Residual bias in a multiphase flow model: Calibration and prediction

When calibrated models produce biased residuals we assume it is due to an inaccurate conceptual model and revise the model, choosing the most representative model as the one with the best-fit and least biased residuals. However, if the calibration data are biased we may fail to identify an acceptable model, or may choose an incorrect model. Conceptual model revision could not eliminate biased residuals during inversion of simulated DNAPL migration under controlled conditions at the Borden site near Ontario, Canada. This paper delineates hypotheses for the source of bias, and explains the evolution of the calibration and resulting model predictions.