The Rod Bundle Heat Transfer (RBHT) project (2020-2023), organized by the OECD/NEA, is dedicated to produce novel experimental data using the Separate Effect Test Facility present at Penn State University in the US. The project aims to benchmark thermal-hydraulic codes in predicting reflood transients under varying boundary conditions. This paper reports on the prediction capabilities of the TRACE system code for selected tests of this project. Uncertainty in selected figure of merits (e.g., temperature profiles, pressure drops) are accounted for using the probability density function of influential thermal-hydraulic parameters that were calibrated on similar experiments pertaining to the FEBA project (1980s). Despite considering these uncertainties, TRACE predictions are significantly biased. As an example, the maximum rod clad temperature at high elevation is consistently overpredicted by several hundred Kelvins. In this paper, we quantify the time-dependent biases and aim at correcting TRACE predictions for experiments performed with different boundary conditions. For that purpose, we train a set of Gaussian Process metamodels on the biases using a subset of RBHT experiments with known boundary conditions and validate the metamodels predictions for other boundary conditions by comparing TRACE bias-corrected predictions to experimental values. Although effective, this approach does not alleviate concerns about the physical models of TRACE for reflood transients and the need to improve them.