ABSTRACT:We present numerical-model experiments to investigate the dynamics of tropical-cyclone amplification and its predictability in three dimensions. For the prototype amplification problem beginning with a weak-tropical-storm-strength vortex, the emergent flow becomes highly asymmetric and dominated by deep convective vortex structures, even though the problem as posed is essentially axisymmetric. The asymmetries that develop are highly sensitive to the boundarylayer moisture distribution. When a small random moisture perturbation is added in the boundary layer at the initial time, the pattern of evolution of the flow asymmetries is changed dramatically, and a non-negligible spread in the local and azimuthally-averaged intensity results. We conclude, first, that the flow on the convective scales exhibits a degree of randomness, and only those asymmetric features that survive in an ensemble average of many realizations can be regarded as robust; and secondly, that there is an intrinsic uncertainty in the prediction of maximum intensity using either maximumwind or minimum-surface-pressure metrics. There are clear implications for the possibility of deterministic forecasts of the mesoscale structure of tropical cyclones, which may have a major impact on the intensity and on rapid intensity changes.Some other aspects of vortex structure are addressed also, including vortex-size parameters, and sensitivity to the inclusion of different physical processes or higher spatial resolution. We investigate also the analogous problem on a β-plane, a prototype problem for tropical-cyclone motion. A new perspective on the putative role of the wind-evaporation feedback process for tropical-cyclone intensification is offered also.The results provide new insight into the fluid dynamics of the intensification process in three dimensions, and at the same time suggest limitations of deterministic prediction for the mesoscale structure. Larger-scale characteristics, such as the radius of gale-force winds and β-gyres, are found to be less variable than their mesoscale counterparts.