IFP Energies Nouvelles – Stage – Chemometrics and Machine Learning for the prediction of petroleum cuts properties

The classic analytical workflow to obtain the properties of the produced petroleum cuts during tests on pilot plant units at IFPEN includes a first stage of distillation of the total effluent in the different cuts of interest. Each cut is then analyzed following normalized analytical methods to obtain their properties. It’s a time- and volume-consuming workflow, and the use of a faster alternative approach is of great interest. Previous studies, carried out at IFPEN, have shown that it is possible to predict these properties with data coming only from the analysis of the total effluent (therefore, without distillation). These analyzes generate either univariate or multivariate data. The methods used to exploit these data therefore range from polynomial regressions to neural networks, including multivariate regressions (PLS). The challenge is then to combine datasets with various dimensions.

The main objective of this internship is to combine chromatographic data (multi-dimensional dataset) with properties of the total effluent (one-dimension dataset for each variable/property) and process data such as pressure, temperature, … (one-dimension dataset for each variable/property). Usually, chemometric methods are used to extract information from chromatographic data and an important step is the preprocessing of these data (correction of the baseline, alignment of the data…). Dimensionality reduction approaches will be investigated, as well as data fusion methods and neural networks