Laboratory testing of unsaturated soil shear strength parameters is often time-consuming, expensive, and requires specialized equipment. This study explores Artificial Neural Networks (ANNs) as an alternative, systematically optimizing the predictive model through a novel, multi-stage analysis investigating activation functions and iteratively tuning hidden layer counts and neuron numbers. A comprehensive evaluation of 195 network configurations was conducted using a dataset of 490 points compiled from 14 soil types, primarily fine-grained soils. Modeling identified the Bayesian regularization (TRAIN BR) function as superior (R=0.97R=0.97R=0.97). Subsequent expansion to three, four, and five hidden layers (with neuron counts from 50 down to 10) determined the most effective architecture. The four-layer Multilayer Perceptron (MLP) network emerged as the optimal configuration, achieving exceptional performance with an overall R2R^2R2 value of 0.98. Model validation utilized rigorous approaches. Initially, reserved samples confirmed the four-layer network’s high accuracy for cohesion. Secondly, predictions were compared with established empirical methods, demonstrating significantly higher accuracy. Finally, five independently prepared samples tested via in-house Direct Shear Testing further validated the model’s reliability. This external validation confirmed close agreement, showing prediction errors ranging from 1% to 11% for friction angle and 3% to 14% for cohesion. While further validation using a wider diversity of soil types and a larger external sample size is required to confirm generalizability, these results firmly establish ANNs as a powerful, accurate, and cost-effective tool for geotechnical engineers providing reliable estimates of unsaturated soil shear strength parameters.