The atmospheric boundary layer (ABL) is a turbulent “vertical conversation” that couples motions across height and time in strongly stability-dependent ways, yet most learning systems neither reveal nor exploit this structure. Here we test whether a deep learning (DL) model can recover these couplings directly from data and whether the resulting insight can guide model design. Using turbulence-resolving large-eddy simulations across neutral, weakly convective and convective regimes, we train a spatiotemporal DL model to predict near-surface windspeed from short sequences of horizontal windspeed at multiple heights. We interrogate trained networks with integrated gradients, temporal/spatial occlusion and region-of-interest attribution to map cones of influence that quantify how predictive information is distributed across height and lag and how it concentrates on coherent turbulent structures. The learned pathways reorganize with stability: neutral cases rely on earlier lags and near-surface streak-like shear structures; weakly convective cases broaden upward and redistribute attribution across intermediate lags; and convective cases become more recency-weighted while emphasizing plume and cell-boundary regions. A temporal-sampling sweep shows that skill improves when input spacing aligns with regime-dependent integral time scales. Together, these results show how explainable deep learning exposes information pathways in the ABL and provides actionable design rules for geophysical models.