Neural network interatomic potentials (NNIPs) enable accurate and efficient Molecular Dynamics (MD) simulations, approaching the fidelity of first-principles calculations. Domain-specific NNIPs trained with active learning (AL) can reproduce material properties with high precision. Despite their potential, current AL methods face challenges: they might include undesirable structures in training datasets, and models like DeepPot-SE may struggle to differentiate between atoms of the same element type that have different roles in complex systems. This work addresses these limitations by introducing expert-guided filtering and atom-specific training. Expert-guided filtering uses domain knowledge to remove undesirable structures from training datasets generated by MD simulations. Atom-specific training prevents the model from confusing atoms by assigning different weights to atoms of the same element type that exhibit different behaviors. We evaluated the proposed approach using a DeepPot-SE model trained for the hydrogen molecule diffusion in a TiO2-SiO2 membrane. As a result, the model achieves improved force prediction accuracy, reducing the Root Mean Squared Error (RMSE) from 137.9 meV/Å to 98.9 meV/Å, which corresponds to an approximate 28% reduction. This enhancement enables the accurate reproduction of hydrogen molecule diffusion and the calculation of the self-diffusion coefficient.