For industrial bioprocesses requiring high cellular densities, the fed-batch operation is the preferred choice as it avoids the accumulation of major fermentation by-products due to metabolic overflow, increasing the process productivity. Reproducible operation at high cell densities is challenging (> 100 gDCW/L), which has precluded rigorous model evaluation describing this process. Here, we evaluated three phenomenological models and proposed a novel hybrid model including a neural network. For this task, we generated highly reproducible fed-batch datasets of a recombinant yeast growing under oxidative, oxygen-limited, and respiro-fermentative metabolic regimes. The models were robustly calibrated to these data using a systematic workflow based on pre- and post-regression diagnostics. Compared to the best-performing phenomenological model, the hybrid model substantially improved predictions by 3.6- and 1.7-fold in the training and test data, respectively. This study illustrates how hybrid modeling approaches can advance our description of complex bioprocesses that could support more efficient operation strategies.