This study explores the application of machine learning to enhance the precision of Scantech GEOSCAN Prompt Gamma Neutron Activation Analysis (PGNAA) for on-belt analysis of bulk materials. Focusing on the indirect measurement of Basicity in iron making, we aim to mitigate the uncertainty inherent in traditional analytical methods, which often rely on error-prone, multi-step calculations. Our approach leverages supervised learning, training a model on spectral analysis data, belt loads, statistical measures, and detector temperatures. Batch correlation serves as the loss function, and a defined acceptable response range acts as a filter, ensuring robust network performance. Traditional PGNAA techniques, including single-peak analysis and Monte Carlo simulations, can be challenging when dealing with elements exhibiting weak or unstable signals. Indirectly calculating these elements using intermediate ratios often leads to cumulative uncertainties. Our machine learning model circumvents this issue by establishing a direct relationship between the spectral analysis data and the Basicity parameter, eliminating the need for intermediate variables. The network architecture comprises multiple layers with tan-sigmoid and log-sigmoid activation functions, optimised for mapping input features to the target Basicity output. Rigorous filtering criteria were implemented during training to reject networks susceptible to overfitting or convergence to local extrema, ensuring reliable online performance. Experimental results demonstrate a significant improvement in accuracy, validated by strong correlation with laboratory data and reduced uncertainty. Further validation on independent data sets confirms the model's robustness across varying operational conditions. This method offers a promising avenue for enhancing the precision and operational efficiency of PGNAA systems, extending their applicability to values previously considered difficult to measure directly. Future work will focus on further network optimisation and validation across a wider range of material compositions.