Schematic representation of the CNN model used in this work. Four-channel magnetic field images (B_los, B_r, B_p, and B_t) with an input size (4, 512, 512) are processed through convolutional and residual blocks, followed by global average pooling and sigmoid classification to predict the probability of flare occurrence within 24 hr. The final convolutional layer, which is used as the target layer for Grad-CAM, is highlighted.