Speaker
Description
We employ machine learning methods to analyze ALMA archive data within the Interstellar Medium and Star Formation categories. To manage the massive data volume, we utilize DBSCAN to extract subcubes including signal detections from the archive datacubes. As a result, we reduced the data size to approximately 5% of its original volume, comprising a total of ~250,000 FITS cubes. We aim to develop an image encoder model capable of interpreting astronomical 3D datacube structures for future downstream tasks. Specifically, we implemented a ResNet-34 architecture and applied the SimCLR framework for self-supervised training with data augmentation. Finally, Principal Component Analysis (PCA) was conducted to evaluate training quality and to verify whether the model is learning genuine physical structures rather than non-physical artifacts or statistical features.
| Participate the oral/poster presentation award competition | No |
|---|