Source-linked AI summary
Improving Direct Physical Properties Prediction of Heterogeneous Materials from Imaging Data via Convolutional Neural Network and a Morphology-Aware Generative Model
Ruijin Cang, Hechao Li, Hope Yao, Yang Jiao, Yi Ren
TL;DR
Predicting material properties from high-dimensional microstructures is data-demanding when samples are costly to acquire. The paper trains a morphology-constrained generative model on limited authentic data to create artificial samples, which more effectively improves structure-property prediction than a state-of-the-art Markov Random Field model.
Problem
High-dimensional microstructures make nonlinear structure-property prediction data-demanding, while acquiring authentic samples can be costly experimentally or computationally.
Method
A variational-autoencoder-based generative model uses convolutional networks and an auxiliary morphology representation to generate samples matching the authentic morphology distribution.
Results
The proposed method is more data-efficient at improving structure-property prediction than a state-of-the-art Markov Random Field method in a sandstone case study.
Takeaways & Limitations
Generated samples can provide extra data for more cost-effective material design when microstructure acquisition is the design bottleneck.
Takeaways & Limitations
Matching a representation space does not guarantee matching the microstructure space, and future work is needed to enforce physical meaningfulness through physics-based loss.
Abstract
from arXiv · showhide
Direct prediction of material properties from microstructures through statistical models has shown to be a potential approach to accelerating computational material design with large design spaces. However, statistical modeling of highly nonlinear mappings defined on high-dimensional microstructure spaces is known to be data-demanding. Thus, the added value of such predictive models diminishes in common cases where material samples (in forms of 2D or 3D microstructures) become costly to acquire either experimentally or computationally. To this end, we propose a generative machine learning model that creates an arbitrary amount of artificial material samples with negligible computation cost, when trained on only a limited amount of authentic samples. The key contribution of this work is the introduction of a morphology constraint to the training of the generative model, that enforces the resultant artificial material samples to have the same morphology distribution as the authentic ones. We show empirically that the proposed model creates artificial samples that better match with the authentic ones in material property distributions than those generated from a state-of-the-art Markov Random Field model, and thus is more effective at improving the prediction performance of a predictive structure-property model.
1. Introduction
Predictive structure-property models can accelerate material design but require costly, high-dimensional microstructure data. The paper proposes morphology-aware generation of artificial samples to address this data bottleneck.
- Predictive models can mimic highly nonlinear physics-based mappings, reducing reliance on numerical simulations or experiments during material design.
- High-dimensional 2D or 3D microstructure inputs make structure-property prediction data-demanding, reducing its value when sample acquisition is costly.
- The proposed generative model learns from few authentic samples and creates an arbitrary amount of artificial samples at negligible computation cost.
- A morphology constraint improves consistency between artificial and authentic samples by matching their morphology distributions.
- A sandstone case study evaluates prediction of Young’s modulus, diffusion coefficient, and permeability coefficient against a Markov Random Field model.
- The model uses a variational autoencoder with convolutional encoder and decoder networks, while an auxiliary network quantifies target morphology.
2. Background
Material representations must balance interpretability, dimensionality, expressiveness, and generation cost, while limited and expensive data make acquisition a central challenge. The paper addresses representation and generation limits by adding a morphology constraint to a differentiable VAE-based model.
- Data science challenges in computational material science: High-dimensional microstructures and costly sample acquisition limit the value of data-demanding predictive structure-property models.These challenges motivate methods that reduce dependence on additional material samples.
- Material representations: Encoder-decoder representations reduce microstructure dimensionality, but must also match authentic and model data distributions.The encoder maps microstructures to reduced representations, while the decoder reconstructs or generates microstructures.
- Material representations: Representation-space agreement does not guarantee microstructure-space agreement, as visually more plausible images can have worse two-point-correlation matching.Figure 1 compares authentic samples with Markov random field, hybrid, and deep-belief-network generations.
- Material representations: Physical descriptors and correlation functions can be interpretable and low-dimensional yet insufficiently expressive for complex morphologies and costly to decode.Random-field generation can require Markov chain Monte Carlo optimization, adding computational cost.
- Proposed representation and generation approach: The proposed model directly enforces morphology matching through a penalty added to a VAE, avoiding additional parameter tuning and supporting differentiable training.It encodes styles as a distribution rather than a single style vector, making the generator more expressive.
- Variational autoencoders: A VAE combines reconstruction with KL-divergence regularization while encoding inputs as latent distributions and decoding latent samples into outputs.The paper explains that feed-forward network parameterizations make gradients available for backpropagation.
