Training and validation performance (RMSE and balanced accuracy, both measured on the final state) of different approaches for planetary collision treatment. Classification is performed as a post-processing step on top of predicted masses. We assume the SPH simulation data to be the ground truth. For single-task learning (lower half of the table), each entry corresponds to the performance of individually trained ML models and column headers indicate optimized tasks, respectively. We use data from Timpe et al. (2020b) to obtain results in the lower half. Best results are indicated in bold, whereas * indicate statistically significant results according to a Wilcoxon test (comparing FFN and RES). Our proposed RES model outperforms the other baseline methods in most experiments.