conductivity surfaces for both structures highlight a high degree of alignment between the reference and the reconstructed 3D microstructures. Furthermore, the thermal conductivity tensors for thirty reconstructed microstructures are calculated and projected onto three principal planes, as shown in Figure 24c-Figure 24e. The analysis reveals that the tensors from the reconstructed samples fluctuate slightly around the reference tensor, with their average value closely approximating the reference. These results underscore the statistical equivalence between the reference and reconstructed microstructures, validating the accuracy of the stochastic microstructure reconstruction process in preserving effective thermal conductivity.

Image not extracted

Figure 24. (a) Volumetric representation of conductivity tensor (reference) (b) Volumetric representation of conductivity tensor (reconstruction) (c) Thermal conductivity tensor on xy-plane (d) Thermal conductivity tensor on yz-plane (e) Thermal conductivity tensor on zx-plane. Comparison of effective thermal conductivity tensors between the reference and reconstructed 3D microstructures of the composite cement past.

Discussion and conclusions

Discussion

The proposed SENN-based framework for stochastic microstructure reconstruction is a versatile tool for advancing microstructure-property analyses in materials science, energy storage, and other related fields. By reducing reliance on costly volumetric microscopy techniques, this framework is particularly valuable for studying opaque multiphase composites, where acquiring sufficient 3D digital microstructures is challenging. Its ability to synthesise statistically equivalent 3D microstructures from limited 2D exemplars enables large-scale statistical analyses, providing a cost-efficient pathway to better understand and optimise the performance and functionality of multiphase composites. The key features of this innovative SENN-based framework include:

  • Multiphase reconstruction: Unlike most reconstruction methods tailored to bi-phase microstructures, this framework specialises in reconstructing multiphase microstructures with complex morphologies. It demonstrates a strong capability to accurately capture the random distribution of phases and the intricate interfaces between them;

  • Accuracy and efficiency: By integrating SENN with Gibbs sampling, the framework achieves precise statistical equivalence between reconstructed and reference microstructures while maintaining computational efficiency. Its hierarchical characterisation and multi-level reconstruction approach further ensure seamless capture and preservation of local, regional, and global features across various length scales;

  • Small data requirement: In contrast to traditional machine learning-based approaches that require extensive 3D datasets, this SENN-based framework can reliably synthesise 3D microstructures using only a limited number of 2D exemplars. This feature is particularly advantageous in cases where obtaining volumetric datasets is prohibitively expensive or technically infeasible;

  • Model interpretability: SENN explicitly encodes morphological statistics, providing clear insights into the model’s predictions and fostering trust in the reconstruction process. This interpretability also enhances the understanding of the relationships between 2D morphological features and the 3D counterparts.

However, certain limitations of the SENN-based framework should be noted. First, the framework relies on the quality and diversity of the input 2D exemplars; inadequate representation of critical features in 2D data may introduce biases in the reconstructed 3D microstructures. Second, the Markov random field assumption used for statistical characterisation may be less effective for multiphase microstructures with extremely strong heterogeneity. Third, while Gibbs sampling is effective for synthesising microstructures, it can become computationally intensive for extremely large datasets or less effective for highly complex morphologies. Future work could address these challenges by exploring more scalable synthesis techniques and enhancing the framework’s adaptability to multiphase composites with evolving morphologies. Furthermore, integrating the framework with advanced optimisation algorithms could enhance its predictive capability, enabling direct linkage between desired macroscopic properties and reconstructed microstructures. Such developments would further expand the framework’s applicability and utility in real-world material design and analysis.

