Abstract: In tokamak plasmas, particles undergo transitions between trapped and passing states under magnetohydrodynamic instabilities. We developed a four-class bidirectional LSTM neural network with self-attention that categorizes particles as consistently passing, consistently trapped, passing-to-trapped, or trapped-to-passing. Using 11-dimensional time-series data (position, momentum, magnetic field, pitch angle parameter, and toroidal angle), the model generates attention weights that highlight critical moments in particle trajectories. Trained on a small labeled subset, our classifier efficiently scales to large-scale simulations, supporting detailed analysis of transition dynamics. We demonstrate this approach using alpha particles with significant finite Larmor radius effects, and the methodology can be adapted to different instability scenarios through retraining.