Oil and gas industry has increasing demands for big data and big compuing capabilities.
Cementitious materials and functional oxides
There are many situations in science and engineering where conventional modelling and simulation approaches cannot deliver sufficient insight to support quantitative predictions of the behaviour of a system of interest. This is typically the case in situations of a highly multivariate nature—that is, of high dimensionality—where properties may be sensitively dependent on the values of many variables.
For example, in the search for oil and gas, one needs to perform extensive drilling operations, an integral part of which is “cementing”, during which the annulus formed between the steel casing dropped into the wellbore is filled with cement. The challenge is to pump the cement slurry to the depth required, whereupon it should set in the annulus as rapidly as possible. Premature setting of the cement, or failure to set, may both lead to very substantial delays in the drilling process, preventing production of the oil or gas, and causing increasing costs as expensive personnel and equipment are kept on rigs for lengthy periods of time.
Cement is a highly complicated material, of inexact chemical composition and properties which vary significantly from batch to batch, and between manufacturers, as well as responding differently to various additives and the local water with which it is mixed at the wellside. No simple mechanistic models exist capable of predicting the performance properties of the material, such as its thickening and setting times. The thickening time is of central importance in the oilfield context: it defines the time after mixing with water beyond which the cement is no longer pumpable. Retarders are usually added to the slurry to extend the thickening time.
Our approach: Big data analytics and machine learning
We have solved this problem through the use of big data analytics methods based on artificial neural networks. By constructing a data base comprised of a wide range of cements, we were able to show that the rapidly performed diffuse reflectance Fourier Transform Infra-Red (FTIR) spectrum of a cement powder can be used to predict the thickening time of the slurry made up from the powder. This remarkable and surprising finding is due to the fact, demonstrated in our work, that the FTIR spectrum contains information on the chemical composition and the particle size distribution of the cement powder. These indices (representing some 40 independent variables) are known to be the key ones which determine the reactivity of the cement when mixed with water and additives.
With a data base established and a trained neural network in place, the prediction of thickening times of cement slurries becomes a very rapid process: one acquires the FTIR spectrum of the powder and passes its digitised form through the neural network, which returns a prediction of the thickening time and other properties of interest essentially instantaneously. This permits a dramatic acceleration of the selection of the appropriate slurry, with ensuing reliability and hence productivity of the well site operations, reducing the occurrence of major operating failures.
The figure above shows the measured and predicted thickening-time curves for a typical Oil-field Cement. The upper continuous curve displays the artificial neural network prediction for a neat slurry, and the continuous lower curve shows similar predictions for a retarded slurry.
Similar machine learning based methods can be applied to the discovery of a range of materials, including the optimisation of the composition of functional metal oxides for applications such as solid oxide fuel cells.
Similar techniques can be applied to very wide ranging problems for which either the data already exists in digital form, or can be readily collected for processing and analytics – from retail and financial services, through smart cities to biomedical informatics.
P. V. Coveney, P. Fletcher, T.L. Hughes, “Using Artificial Neural Networks to Predict the Quality and Performance of Oil-Field Cements”, AI Magazine, 17, (4), 41- 53 (1996)
D. J. Scott, P. V. Coveney, J. A. Kilner, J. C. H. Rossiny, and N. Mc N. Alford, “Prediction of the functional properties of ceramic materials from composition using artificial neural networks”, Journal of the European Ceramic Society, 27, (16), 4425-4435, (2007)
D. Scott, P. V. Coveney, and S. Manos, “Design of Electroceramic Materials Using Artificial Neural Networks and Multi-Objective Evolutionary Algorithms” Journal of Chemical Information and Modeling, 48, (2), 262-273, (2008)