Conclusions

This study presents a novel machine learning-based framework for stochastic microstructure reconstruction, addressing the critical challenge of generating 3D digital microstructures of random multiphase composites from limited 2D exemplars (cross-sectional images). Leveraging the innovative Statistics-Encoded Neural Network (SENN), the framework encodes and extrapolates 2D morphological statistics to synthesise statistically equivalent 3D microstructures, significantly reducing reliance on costly volumetric microscopy techniques. By integrating hierarchical characterisation and multi-level reconstruction, the SENN-based framework captures microstructural features at local, regional, and global scales with high fidelity, achieving statistical equivalence to reference microstructures. The reconstructed 3D microstructures exhibit strong morphological similarity to reference datasets, as confirmed by quantitative descriptors such as volume fraction, surface area density, triple-phase boundary density, and geometrical tortuosity. Furthermore, quantitative analysis of effective macroscopic properties, including stiffness, permeability, diffusivity, and thermal conductivity tensors, further affirmed the statistical equivalence and practical applicability of the reconstructed 3D microstructures for image-based poro/micro-mechanical modelling.

In conclusion, this work makes several key contributions to image-based poro/micro-mechanical modelling and MPR analysis. First, the SENN-based framework provides a cost-efficient solution for generating 3D multiphase microstructures, mitigating the need for extensive volumetric microscopy imaging. Second, its integration of hierarchical characterisation and multi-level reconstruction approaches ensures the accurate representation and capture of microstructural features across multiple relevant length scales, making the framework particularly effective for complex, multiphase composites. Finally, by enabling large-scale statistical analysis of MPRs, the SENN-based framework opens new avenues for materials design and optimisation across diverse applications, including materials science, structural engineering, and electrochemical devices.

Parameter settings for training statistics-encoded neural networks

To train SENN models effectively for approximating 2D morphological statistics, data events extracted from 2D exemplars (cross-sectional images) are randomly divided into three subsets: 70% for training, 20% for validation, and 10% for testing. This data-splitting strategy reduces the risk of overfitting and ensures robust model performance. The theoretical framework for SENN-based characterisation is outlined in § 2, while the parameter settings for training SENN models in the three case studies are provided in Table A1.

Acknowledgements

Wei Tan acknowledges the financial support from the EPSRC New Investigator Award [grant number: EP/V049259/1]. This work was selected by the ERC and funded by UK Research and Innovation (UKRI) under the UK government’s Horizon Europe funding guarantee [grant number EP/Y037103/1].

Data availability

Data will be made available on request.

Declarations

The authors declare that they have no conflict of interest in this paper.

