Doing More with Less: Application of Machine Learning Regularization for Well Log Predictions and Model Building in Sparse Data Environment

Submitted Abstract Objectives/Scope Here in our case study, we have embarked in finding the best methodology and approach in tackling sparse data environment and limited logging via a multi-stage Machine Learning workflow that aims to address these issues, in which the aim is to aid in building the best possible background models to be used in Seismic Inversion. Methods, Procedures, Process For the well logs, two testing conditions were introduced whereby for one case, we would only be using some well data bases for the training while in another case, we would be slowly increasing the training dataset by including the predicted logs into the well database, which is generated via the Machine Learning log to log prediction process. This additional method was added to the mix as to bolster the Machine Learning training dataset and to increase its training pool, at the cost of possibly increasing the bias and heuristics of the ML as this iteration continues with each passing Machine Learning log prediction stage. For the seismic data stage, the data is first conditioned with seismic attributes (Dip, Instantaneous Frequency, Gradient Magnitude, Quadrature, and Envelop) and these were extracted across the well locations at the test wells. Proper seismic well tie and seismic extraction is also done to minimize any errors being carried forward into the ML training and testing phase. Results, Observations, Conclusions This case study has demonstrated the usefulness of applying ML methods via a detailed workflow in building a background model for seismic inversion. Although most of the results are not without its flaws, it is indeed observed that the result tends to get better with the inclusion of an extended flow coupled with a robust and sequenced training database. The way forward from these tests would be to integrate these with Deep Learning methods which would hopefully provide this study with a much more consistent background model for Vp and Vs, as outlined by Mosser (2018) Novel/Additive Information This workflow introduces a two-step workflow to initially populate the sparse data by conducting log predictions using otherprior to predicting the seismic low velocity model, which is then used during seismic inversion to help build a result for it.   Full Paper Summary Well logs are the main drivers and input for Quantitative Interpretation (QI) by serving as one of the hard data of the subsurface available for geoscientists and engineers. It is well known that the distribution of well data along the field is sparse if not available at all for exploration area. In addition, some well log data are not available at certain logging intervals due to cost and technical challenge. Since well log data is critical to build the background models, missing log data interval could lead to inconsistent and even erroneous result in seismic inversion, hence impacting the overall result for the seismic inversion. Machine Learning (ML) application for rock property prediction has been found in the field of geosciences as showcased by Das et al. (2019) and Purves et al. (2021) for instance. However, ML or deep-learning based predictive method require the existence of large amount of input training data in order the predictive model network can generalize well when applied to blind data. Given that the sparsity is found both spatially and laterally, how would these ML methods would work in such conditions? In this paper, we have developed a workflow in handling these two fronts via a multi-stage process that aims to address the data sparsity from patchy logging as well as laterally sparsely located well logs to build the best possible background models to be used in Seismic Inversion.