Abstract Defect detection and localization are key to preventing environmentally damaging wellbore leakages in both geothermal and oil/gas applications. In this study, a multistep, machine learning approach is used to localize two types of thermal defects within a wellbore model. This approach includes a comsol heat transfer simulation to generate base data, a neural network to classify defect orientations, and a localization algorithm to synthesize sensor estimations into a predicted location. A small-scale physical wellbore test bed was created to verify the approach using experimental data. The classification and localization results were quantified using these experimental data. The classification predicted all experimental defect orientations correctly. The localization algorithm predicted the defect location with an average root-mean-square error of 1.49 in. The core contributions of this study are as follows: (1) the overall localization architecture, (2) the use of centroid-guided mean-shift clustering for localization, and (3) the experimental validation and quantification of performance.