Machine Learning to Predict Scale Inhibitor Squeeze Treatments Lifetime

Machine Learning (ML) involves the training of a model to make predictions based on data. A model might be described by several parameters, normally, weights and biases, where training is the process to determine the ideal parameters that comprise model, based on the dataset (a collection of raw data arranged in columns). ML can be divided in four main types, supervised learning, where training is based on labelled data; unsupervised learning, based on unlabelled data, where the goal is to find hidden patterns or structures within the data; semi-supervised learning, it uses a combination of labelled and unlabelled data; and finally, reinforcement learning, the learning is through trial and error, where the model receives reward by taking desirable actions and penalties for undesirable, the goal is to maximise the reward. There are a great variety of ML techniques for various tasks, such as dimensionality reduction, clustering, and regression, commonly applied for image recognition, speech recognition and predicting modelling. Scale inhibitor (SI) squeeze treatment is one of the most common techniques to inhibit the deposition of oilfield scale. It consists in bull-heading a scale inhibitor slug in producing wells, the chemical retains on the rock surface and is slowly released when the well is back in production, the well will be protected from scaling, if the produced chemical concentration is above a certain threshold, usually few ppm. Generally, process modelling is applied to predict the SI return concentration vs water produced, to determine the treatment design lifetime, based on the operating parameters, formation and inhibitor properties, particularly retention capability. Commonly, a historically matched field isotherm results in predictions with significant degree of accuracy, where the main disadvantage is that historical data for the producing well under consideration is necessary. The purpose of this manuscript is to develop a ML model, where the input is the treatment design, which consists of chemical slug volume, chemical concentration, and overflush volume, well formation characteristics (perforation heights, permeability and porosity), and the output the scale inhibitor return concentration vs. volume of water produced. The ML model was trained, tested and blind validated using squeeze treatment data for a field in the North Sea. The dataset consists of 141 field squeeze treatments, 9 for blind validation, and 131 for 80/20 training and testing split. The best ML model Random Decision Forrest (RDF), which resulted in significant accurate prediction over the 9 blind squeeze treatments. This model may be used as proxy for the process modelling of squeeze treatments.