PlanetNet-MMG: A Robust Multi-Modal Graph-Based Deep Learning Model for Exoplanet Candidate Classification

The search for exoplanets has advanced into the era of intelligent automation, yet most deep learning pipelines remain constrained to single-modality inputs or isolated views of astronomical data. We present PlanetNet-MMG, a novel multi-modal deep learning architecture that combines structured stellar metadata, raw lightcurve sequences, and graph-based relational context into a unified classification model. Our approach fuses three powerful encoders: a Tabular Transformer for domain-aware feature projection, a PatchGRU enhanced with a Vision Transformer (ViT) for learning fine-grained temporal patterns in segmented lightcurve patches, and a Graph PARE encoder that models inter-object similarity via a relational graph. Trained on a harmonized dataset derived from Kepler, TESS, and confirmed exoplanet archives, PlanetNet-MMG outperforms all state-of-the-art baselines, achieving a peak test accuracy of 90.4% and a class-averaged AUC of 0.973. Extensive experiments across 10–100 epochs and comparative evaluations against Astronet, ExoNet, OsbornNet, GCN (Lu), and ExoMiner confirm the effectiveness of our multimodal fusion. We further provide interpretability through attention overlays, t-SNE projections, and confidence histograms, reinforcing PlanetNet-MMG’s transparency and reliability for scientific discovery in astrophysics.