Summary In a recent paper, we published a machine learning method to quantitatively predict reservoir fluid gas/oil ratio (GOR) from advanced mud gas (AMG) data. The significant increase of the model accuracy compared to traditional modeling approaches makes it possible to estimate reservoir fluid GOR based on AMG data while drilling, before the wireline operation. This approach has clear advantages because of early access, low cost, and a continuous reservoir fluid GOR for all reservoir zones. This paper releases further study results to predict other reservoir fluid properties in addition to GOR, which is essential for geo-operations, field development plans, and production optimization. Two approaches were selected to predict other reservoir fluid properties. As illustrated by the reservoir fluid density example, we developed machine learning models for individual reservoir fluid properties for the first approach, similar to the GOR prediction approach in the previous paper. As for the second approach, instead of developing many machine learning models for individual reservoir fluid property, we investigated the essential properties for equation of state (EOS) fluid characterization: C6 and C7+ composition and the molecular weight and density of the C7+ fraction. Once these properties are in place, the entire spectrum of reservoir fluid properties can be calculated with the EOS model. The results of reservoir fluid property prediction are satisfactory with both approaches. The reservoir oil density prediction has a mean average error (MAE) of 0.039 g/cm3. The accuracy is similar to the typical density derived from the pressure gradient from wireline logging data. For the essential fluid properties required for EOS model prediction, the overall accuracy is less than the laboratory measurements but acceptable as the early phase estimations. The reservoir fluid properties predicted from the EOS model are similar to the predictions from individual machine learning models. We applied the field measured AMG data into the reservoir fluid property models and achieved good results, as illustrated by the reservoir fluid density example. The previous paper completed the methodology to predict all reservoir fluid properties based on AMG data. This work paves the way to generate a complete reservoir fluid log for all relevant reservoir fluid properties while drilling. The method has a significant business impact, providing full coverage of reservoir fluid properties along the well path in the early drilling phase. The advantage of providing reservoir fluid properties in all reservoir zones while drilling far outweighs the limitation of somewhat reduced reservoir fluid property accuracy.

This paper presents a new approach to GOR prediction from advanced mud gas data for gas flooded reservoirs. This new approach predicts the GOR of the reservoir oil and/or gas while drilling in gas flooded reservoirs and makes real-time well decisions based on identifying the reservoir fluids. The method is currently under extensive verification in real field operations. The new method's potential is significant for accurately mapping resources for in-fill wells, boosting profitability, and lowering carbon footprint in gas flooded fields. This study widens our previous machine learning model from advanced mud gas (Yang et al., 2019) developed for predicting in-situ reservoir fluid properties by extending the methodology to include gas flooded reservoirs. The methodology is verified by compositional modeling of a large field undergoing pressure depletion, water, and gas injection over 30 years. The compositional reservoir model is used to generate advanced mud data (light components C1 to C5) to predict fluid phase properties by machine learning. The predicted fluid properties are then compared to the actual model data. The data points represent different well locations and different times of the field development, i.e., initial state, water injection, gas injection, and pressure depletion. To accurately predict the fluid properties using a machine learning model, the "training database" must contain a wide range of fluids that covers the "compositional window" in the gas injection process. To achieve this, we have introduced synthetic fluid samples into the machine learning fluid database. The synthetic data are generated through slim tube simulation by developing miscibility between original reservoir fluid composition and representative injection gases. We have verified the machine learning model by comparing the GOR from the reservoir simulation model against the predicted GOR at different depths during the different production stages of the field. The GOR prediction shows good agreement with the simulation results.

