A GEFS-ensemble-driven dual-gate machine-learning system for 3-hourly quantitative precipitation forecasting over Brazil: independent dual-reference verification and calibration for flood early warning

Sub-daily quantitative precipitation forecasts (QPF) are a primary input to flood early warning, yet over Brazil they are hampered by convective, intermittent rainfall and a sparse observing network. We present a machine-learning post-processing system (V2.0C) producing 3-hourly QPF on a 0.25° grid by coupling a LightGBM precipitation-occurrence classifier with a Tweedie amount regressor and a dual-gate routing scheme with a physically guarded heavy-rain (≥5 mm) hard route. Its 71-dimensional feature set is dominated by GEFS ensemble statistics. The system is verified over two independent windows — April–May 2026 and the complete austral-summer convective season (October 2025– March 2026) — against the IMERG satellite product (areal) and the INMET automatic-station network (point). Against IMERG, V2.0C attains the highest correlation (0.324 and 0.286) and a heavy-rain CSI (≥10 mm) of 0.063–0.071, 3.3–4.4 times raw GFS, with stable month-to-month skill through the SACZ peak. A common mask analysis localizes the 60–65% ensemble-driven correlation gain to the southern third of the domain, where GEFS statistics are the only direct precipitation predictors available. Controls show that the guarded routing makes a high-recall heavy-rain expert safe to deploy: it adds 6–10% heavy-rain detection over a no-expert control, whereas unguarded use halves heavy-rain CSI. Gauge verification reproduces the heavy-rain advantage (CSI ≥10 mm an order of magnitude above raw GFS) and the wet bias, and the point–area correlation gap is explained by sub-grid representativeness. The detection-oriented bias is removable a posteriori: leakage-free quantile mapping eliminates it (bias +0.02 mm, MAE −69%) without reducing detection, supporting a dual-product strategy for flood-warning and hydrological applications.