smoothing before ICA (or any other analysis step) is primarily intended to increase signal-to-noise ratios. The idea here is to average out some of the unstructured noise in favor of the more structured noise, leading to more uniform components. I don't have first-hand experience with data you are describing, but could see how not smoothing WM could impact the quality of the retrieved components. I would suggest to just go ahead and try it out. The features that AROMA uses to select components should be robust enough. If not, you could apply AROMA to fully smoothed data and use those components and classification to filter out noise from your variably smooth data (using fsl_regfilt).

hth,

Maarten