**4. Discussion**

The proposed multi-task FCRN model was directly developed from the single-task FCRN of Wu et al. [23]. However, this is an open framework, enabling many other networks for seismic impedance prediction to also be explored. We believe seismic reconstruction is a task that is not limited to impedance inversion, and other deep learning tasks such as seismic fault interpretation can also benefit in similar way. Close inspection of Figures 3 and 4 indicates that the almost flat total loss variation range corresponds to a relatively large range of - *σpre*, *σrec* . This means that there maybe not a single optimal weighting for all tasks, which was also an observation of Kendall et al. [36].

For the first two experiments, we attempted to add six levels of Gaussian noise, having an SNR of 0, 5, 15, 25, 35, and 45 dB, to the test set to test the tolerance of the two networks to noise. The test results of the Overthrust model show that the accuracy of the impedance prediction of the multi-task network is lower than that of the single-task network under the 0 dB noise. A possible reason for this is that when we add noise to the test dataset, we should train the model again to obtain a new optimal weight for the two tasks.
