Liquid crystal thermography combined with transient conduction analysis is often used to deduce local values of convective heat transfer coefficients. Neural networks based on the backpropagation algorithm have been successfully applied to predict heat transfer coefficients from a given set of experimentally obtained conditions. Performance characteristics studied on numerous network configurations relevant to this application indicate that a 3-6-3-1 arrangement yields the least errors with convergence improving directly with both the global learning rates and those of individual layers.