Abstract
Understanding the evolution of defects in materials under extreme conditions is vital for nuclear energy sustainability. However, analyzing high-resolution video data from in situ radiation experiments is traditionally a manual, time-intensive process. We present PHASE-FIELD-LAB, an integrated platform that automates phase field model discovery. By combining AI-assisted superpixel annotation with physics-informed machine learning, the system reduces manual labeling time by up to 75% while maintaining physical consistency. The platform enabled the discovery that nano-void defects in irradiated copper fluctuate in size, rather than shrinking monotonically as previously believed.
Introduction
In materials science, the behavior of nano-void defects under high heat and radiation is a critical factor in material degradation. Modern Transmission Electron Microscopy (TEM) generates massive datasets that create an analysis bottleneck. Traditional manual annotation takes months, and standard "black-box" AI models often produce results that violate basic physical laws. There is a clear need for a bridge between raw experimental data and interpretable Partial Differential Equation (PDE) models.
The PHASE-FIELD-LAB Framework
The authors propose PHASE-FIELD-LAB, a platform consisting of three integrated modules: an AI-based annotation tool, a physics-learning module with human-in-the-loop feedback, and a simulation/visualization engine.
The framework creates a closed loop where experimental data is analyzed to learn physics models, which are then validated by experts to design new experiments. The learning module utilizes a specialized architecture called Neuradiff, which combines a "Recognition Net" for defect identification and a "Neural PDE Net" for evolution simulation. Crucially, penalty terms are added to the loss function to ensure learned parameters remain within physically realistic ranges provided by domain experts.
AI-Assisted Annotation
A core innovation of the platform is the annotation module. Instead of manual pixel-by-pixel tracing, it utilizes the Simple Linear Iterative Clustering (SLIC) algorithm to segment images into "superpixels."
Researchers can interactively mark or erase these segments, drastically accelerating the labeling of complex, elongated defects. This approach turns a tedious drawing task into a simple selection task, which is particularly effective for the irregular shapes found in TEM imagery. Compared to standard baselines like the VGG Image Annotator (VIA), this method achieved significantly higher Intersection-over-Union (IOU) scores ($74.4\%$ vs $39.1\%$ for mixed defects).
Results and Scientific Discovery
The platform achieved a 75% reduction in labeling time for elongated voids and a 50% reduction for circular voids. While baseline models frequently violated physical constants, the PHASE-FIELD-LAB learning module satisfied 100% of physical constraints across multiple runs.
Beyond efficiency gains, the platform facilitated a significant scientific discovery. For years, it was accepted that nano-voids shrink monotonically under certain conditions. By enabling the analysis of long-duration videos (a task previously too labor-intensive to perform), the researchers discovered that these defects actually exhibit "random walk" size fluctuations. This finding was subsequently confirmed by manual measurements and phase field simulations.
Conclusion
PHASE-FIELD-LAB demonstrates the power of combining human expertise with physics-constrained AI. By automating the path from raw data to physical models, the platform removes a major bottleneck in materials science. This approach is broadly applicable to any scientific field governed by PDEs, such as fluid dynamics or collective cell migration, offering a path toward faster, more reliable discovery in complex physical systems.
Summary updated