Global Diurnal Precipitation Cycle in the AI Model GraphCast and a 5‐km Unified Model: Challenges and Opportunities

Emanuele S. Gentile, Kieran M. R. Hunt, Lorenzo Tomassini, B. Harvey, O. Martínez-Alvarado

Geophysical Research Letters · 04 May 2026 · Met Office, Princeton University, University of Reading

In one sentence. GraphCast inherits and amplifies ERA5's premature afternoon peak and weak nocturnal boreal-summer rainfall, while a 5-km Met Office model improves night-time precipitation.

Abstract

Abstract This study evaluates the ability of the AI weather forecast model GraphCast to reproduce the global diurnal cycle of boreal summer precipitation, comparing it with Integrated Multi‐satellite Retrievals for GPM (IMERG) satellite observations, the ERA5 reanalysis, and an experimental global 5‐km Met Office Unified Model (UM) which is convection permitting but still retains an active scale‐aware parametrization. ERA5 captures large‐scale rainfall patterns but exhibits a premature afternoon peak and excessively weak nocturnal precipitation over land compared to IMERG. GraphCast, while reproducing realistic mean spatial rainfall distributions, inherits, and amplifies these timing and amplitude biases, concentrating precipitation near‐midday and producing very little nocturnal signal. The 5‐km UM improves nocturnal precipitation across many regions but overestimates rainfall over oceans and initiates convection too early over land, likely due to the still active convective parametrization. Our analysis shows that these contrasting behaviors highlight both challenges and opportunities for improving precipitation prediction through hybrid AI‐physics approaches.

Labels

Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.