Use of a multiproduct pipeline operation inevitably produces an interface mixture or TRANSMIX which either has to be returned to refinery to be re-processed or downgraded into one of the adjacent batches, which is an additional cost as product is lost. In this paper, a Non-linear Autoregressive External Inputs (NARX) model is used for estimating the mixing volumes. Comparison with other models shows in the literature, the proposed neural network is reliable over a wide range of flow conditions and data. The model was fed using field data, 10 hidden layers and three subsets for the training were used: 50%, 65% and 85%. The NARX was capable of simulating the system and demonstrate a good performance comparable with the traditional methods and real data with a maximum error of 4% at 85% training data. (C) 2018 Elsevier Ltd. All rights reserved.