Tokyo team's 220-metre global simulation erases spurious 'popcorn' rain

At 220 metres ^[3], a grid cell is about 4.5 times narrower than the 1-kilometre end of the storm-resolving range ^[16]. Over the same patch of ground, that comes to roughly 21 times as many grid columns ^[17]. "This finer resolution allows us to represent the internal structure of convective clouds, including individual updrafts and downdrafts, much more explicitly than before," said Shuhei Matsugishi, a project researcher at the University of Tokyo's Atmosphere and Ocean Research Institute ^[2][15].

Popcorn rain is the name for a long-standing bias in current storm-resolving models. These are intense, localised bursts of rain that appear in the simulation and do not occur in reality ^[5]. In the 220-metre run, according to the university's account, rainfall looked more realistic, without those unrealistically intense local events ^[13]. The thing this doesn't tell you is how large the improvement was, or whether the same day was rerun at coarser spacing, the control that would isolate grid spacing as the cause.

Matsugishi said computational cost was the hardest part of the work ^[14]. The simulation held almost 1 trillion three-dimensional grid points and covered eight hours of 5 August 2016 ^[7][8]. Spread over those eight hours, its electricity use works out to about five years of an average Japanese household's consumption for each simulated hour ^[18].

"At present, a global simulation at 220-meter (720-foot) resolution is far too computationally expensive to replace operational weather forecasting systems," Matsugishi said ^[10].

He sees the near-term use in research. "For now, these simulations are better suited to research experiments. For example, they can be used to investigate the detailed structure of tropical convection and heavy rainfall, and to provide a high-resolution reference against which coarser climate models can be evaluated and improved," he said ^[11].

I think the result supports a narrow claim, and a good one. In one simulated case, a known rain error in global models went away once the grid was fine enough to resolve individual updrafts ^[5][2][8]. Showing that a global model can forecast extreme local rainfall would need many more cases, scored against observations, at a cost operational centres could carry ^[10]. The work is published in Geophysical Research Letters ^[6].