Flash drought (FD), so-named for its abrupt and unforeseen onset, poses a significant challenge to forecasting, as current Numerical Weather Prediction (NWP) shows limited skill in the sub-seasonal to seasonal timescale (S2S, 2-week to 2-month range). Here, we present various data-driven deep learning (DL) frameworks designed to bridge this S2S FD forecasting gap and uncover underlying drought-inducing mechanisms via interpretability. We developed multiple spatiotemporal DL models (e.g., ConvLSTM, U-Net) and a Bayesian model averaging (BMA) ensemble to forecast pentad-scale (5-day) Standardized Soil-moisture Index (SSI), serving as the basis for subsequent FD identification. These models leverage diverse drought-related precursors, including compound drought-heatwave, evaporative stress, vapor pressure deficit (VPD), and vegetation conditions. Evaluating performance across basins with varied climate regimes, we found that Artificial-Intelligence-based methods offer enhanced SSI forecast reliability over NWP, particularly for weather-scale (1-3 pentads). Notably, the BMA ensemble provided reliable SSI forecasts up to 12 pentads (similar to 60 days, spanning the entire FD lifecycle), outperforming advanced physics-based NWP and pixel-wise benchmark models. Occlusion heatmap reveals that DL models leverage physically plausible precursors for predicting subsequent FD events. Through SHAP analysis, three primary FD-inducing patterns were identified: water-dominated (e.g., precipitation), energy-dominated (e.g., VPD), and multi-driver composite. Representative regions are arid climate, snow climate, and near-equatorial areas (e.g., equatorial and warm climate), with widespread interactions among drivers. This study demonstrates explainable DL models potent tools for advancing SSI forecasting and dissecting complex hydro-climatological drivers relevant to FD assessment, offering novel insights for improved early warning systems.