3. Proposed Models
The proposed models combine convolutional VAE encoding and decoding with a VGG-based morphology model and a CNN predictive model. Training uses reconstruction, KL, style, and mode-collapse losses, with paired artificial and authentic samples.
- Network specifications: The generative model combines a VAE for image encoding and decoding with an auxiliary network that computes morphology style vectors.The VAE encoder and decoder use convolutional neural networks, while the decoder mirrors the encoder architecture.
- Network specifications: The predictive model uses a CNN with four convolutional layers and max-pooling layers with stride 2 to predict material properties from microstructures.
- Network specifications: A VGG network supplies morphology features, with the first four layers used because deeper layers capture styles at larger length scales.The feature-map choice is empirical, and style vectors concatenate vectorized Gram matrices.
- Model training: Generative-model training minimizes reconstruction, KL-divergence, style, and mode-collapse losses.The mode-collapse term forces generated samples to differ in their activations in the style network, while the style loss compares generated and authentic image sets.
- Model training: Each training iteration samples 20 authentic microstructures and generates 20 artificial ones, pairing artificial samples with authentic samples to guide morphology targets.The model parameters are optimized through stochastic gradient descent.
- Model training: The predictive model uses training, validation, and test sets, monitors validation performance for overfitting, and evaluates test performance with R-square.Adam is used with a learning rate of 0.001.
4. Case Study and Results
The sandstone case study evaluates whether morphology-aware generations improve microstructure fidelity and structure-property prediction. Across Young’s modulus, diffusivity, and permeability, the proposed approach outperforms the MRF benchmark in generation quality and predictive-data utility.
- Case study: The case study evaluates Young’s modulus, diffusivity, and fluid permeability, which depend on distinct sandstone microstructural features.Young’s modulus is mainly sensitive to rock-phase morphology and volume fraction; diffusivity to local pore-size distribution; and permeability to pore connectivity and interface morphology.
- Microstructure generation: The proposed model creates sandstone microstructures with better visual and statistical similarity to authentic samples than the MRF approach.With only 200 training samples, style loss improves standard VAE generation quality, while MRF generations show unrealistically small stone-phase particles.
- Prediction experiment: 1000 artificial microstructures are generated separately by the proposed method and MRF, and their true properties are computed for predictive-model augmentation.Baseline models use 100 training and 100 validation points, with performance evaluated on a separate 100-point test set.
- Prediction experiment: Additional samples from 50 to 1000 are randomly added to the original training set, with means and variances of test R-squares reported over 10 draws.At size 1000, all additional points are used together rather than randomly subsampled.
- Results: The proposed generations have property distributions similar to authentic samples, whereas MRF generations have significantly smaller variances.The paper attributes the MRF discrepancy to insufficient variance in generation volume fractions, which can produce a gap in downstream prediction performance.
5. Discussion
The proposed generative model provides additional material data for predictive modeling and more cost-effective microstructure design. Its morphology constraint improves consistency with authentic samples, while future work could incorporate physics-based losses to enforce physical meaningfulness.
- Utilities of the proposed method: The model captures low-dimensional representations of high-dimensional material data and supplies extra samples for structure-property prediction.This can be particularly valuable when acquiring microstructure data is expensive.
- Utilities of the proposed method: The model can facilitate more cost-effective design of optimal and feasible microstructures by improving predictive structure-property models with additional data.The paper also suggests active learning to sample microstructures with high prediction uncertainty.
- Scope: The approach is generally applicable across material systems, although minor architecture tuning may be needed for different morphological complexities.Its stated applicability is independent of any specific material system.
- Future directions: Future work could replace style loss with physics-based loss to further improve limited-data generation and enforce physically meaningful samples.The proposed direction measures similarity through material properties or equilibrium-condition violations.
6. Conclusions
The paper presents a low-cost method for generating many artificial sandstone microstructures from few training samples. Incorporating style loss improves their usefulness for structure-property prediction compared with a state-of-the-art Markov Random Field method.
- The method generates an arbitrary amount of artificial microstructure samples with low computation cost from a small training set.
- Incorporating a style loss significantly improves the quality of the generated artificial microstructures.
- The sandstone case study shows greater data efficiency for improving structure-property mapping performance than a state-of-the-art Markov Random Field method.
- The findings motivate future investigation of physically meaningful generative models for accelerating microstructure-mediated material design.