In my latest publication I explored the likelihood of failure in the cement sheath based on cement evaluation logs. These logs can point out heterogeneous features in the cement, such as microannulus and channels, and I can replicate these features in mechanical models. One issue that I came across is that, in many wells, these tools are not ran. To overcome this limitation I have been generating synthetic logs based on data from wells in the same field; the study was presented at the 55th US Rock Mechanics / Geomechanics Symposium. The utilization of  synthetic logs (or pseudo logs) is quite common in petrophysics and reservoir characterization to address lack of data. But unlike these logs, in cement bond logs (CBL) there is no correlation between depth or reservoir parameters and the cement bond. For this reason we chose not to use classical machine learning, where we look for the “best guess”, and employed a Bayesian approach. We used Gaussian Process Regression where, from the training data, the algorithm compute predictive distributions for new test inputs. With this framework I cannot locate the exact location of potential failure, but can quantify uncertainty in the model and exploit this uncertainty in order to make predictions on new wells. Certainly more precise than a simply random placement of cement properties.