Project Description

Modeling Multi-Band Lightcurves for Astronomical Surveys
Project Description
Astronomical surveys involve collecting observations across multiple wavelengths, referred to as bands. Each band corresponds to a specific segment of the electromagnetic spectrum, offering diverse information about celestial bodies. This multi-band approach enables astronomers to deduce key properties of astronomical objects, such as their temperature, composition, and age.
Developing models capable of accurately representing astronomical lightcurves—time- series data of brightness variations—becomes particularly challenging when working with multi-band observations. A simple strategy, based on Astromer [1], is to input multi-band data into a transformer-based encoder as multi-dimensional features. However, the use of different physical filters across observation bands often results in data acquisition occurring at different times, leading to asynchronous datasets. Conventional transformer architectures are ill-suited to address this issue.
This project seeks to investigate alternative methods for effectively processing multi – band lightcurve data. Two specific approaches are proposed:
• Late Fusion with Embedding Mixing: In this method, transformer encoders are independently trained for each band, inspired by the architecture of Astromer. Instead of merging multi-band data initially, embeddings are generated for each band separately. These embeddings are then combined using techniques like embedding mixing [2] during the final task, such as classification or regression. This decoupled processing accommodates asynchronous data while preserving the benefits of transformer-based encoders.
• Multi-Attention Layers: This approach focuses on extending the transformer architecture with multi-attention mechanisms. These layers facilitate the model’s ability to simultaneously attend to temporal and inter-band relationships, potentially capturing the intricate dependencies in asynchronous observations.
Further areas for exploration include:
• Examining self-attention mechanisms tailored specifically for asynchronous data.
• Applying the proposed models to real-world astronomical datasets to evaluate their practical effectiveness.
浙大学霸代写 加微信 cstutorcs
References
[1] ASTROMER—A transformer-based embedding for the representation of light curves. Astronomy & Astrophysics. Retrieved from https://www.aanda.org/articles/aa/pdf/2023/02/aa43928-22.pdf
[2] Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., … Polosukhin, I. (2021). Attention is all you need. arXiv:1706.03762. Retrieved from https://arxiv.org/abs/1706.03762
• Single-band ASTROMER code: astromer-science/python-library: Python library including the stable ASTROMER
• Multi-band Dataset for pre-training (MACHO – large unlabelled dataset): macho_filtered_400
• Multi-band Dataset for fine-tuning and classification (Alcock – small, labelled dataset): alcock
Use the single-band ASTROMER code to build a multi-band ASTROMER model with this architecture.
Be as creative as possible. Bonus points for novel ideas or architectural improvements.
Code Help
• Achieve at least the performance metrics of the single-band model in the ASTROMER paper!
• Achieve at least 70% validation accuracy on classifying the Multi-band Alcock dataset.
Computer Science Tutoring