Higher price and declining ‘easy’ reserves have forced many oil and gas producers to increasingly focus on producing from challenging unconventional reservoirs or by redeveloping existing or abandoned fields to target an improved recovery. This can especially be seen from the widespread proliferation of fracturing operations, EOR and IOR processes as well as the widespread field redevelopment projects across the Middle East. However, in many operations, especially in deep reservoirs cement layer failures and well integrity issues are commonly observed. In line with this trend, demand for increased understanding of cement behavior over the course of the lifetime of the well has also increased.
Cement layer properties and the rate at which they change, will depend on: cement composition, environmental conditions, production conditions, and many other factors. Because of this, it may be more optimal to create a diverse batch of controlled composition cement tested over set time intervals and implement machine learning algorithms to populate the missing data; both as a function of time and cement composition. Research done as part of this project has developed a dataset of cement properties from controlled laboratory testing and various applications of machine learning will be applied on cement properties in order to create artificial data. The dataset contains laboratory testing of over 1000 samples comprised of class H and C cement, cured at both 25 and 75 degrees Celsius. Each sample has a recorded density, UPV, and UCS, as well as dimensional measurements. The dataset is properly organized for each method of analysis, with 30% of the dataset removed to function as the validation group. The tests are run with two versions of the dataset, one identical for all and one with a unique set of removed samples. After all testing is complete; the resultant information is organized and analyzed appropriately.
Machine learning methods show promise in accurately forecasting both the UCS and UPV for cement given enough information. It appears a combination of methods may be the best, as some methods are better at populating missing data across the same time interval, and other methods are better forecasting forward in time. Investigations into the correct combination of these methods, as well as calculating the statistical confidence of each forecasted dataset will likely result in the best method of for generating accurate data regarding any particular cement slurry.
While there is notable work regarding machine learning ant cement properties, a significant portion of that work is in surface construction in the fields of civil and industrial engineering. This work also works to highlight any effect curing conditions may have on results.