THE NEXT-GENERATION NWP MODELS.The meeting provided a good opportunity for various scientists to describe recent modeling improvements. At the Japan Meteorological Agency (JMA), the improved nonhydrostatic model has led to the better prediction of heavy rainfall in Japan. That was the conclusion of Kazuo Saito of the Meteoro- PROBLEMATIC ISSUES. Problems with physics parameterizations in the models that emerged during the workshop include the following: resolution dependency of each physical process, deterministic versus stochastic approaches, and use of observations.Resolution dependency of physics. Physical parameterization schemes developed at one scale may no longer be valid at smaller scales, because computer power increases and grid sizes decrease. Cumulus schemes are a current example, and PBL schemes may be another in the future. Chun stated that GWDC was needed in mesoscale models as long as the need for a cumulus parameterization scheme was valid. Hong stated that parameterized convection could still be useful even at grid sizes of a few kilometers because it helps initiate mesoscale circulation. Dudhia and Lee stated that 4-km grid spacing could be sufficient for microphysics alone in strongly forced midlatitude convective systems. Even with 5-km grid spacing, a convective parameterization is required to properly simulate convective rains and keep radiation and convection in equilibrium. Based on the CRM results and data assimilation issues in high resolutions, Tao argued that 3 km might be the cutoff for cumulus schemes. Hong stated that PBL and shallow convection might be valid with a coarser resolution than is needed for LES. Work at NCAR has shown that the PBL scheme might still give reasonable results at 50-250-m grid sizes by suppressing resolved eddies, thereby eliminating the problem of double counting. Tao said we should separately decide which schemes to apply at various scales (5, 3 km, etc.), while Chen pointed out the need to develop new schemes that gradually change with scale.Deterministic versus ensemble versus stochastic approaches. Deterministic approaches to modeling imply refinement of parameterizations, addition of complexity, and superparameterization, whereas an ensemble approach can be based on uncertainties in initial conditions or physics schemes. Meanwhile, stochastic approach incorporates randomness, such as Grell's ensemble cumulus approach, the PDF approach, or random number uses. For a given computing resource the choice can be as simple as deciding whether to use high-resolution deterministic or lower-resolution ensemble approaches. A recent convection workshop (Tao et al. 2003) found a need to improve ID schemes, a need for multimoment approaches in these schemes, and considered superparameterization a promising approach. Hong pointed out that ensemble means produce better large-scale patterns (such as 500-hPa height), but deterministic method produces better precipitation, such as heavy rain. Hong stated that a stochastic approach might be able to combine advantages ...