Error Modelling in Bayesian CSEM Inversion

Mis-characterisation of the noise has significant potential to disrupt reservoir parameter estimates and uncertainties in geophysical inversion. In all form of geophysical inversion, the "effective noise" used in the data misfit absorbs effects from approximate forward modelling in additional to environmental processing and measurement noise. For risk assessment, inversions require parameter uncertainties, and these are best approached from a Bayesian angle. But parameter uncertainty estimates are dependent on the noise model at leading order. Modelling noise in particular can be strongly correlated, and will corrupt parameter uncertainty estimates if the correlations are not taken into account. In CSEM inversion, structural and resistivity parameters can be particularly difficult to disentangle, and their separation is rather vulnerable to systematic components of the noise. We present several ideas to manage the effect, two of which are easily incorporated into standard optimization and sampling schemes.