Use this URL to cite or link to this record in EThOS:
Title: Automated lithofacies predictions from well logs
Author: Martin, Richard
Awarding Body: University of Edinburgh
Current Institution: University of Edinburgh
Date of Award: 2004
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
This PhD study uses two data sets to investigate lithofacies predictions via neural analysis. A neural network model is developed to predict lithofacies in the Palaeocene submarine fan deposits of the Lomond Field in the Central North Sea. When the results from the optimum model are compared directly to core in many cases a close match is found, even in intervals where the model interpretation does not match the ‘human’ interpretation. In wells where predictions do not match the core this is shown to be a consequence not of the methodology itself but because of characteristics of the data collected in the training wells. This is due to variations in tool type, interval depth and fluid saturation. When this occurs, prior calibration of logs in the training data, for example through making a correction for fluid saturation, can improve performance. In all cases the observation of the activation levels of all output nodes in the network can qualitatively describe any uncertainty in the results, thus aiding interpretation. Using a second data set of wells form the deep marine West Delta Deep Concession area, Nile delta, Egypt more neural network models are developed to predict ‘image’ facies. Model inputs consist of conventional logs and derived logs from statistical and power spectral analysis of the pad data from the Fullbore Formation Microlager (FMI) tool. Although results are similar, slight improvement is seen if separate networks are trained to predict image facies that belong to specific lithological groups, rather than a single network trained to predict all facies. Where sandy thin beds (<1cm in thickness) occur within thicker shale units further neural networks are needed to discriminate these, trained with different inputs from previous networks.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available