References

  1. [1]S Torquato. Modeling of physical properties of composite materials. International Journal of Solids and Structures, 37(1-2):411–422, 2000.
  2. [2]Xiaolei Yin, Wei Chen, Albert To, Cahal McVeigh, and Wing Kam Liu. Statistical volume element method for predicting microstructure–constitutive property relations. Computer Methods in Applied Mechanics and Engineering, 197(43-44):3516–3529, 2008.
  3. [3]Jian Li and Stephan Rudy​​kh. Tunable microstructure transformations and auxetic behavior in 3d-printed multiphase composites: the role of inclusion distribution. Composites Part B: Engineering, 172:352–362, 2019.DOI
  4. [4]Ahmet Cecen, Hanjun Dai, Yuksel C Yabansu, Surya R Kalidindi, and Le Song. Material structure-property linkages using three-dimensional convolutional neural networks. Acta Materialia, 146:76–84, 2018.
  5. [5]TWJ Kwok and D Dye. A review of the processing, microstructure and property relationships in medium mn steels. International Materials Reviews, 68(8):1098–1134, 2023.DOI
  6. [6]Jinlong Fu, Min Wang, Bin Chen, Jinsheng Wang, Dunhui Xiao, Min Luo, and Ben Evans. A data-driven framework for permeability prediction of natural porous rocks via microstructural characterization and pore-scale simulation. Engineering with Computers, 39(6):3895–3926, 2023.
  7. [7]Yangjian Xu, Huanbo Weng, Xiaozhe Ju, Hongshi Ruan, Junjun Chen, Chenyu Nan, Jing Guo, and Lihua Liang. A method for predicting mechanical properties of composite microstructure with reduced dataset based on transfer learning. Composite Structures, 275:114444, 2021.
  8. [8]Chenyu Nan, Hongshi Ruan, Xiaozhe Ju, Junhan Hu, Lihua Liang, and Yangjian Xu. Transfer-learning-based strategy for enhancing prediction accuracy and computational efficiency of nonlinear mechanical properties in composite materials. Composites Science and Technology, 246:110388, 2024.
  9. [9]Marat I Latypov, Laszlo S Toth, and Surya R Kalidindi. Materials knowledge system for nonlinear composites. Computer Methods in Applied Mechanics and Engineering, 346:180–196, 2019.
  10. [10]Faisal H Bhuiyan, Seyed Hamid R Sanei, and Ray S Fertig III. Predicting variability in transverse effective elastic moduli and failure initiation strengths in ud composite microstructures due to randomness in fiber location and morphology. Composite Structures, 237:111887, 2020.DOI
  11. [11]Anna Madra, Jérôme Adrien, Piotr Breitkopf, Eric Maire, and François Troc hu. A clustering method for analysis of morphology of short natural fibers in composites based on x-ray microtomography. Composites Part A: Applied Science and Manufacturing, 102:184–195, 2017.
  12. [12]Yoshimasa Urushihara, Lei Li, Junji Matsui, and Takashi Nishino. In situ observation of filler displacement during tensile deformation of nanosilica-filled natural rubber using field-emission scanning electron microscope. Composites Part A: Applied Science and Manufacturing, 40(2):232–234, 2009.DOI
  13. [13]Martin Ostoja-Starzewski. Material spatial randomness: From statistical to representative volume element. Probabilistic Engineering Mechanics, 21(2):112–132, 2006.
  14. [14]Srinivas Sriramula and Marios K Chryssanthopoulos. Quantification of uncertainty modelling in stochastic analysis of frp composites. Composites Part A: Applied Science and Manufacturing, 40(11):1673–1684, 2009.DOI
  15. [15]Miguel A Bessa, Ramin Bostanabad, Zeliang Liu, Anqi Hu, Daniel W Apley, Catherine Brinson, Wei Chen, and Wing Kam Liu. A framework for data-driven analysis of materials under uncertainty: Countering the curse of dimensionality. Computer Methods in Applied Mechanics and Engineering, 320:633–667, 2017.DOI
  16. [16]Ramin Bostanabad, Yichi Zhang, Xiaolin Li, Tucker Kearney, L Catherine Brinson, Daniel W Apley, Wing Kam Liu, and Wei Chen. Computational microstructure characterization and reconstruction: Review of the state-of-the-art techniques. Progress in Materials Science, 95:1–41, 2018.DOI
