Chapter 8 - Comparative study of shallow and deep machine learning models for synthesizing in situ NMR T2 distributions

Abstract

Precise assessment of the in situ pore size distribution is critical for successful reservoir development. Nuclear magnetic resonance (NMR) T2 distribution logging enables the in situ estimations of reservoir pore size, fluid mobility, and fluid composition. However, the NMR logs are not acquired in all the wells due to operational or financial constraints. In this study, we compare the performance of shallow-learning and deep-learning models in synthesizing the NMR T2 distribution logs. Six shallow-learning models and four deep-learning models are trained and tested on data acquired from a 300-ft depth interval of a shale formation. Both raw “easy-to-acquire” conventional well logs and the inverted formation mineral and fluid composition logs are used for the synthesis of the NMR T2 distribution logs. Deep-learning models perform better than shallow-learning models. We implement a two-step training process to boost the performance of the deep-learning models. Deep-learning models have better performance when trained with inverted logs as compared that with raw logs.