Two-dimensional porous boron nitride (BN) is widely used in electronics, catalysis, and membranes. Its thermal conductivity is influenced by hole density and distribution. This study uses molecular dynamics simulations with a hybrid machine learning approach to predict BN's thermal conductivity. The hybrid model, combining convolutional neural networks and multilayer perceptrons, achieves high accuracy (RMSE = 0.01, R² = 0.99). A novel inverse design method links hole distribution to enhanced conductivity, optimizing designs with just 1024 samples. These findings demonstrate the power of machine learning in advancing physical insights and solving complex design challenges efficiently.