Objective Evapotranspiration is the core link in the coupled atmosphere–water–heat system, and assessing its spatiotemporal variability is essential for sustainable water-resources management in inland arid regions. Methods:Taking the Manas River Basin—the “magnifying glass” of hydro-atmospheric coupling in the Central Asian arid belt—as the study area, we analyzed potential evapotranspiration (PET) from 1980 to 2023 using the Theil–Sen median slope and Mann–Kendall tests. To probe the mechanisms underlying PET evolution, we fused GeoDetector with random-forest regression, creating a “microscope” ensemble framework. Results:Basin-mean air temperature increased at 0.179 °C dec -1, whereas mean wind speed and relative humidity declined significantly by −0.040 m s -1 dec -1 and −0.914 % dec -1, respectively; sunshine duration decreased by −0.148 h dec -1. PET showed a significant upward trend of 1.319 mm yr -1, with the fastest rise in desert zones (1.388 mm yr -1), followed by oasis plains (1.292 mm yr -1) and mountains (1.285 mm yr -1). Spatially, PET was higher in the north than in the south and higher in the west than in the east, ranging from 120.92 mm to 1081.15 mm. GeoDetector identifies temperature, precipitation, and wind speed as the top three factors, while Random Forest recognizes temperature, sunshine hours, and relative humidity as the top three. Both methods place temperature as the primary factor, but they show complementary differences in the secondary factors. These findings furnish a scientific basis for designing crop-irrigation schedules and enhancing water-resource efficiency in arid inland irrigation districts.