Figure 8 shows the accuracy of our generated model compared to the validation data (that is, the 30% test data from benchmarking). We use Mean Absolute Percentage Error (MAPE) as our evaluation metric. As you can see, the predicted memory consumption results (from the Pandora symbolic regression tool) for the various image dimension/batch sizes match closely to the measured data. On average, the model had a MAPE of 4.15%. Compared to results generated by the base analytical model in Case-study application: 3D U-Net and Figure 5, the performance model generated through symbolic regression performed significantly better.