State of the art thermodynamic models, such as the Perturbed-Chain Statistical Associating Fluid Theory (PC-SAFT), require a thorough parametrization (three pure-component parameters for nonassociating molecules) of the molecules considered. In our previous work (J. Habicht, C. Brandenbusch, G. Sadowski, Fluid Phase Equilibria, 2023, 565, 113657), we introduced a Machine Learning approach for a predictive parametrization of nonassociating components. Within this approach, training is performed using a Huber-loss function, comparing the ML-predicted parameter set with the original one, e.g., from literature. However, often multiple pure-component parameter sets exist for one molecule. This fact makes the training to only one "true" parameter set questionable. Within this work, we thus performed a detailed analysis on the fact of multiparameter set existence. We further expanded our ML-approach by developing a choice of two physics-informed loss functions that allow for the consideration of multiple "true" parameter sets during training. Results indicate that reliable pure-component parameters have a certain orientation when plotted in the three-dimensional parameter space. The results of this work will lead to a more reliable ML-based parametrization and ensure the prediction of optimized purecomponent parameters for a given molecule.