Soil microbial biomass (SMB-C) is one of the most frequently used parameters for the assessment of soil quality, but no threshold values have ever been proposed. We challenged the problem of a reliable numerical estimation of the SMB-C based on the knowledge of physicochemical soil properties. The aim was to evaluate artificial neural network (ANN) modelling for the prediction of SMB-C from a range of physical and chemical properties. The dataset used is composed of 231 vineyard soils of widely different characteristics and exposed to different temperature and moisture regimes. Each soil was described by ten physicochemical parameters: sand, clay, soil organic matter, total N, C/N ratio, pH, EC, exchangeable Na, active lime and total Cu. The ANN followed the topology: one input layer (1 to 11 nodes), one hidden layer (2 center dot n nodes) and one output node (SMB-C). Each soil sample was validated against the other 230 samples. The ANN model showed a much better fit than the linear model. The divergence between measured and predicted SMB-C was greatly restrained using the nonlinear approach, testifying the ability of the ANN to adapt to the highly variable dataset. The ANN analysis confirmed the primary importance of SOM for SMB-C prediction, being present in all of the best five models with the lowest root mean square relative error and in four out of five models with the lowest root mean square error. The prediction capability of SMB-C by ANN was limited at high SMB-C values, but the method can potentially be improved by expanding the dataset and introducing more parameters regarding soil physical properties and management.