Accurate void fraction prediction in vertical rod bundles is crucial for determining two-phase flow parameters, ensuring the safety of nuclear reactor cores. The drift flux model, commonly used for void fraction prediction, requires developing its constitutive correlations. While ANN models outperform traditional fitting methods, their physics-independent nature and lack of inherent feature selection limit their reliability in critical applications. This study proposes a Knowledge-Informed Neural Network (KINN) to develop drift flux constitutive correlations by embedding domain knowledge into the ANN structure as custom hidden layers. Two KINN frameworks have been introduced each incorporates different feature selection techniques. KINN-SHAP is a two-stage approach that use SHAP model for identifying the top k features and update the feature set for second training stage. KINN-DN is an automatic approach that use an innovative single neuron fully attached to input feature for filtering out less informative features in a single training stage. Both Models show close generalization on test set (roughly MRE = 0.064) with dropping 4 features besides superior performance against other drift flux models in literatures.