Textual interpretation of transient image classifications from large language models.

Area of Science:

• Astronomy and Astrophysics
• Artificial Intelligence
• Machine Learning

Background:

• Astronomical surveys generate vast amounts of transient event data.
• Distinguishing genuine astrophysical signals from imaging artifacts is a significant challenge.
• Current methods like convolutional neural networks (CNNs) lack interpretability due to opaque latent representations.

Purpose of the Study:

• To evaluate the performance of large language models (LLMs) in classifying optical transient survey data.
• To assess if LLMs can achieve comparable accuracy to CNNs while providing interpretable outputs.
• To develop a framework for natural language-based classification and querying of astronomical transient candidates.

Main Methods:

• Utilized Google's LLM, Gemini, for classification on three optical transient survey datasets (Pan-STARRS, MeerLICHT, ATLAS).
• Employed a few-shot learning approach with 15 examples and concise instructions.
• Implemented a secondary LLM for assessing the coherence and refining the output of the primary classification model.

Main Results:

• LLMs achieved an average accuracy of 93% across diverse datasets, rivaling CNN performance.
• LLMs generated direct, human-readable descriptions for each transient candidate.
• The framework demonstrated effective iterative refinement and bypassed traditional training pipelines.

Conclusions:

• LLMs offer a promising, interpretable alternative for classifying astronomical transients.
• This approach enhances the understanding of transient events by providing textual explanations.
• LLM-based classification can bridge the gap between automated detection and human comprehension in astronomical data analysis.