Impact toughness is one of the most important mechanical properties of welds. For some materials and temperatures, impact tests have been found to predict the likelihood of brittle fracture more successfully than tension tests or other tests used in material specifications. Assessment and understanding of the impact properties of welds is central to the avoidance of catastrophic failure of welded steel structures. Neural network analysis has been used in the present work to address factors that are beneficial or detrimental to the low temperature impact properties of flux cored arc steel welds and to provide practical guidance on improving impact properties. Three main fields of variables were incorporated into the training of the neural network model: chemical composition, microstructural features, and non-metallic inclusions. Hardness and test temperature were also included as input variables. In total, 31 input factors were used in a neural network based on a back propagation algorithm, with impact energy being the output variable. Sensitivity analysis of the neural network model revealed that the direct influence of the non-metallic inclusion characteristics, density, and distribution on toughness was relatively small, whereas particular compositional and microstructural factors exerted significant effects. © 2003 IoM Communications Ltd.