Uncertainty Propagation in a Multiscale CALPHAD-Reinforced Elastochemical Phase-Field Model

Vahid Attari, Pejman Honarmandi, Thien Duong, Daniel Sauceda, Douglas Allaire, Raymundo ArróyaveSSRN 2019 · DOI

TL;DR

This paper introduces a novel framework for quantifying and propagating uncertainties in computational materials science, specifically for the CALPHAD-reinforced elastochemical phase-field model. By integrating Bayesian inference and advanced sampling techniques, the authors demonstrate how uncertainties in fundamental thermodynamic parameters can significantly impact predicted microstructures, offering a more robust approach to materials design.

What Problem Does This Paper Address?

Computational materials science, particularly through Integrated Computational Materials Engineering (ICME), aims to establish quantitative relationships between material processing, structure, and properties. However, the confidence in these predictions is often limited by uncertainties in the input parameters and models. Traditional methods for propagating these uncertainties are computationally expensive and struggle with the complex, non-linear behavior of models like phase-field simulations. This paper tackles the challenge of propagating uncertainties through a multi-scale model chain, specifically coupling CALPHAD (for thermodynamic assessment) with a phase-field model, which is crucial for understanding microstructure evolution in materials like Mg₂(SixSn₁−x) thermoelectrics. The inherent uncertainty in CALPHAD parameters and the sensitivity of phase-field models to these inputs create a significant gap in reliable materials design.

Methods

The authors propose a framework that combines several key techniques:

  • Bayesian Inference (MCMC): For CALPHAD parameter uncertainty quantification, a Markov Chain Monte Carlo (MCMC) approach, specifically the Metropolis-Hastings algorithm, is employed. This allows for probabilistic calibration of model parameters by updating prior knowledge with observed data.
  • Gaussian Copulas for Sampling: To efficiently propagate uncertainties from multiple, potentially correlated, input parameters (thermodynamic, microelastic, and kinetic) into the phase-field model, a Gaussian copula approach is used. This method generates samples that preserve both the marginal distributions and the pairwise correlations of the input parameters, which is crucial for accurate uncertainty propagation in high-dimensional spaces.
  • High-Throughput Phase-Field Simulations: A large number (200,000) of phase-field simulations are performed using advanced sampling schemes. This "high-throughput" approach generates a comprehensive dataset of microstructures under varying input conditions.
  • Machine Learning for Data Analysis: Given the high dimensionality of both the input parameter space and the output microstructure data, machine learning techniques (hierarchical clustering, various classifiers) are used to analyze the generated microstructures, identify key quantities of interest (QoIs), and understand the relationships between input uncertainties and microstructure characteristics.

Datasets

The paper does not explicitly mention pre-existing datasets for training or evaluation. Instead, it focuses on generating a large synthetic dataset of microstructures through high-throughput phase-field simulations. This generated dataset, comprising 200,000 time series of synthetic microstructures, is made available in the Open Phase-field Microstructure Database (OPMD).

Key Insights and Ideas

  • CALPHAD Uncertainty is Critical: Uncertainties in CALPHAD thermodynamic parameters, even if seemingly small, can propagate to significantly impact the predicted phase diagram and, consequently, the resulting microstructures.
  • Multiscale Uncertainty Propagation is Essential: A robust ICME framework requires propagating uncertainties not just within a single model but across the entire chain of models, from thermodynamics to microstructure evolution.
  • Sophisticated Sampling is Necessary: Naive sampling methods are computationally prohibitive for complex models. Advanced techniques like Gaussian copulas are vital for efficiently exploring high-dimensional parameter spaces while preserving statistical correlations.
  • Microstructure is Highly Sensitive: Even small variations in input parameters (e.g., elastic constants) can lead to qualitatively different microstructural morphologies and topologies.
  • Machine Learning for Understanding Complexity: The high dimensionality of the problem necessitates machine learning for analyzing the vast amount of generated microstructure data and identifying meaningful patterns and relationships.

Key Results

  • The framework successfully quantifies uncertainties in CALPHAD parameters and propagates them to the phase diagram and Gibbs free energy curves.
  • High-throughput phase-field simulations generated a dataset of 200,000 microstructures, revealing the impact of input parameter uncertainties on microstructure evolution.
  • Hierarchical clustering of the microstructures identified two main categories: "decomposed" and "not-decomposed," which are linked to different phonon scattering mechanisms relevant for thermoelectric performance.
  • Classification analysis demonstrated the ability to define decision boundaries in the parameter space that distinguish between these microstructure categories, offering insights for materials design.

Limitations and Future Work

  • Dimensionality of Input/Output Space: The paper acknowledges the challenge posed by the high dimensionality of both the input parameter space and the output microstructure space. While machine learning is used, further exploration of dimensional reduction techniques (e.g., PCA) is suggested for more rigorous classification.
  • Generalizability of QoIs: Developing general quantities of interest (QoIs) that can effectively characterize microstructures and their relationship to material properties remains an ongoing challenge.
  • Future Work: The authors plan to further explore machine learning frameworks for analyzing the microstructure dataset and to investigate the development of inverse maps connecting microstructure features to input parameters. They also aim to extend this framework to other materials systems and properties.

Summary updated

BibTeX
@misc{attari2019uncertainty,
  author = {Vahid Attari and Pejman Honarmandi and Thien Duong and Daniel Sauceda and Douglas Allaire and Raymundo Arróyave},
  title = {{Uncertainty Propagation in a Multiscale CALPHAD-Reinforced Elastochemical Phase-Field Model}},
  year = {2019},
  doi = {10.2139/ssrn.3427526},
  howpublished = {SSRN},
}
Figure