Artificial neural networks were used to model biosurfactant production process by Candida lipolytica UCP 0988, using diesel oil as carbon source. The modeling was carried out using data of 24 full factorial design composed by 20 runs, including four replicates at the center point - having the concentrations of sea water, urea, ammonium sulfate and potassium phosphate monobasic as independent variables and as response variable the surface tension of biosurfactant production media filtrates with 168 hours. Several models were developed using different databases, topologies, transfer functions and training algorithms. A model with topology 4-4-1 - having concentrations of sea water, urea, ammonium sulfate and potassium phosphate monobasic input variables and surface tension as output variable - trained with the backpropagation algorithm based on Levenbeg-Marquadt was selected as the best performance in the estimation of surface tension. The root mean square error and correlation coefficient for the validation set of this model were 0.25433 and 0.97433, respectively. These results confirm the generalization ability and efficiency of artificial neural networks as tools for development of biosurfactant production media