  17. [17]Jinlong Fu, Min Wang, Dunhui Xiao, Shan Zhong, Xiangyun Ge, Minglu Wu, and Ben Evans. Hierarchical reconstruction of 3D well-connected porous media from 2D exemplars using statistics-informed neural network. Computer Methods in Applied Mechanics and Engineering, 410:116049, 2023.DOI
  18. [18]Jinlong Fu, Dunhui Xiao, Dongfeng Li, Hywel R Thomas, and Chenfeng Li. Stochastic reconstruction of 3D microstructures from 2D cross-sectional images using machine learning-based characterization. Computer Methods in Applied Mechanics and Engineering, 390:114532, 2022.
  19. [19]Mohamed H Hamza and A Chattopadhyay. Multi deep learning-based stochastic microstructure reconstruction and high-fidelity micromechanics simulation of time-dependent ceramic matrix composite response. Composite Structures, 345:118360, 2024.DOI
  20. [20]S Torquato and CLY Yeong. Reconstructing random media ii. three-dimensional media from two-dimensional cuts. Physical Review E, 58(1):224–233, 1998.DOI
  21. [21]Yang Ju, Jiangtao Zheng, Marcelo Epstein, Les Sudak, Jinbo Wang, and Xi Zhao. 3d numerical reconstruction of well-connected porous structure of rock using fractal algorithms. Computer Methods in Applied Mechanics and Engineering, 279:212–226, 2014.
  22. [22]ZR Liang, CP Fernandes, FS Magnani, and PC Philippi. A reconstruction technique for three-dimensional porous media using image analysis and fourier transforms. Journal of Petroleum Science and Engineering, 21(3-4):273–283, 1998.
  23. [23]Pål-Eric Øren and Stig Bakke. Process based reconstruction of sandstones and prediction of transport properties. Transport in Porous Media, 46(2):311–343, 2002.DOI
  24. [24]Alireza Hajizadeh, Aliakbar Safekordi, and Farhad A Farhadpour. A multiple-point statistics algorithm for 3d pore space reconstruction from 2d images. Advances in Water Resources, 34(10):1256–1267, 2011.DOI
  25. [25]Pejman Tahmasebi and Muhammad Sahimi. Reconstruction of three-dimensional porous media using a single thin section. Physical Review E—Statistical, Nonlinear, and Soft Matter Physics, 85(6):066709, 2012.
  26. [26]Xingchen Liu and Vadim Shapiro. Random heterogeneous materials via texture synthesis. Computational Materials Science, 99:177–189, 2015.DOI
  27. [27]Hongyi Xu, Dmitry A Dikin, Craig Burkhart, and Wei Chen. Descriptor-based methodology for statistical characterization and 3d reconstruction of microstructural materials. Computational Materials Science, 85:206–216, 2014.DOI
  28. [28]Guilhem Mollon and Jidong Zhao. 3d generation of realistic granular samples based on random fields theory and fourier shape descriptors. Computer Methods in Applied Mechanics and Engineering, 279:46–65, 2014.
  29. [29]Ming Yang, Anand Nagarajan, Bowen Liang, and Soheil Soghrati. New algorithms for virtual reconstruction of heterogeneous microstructures. Computer Methods in Applied Mechanics and Engineering, 338:275–298, 2018.DOI
  30. [30]Andries P Engelbrecht. Computational intelligence: an introduction. John Wiley & Sons, 2007.
  31. [31]Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436–444, 2015.
  32. [32]Veeraraghavan Sundararaghavan and Nicholas Zabaras. Classification and reconstruction of three-dimensional microstructures using support vector machines. Computational Materials Science, 32(2):223–239, 2005.
  33. [33]Ramin Bostanabad, Anh Tuan Bui, Wei Xie, Daniel W Apley, and Wei Chen. Stochastic microstructure characterization and reconstruction via supervised learning. Acta Materialia, 103:89–102, 2016.
  34. [34]Ramin Bostanabad, Wei Chen, and Daniel W Apley. Characterization and reconstruction of 3d stochastic microstructures via supervised learning. Journal of Microscopy, 264(3):282–297, 2016.
  35. [35]Jinlong Fu, Shaoqing Cui, Song Cen, and Chenfeng Li. Statistical characterization and reconstruction of heterogeneous microstructures using deep neural network. Computer Methods in Applied Mechanics and Engineering, 373:113516, 2021.
  36. [36]Lukas Mosser, Olivier Dubrule, and Martin J Blunt. Reconstruction of three-dimensional porous media using generative adversarial neural networks. Physical Review E, 96(4):043309, 2017.
