End-to-End Phase Field Model Discovery Combining Experimentation, Crowdsourcing, Simulation and Learning

Md. Nasim, Xinghang Zhang, Anter El-Azab, Yexiang Xue

Introduction

In materials science, understanding the evolution of nano-structures like void defects under extreme conditions is critical. In situ radiation experiments generate massive amounts of high-resolution video data, but extracting meaningful physics models from this data remains a significant challenge. Traditional methods suffer from an annotation bottleneck, the "black box" nature of deep learning, and a lack of integrated tools for real-time collaboration between AI and human experts.

To address these challenges, we present PHASE-FIELD-LAB, an end-to-end platform designed to automate the discovery of phase field physics models. By integrating streamlined annotation, physics-informed neural networks, and interactive simulation, the framework significantly reduces manual effort while ensuring scientific validity.

The PHASE-FIELD-LAB workflow integrates experimental data analysis, human feedback, and experimental design.

The PHASE-FIELD-LAB Framework

The platform operates as a closed-loop system consisting of three primary modules:

1. AI-Powered Annotation

A major bottleneck in analyzing TEM videos is the pixel-wise labeling of thousands of frames. PHASE-FIELD-LAB utilizes the Simple Linear Iterative Clustering (SLIC) algorithm to group pixels into "superpixels." This allows human annotators to interactively mark or erase clusters rather than tracing individual pixels, transforming a tedious task into an efficient process.

The annotation workflow uses SLIC superpixels to allow rapid human marking of void defects.

2. Physics-Informed Learning

The core engine, Neuradiff, combines a "Recognition Net" for defect identification with a "Neural Differential Equation Net" to model temporal evolution. Unlike standard deep learning models that may violate physical laws, PHASE-FIELD-LAB incorporates a physics-informed loss function. This function penalizes parameters that fall outside expert-defined ranges, ensuring the discovered models are scientifically plausible.

3. Simulation and Visualization

The platform enables real-time visualization by solving the discovered Partial Differential Equations (PDEs) using finite difference methods. This allows researchers to compare predicted evolution directly against experimental footage, facilitating rapid model validation and feedback.

Key Results and Insights

The framework was validated using in situ TEM videos of Copper (Cu) under irradiation at 350°C, featuring both circular and elongated nano-void defects.

  • Efficiency and Accuracy: Compared to the standard VGG Image Annotator (VIA), PHASE-FIELD-LAB reduced labeling time by 50% for circular voids and 75% for complex elongated voids. Furthermore, it achieved a higher Intersection-over-Union (IOU) score (74-75% vs. 39-53%), as the superpixel approach better handles noisy boundaries.
  • Physical Consistency: While baseline models often produced "impossible" physics parameters, models trained within this framework consistently stayed within physically plausible bounds, maintaining a low Mean Squared Error (MSE) of approximately $3.3 \times 10^{-4}$.
  • Scientific Discovery: The tool enabled the discovery of non-monotonic evolution in nano-voids. Researchers observed that void sizes fluctuate in a "random walk" pattern rather than shrinking linearly—a phenomenon previously obscured by manual analysis noise.

Conclusion

PHASE-FIELD-LAB demonstrates that embedding domain expertise directly into the machine learning pipeline—both through interactive annotation and physics-informed constraints—can drastically accelerate scientific discovery. Future work will focus on expanding the framework to other PDE-based systems and increasing automation through active learning.

Summary updated

BibTeX
@inproceedings{nasim2024end,
  author = {Md. Nasim and Xinghang Zhang and Anter El-Azab and Yexiang Xue},
  title = {{End-to-End Phase Field Model Discovery Combining Experimentation, Crowdsourcing, Simulation and Learning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i21.30342},
}
Figure