I see that TGraphSmoother has Nadaraya-Watson kernel regression smoothing (and other forms of kernel-based smoothing). However, it looks to me like this will only work properly on a TGraph, not a TGraph2D.
Are there any tools appropriate for smoothing a TGraph2D? The data in my TGraph2D is noisy, so I need more than just interpolation. Or is the best choice to develop my own adaptation of the TGraphSmoother algorithm with support for TGraph2Ds?
There are some tool that can handle multi-dimensional pdf smoothing in RooFit, for example RooNDKeysPdf - see https://root.cern.ch/doc/master/classRooNDKeysPdf.html . To use it with a TGraph2D, the procedure would be to make a RooDataSet with a data point corresponding to each point in your graph, then use the createHistogram method to get a 2d histogram (TH2) of the smoothed distribution, from which you could re-create your graph. This might be overkill for your problem though..