Setting the standard for machine learning in phase field prediction: a benchmark dataset and baseline metrics

Area of Science:

• Computational Materials Science
• Mesoscale Modeling
• Machine Learning Applications

Background:

• Phase field models are crucial mesoscale tools for simulating microstructure evolution.
• Accelerating these simulations with machine learning (ML) requires extensive, standardized datasets.
• Current ML algorithm development is hindered by a lack of accessible benchmarking data.

Purpose of the Study:

• To introduce a novel, well-documented dataset for benchmarking ML algorithms in phase field modeling.
• To provide a resource facilitating the development and comparison of new ML methods.
• To validate the dataset's utility through a benchmark study.

Main Methods:

• Development of an accessible and documented dataset for phase field simulations.
• Benchmarking using U-Net regression, a convolutional neural network architecture.
• Validation across multiple domain sizes to assess generalization.

Main Results:

• The U-Net regression benchmark achieved competitive error metrics.
• The model demonstrated generalization capabilities across different simulation domain sizes.
• The dataset proved effective for validating ML performance in phase field contexts.

Conclusions:

• The introduced dataset is a valuable resource for advancing ML in phase field simulations.
• U-Net regression shows promise for accelerating and improving microstructure analysis.
• The work highlights opportunities for developing novel ML methods in this domain.