Downhole fluid sampling is ubiquitous during exploration and appraisal because formation fluid properties have a strong impact on field development decisions. Efficient planning of sampling operations and interpretation of obtained data require a model-based approach. We present a framework for forward and inverse modeling of filtrate contamination cleanup during fluid sampling. The framework consists of a deep learning (DL) proxy forward model coupled with a Markov Chain Monte Carlo (MCMC) approach for the inverse model. The DL forward model is trained using precomputed numerical simulations of immiscible filtrate cleanup over a wide range of in situ conditions. The forward model consists of a multilayer neural network with both recurrent and linear layers, where inputs are defined by a combination of reservoir and fluid properties. A model training and selection process is presented, including network depth and layer size impact assessment. The inverse framework consists of an MCMC algorithm that stochastically explores the solution space using the likelihood of the observed data computed as the mismatch between the observations and the model predictions. The developed DL forward model achieved up to 50% increased accuracy compared with prior proxy models based on Gaussian process regression. Additionally, the new approach reduced the memory footprint by a factor of ten. The same model architecture and training process proved applicable to multiple sampling probe geometries without compromising performance. These attributes, combined with the speed of the model, enabled its use in real-time inversion applications. Furthermore, the DL forward model is amendable to incremental improvements if new training data becomes available. Flowline measurements acquired during cleanup and sampling hold valuable information about formation and fluid properties that may be uncovered through an inversion process. Using measurements of water cut and pressure, the MCMC inverse model achieved 93% less calls to the forward model compared to conventional gradient-based optimization along with comparable quality of history matches. Moreover, by obtaining estimates of the full posterior parameter distributions, the presented model enables more robust uncertainty quantification.

The high frequency and long duration of well production data have high value for inferring reservoir properties, monitoring reservoir conditions and detecting changes in the well flow conditions. Unlike traditional well test data where the well flow-rate and the pressure response are carefully monitored, long-term pressure data is subject to abrupt changes in flow-rate, noise caused by hardware failures, missing data and physical changes in the reservoir. In this work, methodology to interpret long-term flow-rate and pressure data was developed using wavelet multiresolution analysis in combination with deep learning algorithms. The methodology requires little to no data preprocessing, no explicit knowledge of a physical model and has high robustness to noise. The proposed methodology was found to exploit the ability of wavelets to capture and synthetize the relevant behavior of the reservoir performance history at multiple scales. This was achieved by applying the Maximal Overlap Wavelet Transform (MODWT) to long-term flow-rate data and using the output of the MODWT as the input for a Recurrent Neural Network (RNN) which creates a function map between flow-rate and pressure as a function of time. For the training of the RNN, synthetic well data were used. The application of wavelet transforms allowed the neural network to perform an automatic noise filtering and simplified the training process while requiring no knowledge of noise levels, little to no manual preprocessing of data and no knowledge of the reservoir's physical model. Moreover, the mathematical formulation of the wavelet transforms allows the neural network to take advantage of the multiresolution properties, opening the door to executing the whole analysis procedure in a simplified way. The proposed methodology aids in the inference of effects such as changes in skin factor, pressure depletion or changes in the reservoir model from the available production data, even when those data are noisy or incomplete. By using production data, the inference can be done without loss of production and used to support oil production enhancement operations. Moreover, the coupling of a neural network with multiresolution analysis eliminates the need for feature engineering and performs automatic denoising of the data, which were found to be extremely desirable properties for a methodology to be scalable and applicable to real well datasets.

Knowledge of reservoir heterogeneity and connectivity is fundamental for reservoir management. Methods such as interference tests or tracers have been developed to obtain that knowledge from dynamic data. However, detecting well connectivity using interference tests requires long periods of time with a stable reservoir pressure and constant flow-rate conditions. Conversely, the long duration and high frequency of well production data have high value for detecting connectivity if noise, abrupt changes in flow-rate and missing data are dealt with. In this work, a methodology to detect interference from longterm pressure and flow-rate data was developed using multiresolution analysis in combination with machine learning algorithms. The methodology presents high accuracy and robustness to noise while requiring little to no data preprocessing. The methodology builds on previous work using the Maximal Overlap Wavelet Transform (MODWT) to analyze long-term pressure data. The new approach uses the ability of the MODWT to capture, synthesize and discriminate the relevant reservoir response for each individual well at different time scales while still honoring the relevant flow-physics. By first applying the MODWT to the flow rate history, a machine learning algorithm was used to estimate the pressure response of each well as it would be in isolation. Interference can be detected by comparing the output of the machine learning model with the unprocessed pressure data. A set of machine learning, and deep learning algorithms were tested including Kernel Ridge Regression, Lasso Regression and Recurrent Neural Networks. The machine learning models were able to detect interference at different distances even with the presence of high noise and missing data. The results were validated by comparing the machine learning output with the theoretical pressure response of wells in isolation. Additionally, it was proved that applying the MODWT multiresolution analysis to pressure and flow-rate data creates a set of "virtual wells" that still follow the diffusion equation and allow for a simplified analysis. By using production data, the proposed methodology allows for the detection of interference effects without the need of a stabilized pressure field. This allows for a significant cost reduction and no operational overhead because the detection does not require well shut-ins and it can be done regardless of operation opportunities or project objectives. Additionally, the long-term nature of production data can detect connectivity even at long distances even in the presence of noise and incomplete data.