*3.3. Experiment on the Volve Model*

The field Volve data used in this paper are from the open-source code of Das et al. [19]. The Volve field shown in Figure 14a is located in offshore Norway and is a clastic reservoir. There are 1300 labeled traces in this dataset, which were generated using the data augmentation method in [19] based on the statistical characteristics of the well position log data shown in Figure 14b. We randomly selected 750 traces as the training set and the remaining 550 traces were set to be the validation set. We used the single true well log data to test the performance of the two networks. Unlike the previous two datasets, the kernel size was set as 80 to adapt the 160 time sampling points of the Volve model. The number of epochs was set to 500, and the other hyperparameters and network structure were set to be the same as in the former two models. Similarly, when the validation loss is minimal during the 500 epochs, the corresponding *σpre* **=** 0.829 and *σrec* **=** 0.828, then we can calculate that the optimal weight for impedance prediction and seismic data reconstruction as 0.499:0.501. Figure 15 shows the predicted results of impedance prediction by the single-task model (blue) and the multi-task model (green) under the optimal weight. We can see that the impedance predicted by the multi-task model under the optimal weight matches the true impedance (red) better than the single-task model, especially in the time range between 40 and 50 ms. The blue and green dotted lines in Figure 15 represent residuals between the true impedance and the impedance predicted by the two networks. The MSE of the impedance predicted by the multi-task model under the approximate optimal weight is 0.0239, whereas the MSE predicted by the single-task model is 0.0352. In addition, the PCC between the ground truth and the values predicted by the multi-task model and the single-task model are 0.872 and 0.832, respectively. The above two metrics prove the superiority of the multi-task model under the optimal weight. It is worth noting that the test MSE and PCC in reference [19] are 0.0255 and 0.82, which is comparable with our single-task model, but is slightly inferior to our multi-task model. The Volve field data test shows that the proposed method in this paper improves the accuracy of impedance prediction on real seismic data.

**Figure 14.** (**a**) The Volve field, as shown on a map, is located in the offshore North Sea area and (**b**) seismic data with a well trajectory from the Volve field (figure is from reference [19]).

**Figure 15.** Impedance predicted for the well location of the Volve model by the two networks.