  37. [37]Junxi Feng, Qizhi Teng, Bing Li, Xiaohai He, Honggang Chen, and Yang Li. An end-to-end three-dimensional reconstruction framework of porous media from a single two-dimensional image based on deep learning. Computer Methods in Applied Mechanics and Engineering, 368:113043, 2020.
  38. [38]Alexander Henkes and Henning Wessels. Three-dimensional microstructure generation using generative adversarial neural networks in the context of continuum micromechanics. Computer Methods in Applied Mechanics and Engineering, 400:115497, 2022.
  39. [39]Yuzhu Wang, Christoph H Arns, Sheik S Rahman, and Ji-Youn Arns. Porous structure reconstruction using convolutional neural networks. Mathematical Geosciences, 50(7):781–799, 2018.
  40. [40]Satoshi Noguchi and Junya Inoue. Stochastic characterization and reconstruction of material microstructures for establishment of process-structure-property linkage using the deep generative model. Physical Review E, 104(2):025302, 2021.
  41. [41]Ruijin Cang, Yaopengxiao Xu, Shaohua Chen, Yongming Liu, Yang Jiao, and Max Yi Ren. Microstructure representation and reconstruction of heterogeneous materials via deep belief network for computational material design. Journal of Mechanical Design, 139(7):071404, 2017.
  42. [42]Eric Laloy, Romain Hérault, John Lee, Diederik Jacques, and Niklas Linde. Inversion using a new low-dimensional representation of complex binary geological media based on a deep neural network. Advances in Water Resources, 110:387–405, 2017.
  43. [43]Xiaolin Li, Yichi Zhang, He Zhao, Craig Burkhart, L Catherine Brinson, and Wei Chen. A transfer learning approach for microstructure reconstruction and structure-property predictions. Scientific Reports, 8(1):13461, 2018.
  44. [44]Kelly M Guan, Timothy I Anderson, Patrice Creux, and Anthony R Kovscek. Reconstructing porous media using generative flow networks. Computers & Geosciences, 156:104905, 2021.
  45. [45]Fan Zhang, Xiaohai He, Qizhi Teng, Xiaohong Wu, Junfang Cui, and Xiucheng Dong. Pm-arnn: 2d-to-3d reconstruction paradigm for microstructure of porous media via adversarial recurrent neural network. Knowledge-Based Systems, 264:110333, 2023.
  46. [46]Yilin Zheng, Zhuo Li, and Zhigong Song. A generative machine learning model for the 3d reconstruction of material microstructure and performance evaluation. Computer Methods in Applied Mechanics and Engineering, 430:117224, 2024.
  47. [47]Kang-Hyun Lee and Gun Jin Yun. Denoising diffusion-based synthetic generation of three-dimensional (3d) anisotropic microstructures from two-dimensional (2d) micrographs. Computer Methods in Applied Mechanics and Engineering, 423:116876, 2024.
  48. [48]Yi Gao, Yang Jiao, and Yongming Liu. Ultra-efficient reconstruction of 3d microstructure and distribution of properties of random heterogeneous materials containing multiple phases. Acta Materialia, 204:116526, 2021.
  49. [49]Steve Kench and Samuel J Cooper. Generating three-dimensional structures from a two-dimensional slice with generative adversarial network-based dimensionality expansion. Nature Machine Intelligence, 3(4):299–305, 2021.
  50. [50]Stan Z Li. Markov random field modeling in image analysis. Springer Science & Business Media, 2009.
  51. [51]Shaoqing Cui, Jinlong Fu, Song Cen, Hywel R Thomas, and Chenfeng Li. The correlation between statistical descriptors of heterogeneous materials. Computer Methods in Applied Mechanics and Engineering, 384:113948, 2021.
  52. [52]Peter B Corson. Correlation functions for predicting properties of heterogeneous materials. ii. empirical construction of spatial correlation functions for two-phase solids. Journal of Applied Physics, 45(7):3165–3170, 1974.
  53. [53]George Casella and Edward I George. Explaining the gibbs sampler. The American Statistician, 46(3):167–174, 1992.
  54. [54]Adrian E Raftery, Steven Lewis, et al. How many iterations in the gibbs sampler. Bayesian Statistics, 4(2):763–773, 1992.
