Digital twin applications for water resource recovery facilities require frequent model recalibration to maintain predictive accuracy under dynamic operational conditions. Current calibration methodologies face critical limitations: manual protocols demand extensive expert intervention and iterative parameter adjustments spanning weeks to months, while automated optimization algorithms impose elevated computational burdens that struggle to converge within practical timeframes. This study introduces Expert Systems with Neuro-Evolution of Augmenting Topologies (ES-NEAT), integrating genetic algorithms, artificial neural networks, and transfer learning to preserve and transfer calibration knowledge across recalibration scenarios. Application to the full-scale Eindhoven WRRF over six months, calibrating 33 parameters across multiple temporal scenarios, demonstrated 72.1% and 49.0% Kling-Gupta Efficiency improvement over manual calibration for tank-in-series and compartmental model structures, respectively. Transfer learning reduced subsequent recalibration computational time by 50-70% while maintaining prediction accuracy, transforming initial 10-12 hour optimizations into 3-6 hour recalibrations through knowledge preservation. Performance degradation analysis established 2-month optimal recalibration intervals under observed operational variability. The methodology enables practical digital twin implementation by transforming recalibration from episodic expert-dependent burden into continuous, automated learning processes operating at timescales matching operational decision-making needs.