NovaSpecNet is a deep learning framework designed to classify spectral phases of nova eruptions automatically. It combines 1D-CNNs for spectral feature extraction with BiLSTMs for temporal modeling, enhanced by multi-head attention mechanisms. The framework is trained using continuum normalization, spectral interpolation, and data augmentation techniques specific to astronomical spectroscopy, aiming for over 90% macro-averaged F1-score.
Key findings
NovaSpecNet automates classification of nova eruption spectral phases with high accuracy.
Combines 1D-CNNs for feature extraction and BiLSTMs for temporal modeling.
Utilizes multi-head attention mechanisms for enhanced classification.
Targets a classification performance exceeding 90% macro-averaged F1-score.
Limitations & open questions
The framework's performance is dependent on the quality and quantity of training data.
Real-time classification for time-domain surveys is a future goal that requires further validation.