  55. [55]Edward H Adelson, Charles H Anderson, James R Bergen, Peter J Burt, and Joan M Ogden. Pyramid methods in image processing. RCA Engineer, 29(6):33–41, 1984.
  56. [56]R Li, XJ Liu, H Wang, DQ Zhou, Y Wu, and ZP Lu. Formation mechanism and characterization of nanoporous silver with tunable porosity and promising capacitive performance by chemical dealloying of glassy precursor. Acta Materialia, 105:367–377, 2016.
  57. [57]Matthias Neumann, Markus Osenberg, André Hilger, David Franzen, Thomas Turek, Ingo Manke, and Volker Schmidt. On a pluri-gaussian model for three-phase microstructures, with applications to 3d image data of gas-diffusion electrodes. Computational Materials Science, 156:325–331, 2019.
  58. [58]A Bertei, E Ruiz-Trejo, K Kareh, V Yufit, X Wang, F Tariq, and NP Brandon. The fractal nature of the three-phase boundary: a heuristic approach to the degradation of nanostructured solid oxide fuel cell anodes. Nano Energy, 38:526–536, 2017.
  59. [59]Jinlong Fu, Hywel R Thomas, and Chenfeng Li. Tortuosity of porous media: Image analysis and physical simulation. Earth-Science Reviews, 212:103439, 2021.
  60. [60]Serkan Toros. Microstructural finite element modeling of redox behavior of ni–ysz based ceramic sofc anodes. Ceramics International, 42(7):8915–8924, 2016.
  61. [61]Xuhao Liu, Shihao Zhou, Zilin Yan, Zheng Zhong, Naoki Shikazono, and Shotaro Hara. Correlation between microstructures and macroscopic properties of nickel/yttria-stabilized zirconia (ni-ysz) anodes: Meso-scale modeling and deep learning with convolutional neural networks. Energy and AI, 7:100122, 2022.
  62. [62]Samuel J Cooper, Antonio Bertei, Paul R Shearing, JA Kilner, and Nigel P Brandon. Taufactor: An open-source application for calculating tortuosity factors from tomographic data. SoftwareX, 5:203–210, 2016.
  63. [63]Julien Sanahuja, Luc Dormieux, and Gilles Chanvillard. Modelling elasticity of a hydrating cement paste. Cement and Concrete Research, 37(10):1427–1439, 2007.
  64. [64]Alex Ricklefs, Alexander M Thiele, Gabriel Falzone, Gaurav Sant, and Laurent Pilon. Thermal conductivity of cementitious composites containing microencapsulated phase change materials. International Journal of Heat and Mass Transfer, 104:71–82, 2017.
  65. [65]Kenjiro Terada, T Miura, and Noboru Kikuchi. Digital image-based modeling applied to the homogenization analysis of composite materials. Computational Mechanics, 20(4):331–346, 1997.
  66. [66]Fang, Bassam El Said, Dmitry Ivanov, and Stephen R Hallett. Smoothing artificial stress concentrations in voxel-based models of textile composites. Composites Part A: Applied Science and Manufacturing, 80:270–284, 2016.
  67. [67]Lion Marc Evans, Emrah Sözümert, Bethany E Keenan, Charles E Wood, and Anton du Plessis. A review of image-based simulation applications in high-value manufacturing. Archives of Computational Methods in Engineering, 30(3):1495–1552, 2023.
  68. [68]Dietmar Ulrich, Bert van Rietbergen, Harrie Weinans, and Peter Rüegsegger. Finite element analysis of trabecular bone structure: a comparison of image-based meshing techniques. Journal of Biomechanics, 31(12):1187–1192, 1998.
  69. [69]Patrick Perré. Meshpore: a software able to apply image-based meshing techniques to anisotropic and heterogeneous porous media. Drying Technology, 23(9-11):1993–2006, 2005.
  70. [70]Li Zheng, Siddhant Kumar, and Dennis M Kochmann. Data-driven topology optimization of spinodoid metamaterials with seamlessly tunable anisotropy. Computer Methods in Applied Mechanics and Engineering, 383:113894, 2021.
  71. [71]Jinlong Fu, Jiabin Dong, Yongliang Wang, Yang Ju, D Roger J Owen, and Chenfeng Li. Resolution effect: An error correction model for intrinsic permeability of porous media estimated from lattice boltzmann method. Transport in Porous Media, 132(3):627–656, 2020.
  72. [72]Evan John Ricketts. Stochastic periodic microstructures for multiscale modelling of heterogeneous materials. Transport in Porous Media, 151(6):1313–1332, 2024.

Paper details

vlnc:W_zts4wqvf
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents