Setting the standard for machine learning in phase field prediction: a benchmark dataset and baseline metrics
Abstract Phase field models are an important mesoscale method that serves as a bridge between the atomic scale and the macroscale, used for modeling complex phenomena at the microstructure level. Machine learning can be employed to accelerate these simulations, enabling faster and more efficient ana...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Nature Portfolio
2024-11-01
|
| Series: | Scientific Data |
| Online Access: | https://doi.org/10.1038/s41597-024-04128-9 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract Phase field models are an important mesoscale method that serves as a bridge between the atomic scale and the macroscale, used for modeling complex phenomena at the microstructure level. Machine learning can be employed to accelerate these simulations, enabling faster and more efficient analyses. However, the development of new machine learning algorithms depends on access to extensive datasets. This work introduces an accessible and well-documented dataset aimed at benchmarking new machine learning algorithms. We validate the dataset with a benchmark using U-Net regression, a widely used neural network architecture. Although direct comparisons are limited by the lack of existing benchmarks, our model’s error metrics are competitive with previous work and generalize across multiple domain sizes. This contribution provides a valuable resource for future efforts in machine learning model development for phase field simulations and demonstrates the potential of U-Net regression, highlighting the scope for novel method development in this area. |
|---|---|
| ISSN: | 2052-4463 |