Standard mud gas data is part of the basic mudlogging service and is mainly used for safety. Although the data is available for all wells, it is not used for real-time fluid typing due to poor prediction accuracy. We developed a new method recently based on a large in-house reservoir fluid database and significantly improved the fluid typing accuracy from standard mud gas data. The new technology unlocks the large potential of utilizing standard mud gas data for thousands of wells. The standard mud gas data has limited gas components that can be detected confidently (usually from C1 to C3). In addition, they are raw data without recycling correction and extraction efficiency correction. Following the traditional geochemical analysis methods, some key parameters (C1/C2, C1/C3, Bernard ratio) have a universal threshold to distinguish gas and oil. The main reasons for the poor fluid typing accuracy are due to 1) lacking C1 to C3 composition correction for wells with oil-based mud; 2) geochemical parameters based on C1-C3 are field dependent. Based on the reservoir fluid database analysis, we divide the reservoir fluids from different fields into two categories. For Type I fields, there are large differences between C1 to C3 component ratios for oil and gas. When water-based mud is used, C1 to C3 component ratios from standard mud gas can be utilized directly to identify oil and gas. When oil-based mud is used, we developed a new method to achieve corrected standard mud gas composition for fluid typing using pseudo extraction efficiency correction based on Equation of State. For Type II fields, there are severe overlapping of C1 to C3 component ratios for oil and gas. The overlapping is the main reason for the poor fluid typing accuracy. We recommend utilizing a heated degasser when drilling into Type II fields to provide additional data of C4 and C5 for accurate fluid typing. Johan Castberg is a Type I field, and we achieved excellent fluid typing results for 14 wells using water-based mud and oil-based mud. There is no additional data acquisition cost for standard mud gas data, which is available for all wells. The new method makes accurate fluid typing possible for real-time well decisions like well placement, completion, and sidetrack. The innovation created significant business opportunities based on the standard mud gas, which has been regarded as not applicable data for accurate fluid typing for many decades.

Advanced mud gas logging has been used in the oil industry for about 25 years. However, it has been challenging to predict reservoir fluid properties quantitatively (e.g., gas oil ratio – GOR) from only the advanced mud gas data (AMG) while drilling. Yang et al. proposed the first accurate GOR predictive model in 2019 from advanced surface data based on a machine learning algorithm. Since then, the method has been applied to both conventional and unconventional fields with good results. For our Norwegian operational units, we are developing a real-time service for fluid identification to optimize fluid sampling in exploration wells and support production drilling. Here, quantitative information about reservoir fluids will support the teams to take wellinformed decisions with respect to well placement, petrophysical log interpretation, and optimizing production by improving the selection of perforation intervals. We utilize a standard wellbore software platform to integrate the following data for fluid identification: AMG data, various AMG QC tools, normalized total gas response, GOR prediction, and petrophysical logs from logging while drilling (LWD). The proposed work approach integrates the information from multiple disciplines and makes the real-time fluid identification task much more reliable for operational decisions. We selected two field cases to demonstrate the approach of integrating AMG data and petrophysical logs. The first field case is an exploration well with multiple reservoir zones planned as production targets. The integrated approach shows reservoir fluids from all reservoir zones are almost identical. Consequently, we reduced the sampling program and only sampled at the best reservoir zone for cost efficiency. The second field case is a mature field being produced by pressure support from water, gas or water alternating gas injections. When a new production well is drilled, there is always a question of whether it encounters any injection gas. We applied the new approach to several production wells and obtained satisfying result. The latest information from the predictive GOR model solved many puzzles in petrophysical interpretations. This paper presents a new approach for reservoir fluid identification by integrating advanced mud gas data and petrophysical logs while drilling. This new approach makes real-time operational adjustments possible based on reservoir fluid identification along the well. The business potential is significant for accurately mapping resources for in-fill wells, boosting profitability, and lowering carbon footprint.

In a recent paper, we published a machine learning method to quantitatively predict reservoir fluid gas oil ratio (GOR) from advanced mud gas (AMG) data. The significant increase of the model accuracy compared to traditional modeling approaches makes it possible to estimate reservoir fluid GOR based on AMG data while drilling, before the wireline operation. This approach has clear advantages due to early access, low cost, and a continuous reservoir fluid GOR for all reservoir zones. In this paper, we release further study results to predict other fluid properties besides GOR, which is essential for geo-operations, field development plans, and production optimization. We use two approaches to predict other fluid properties. For the first approach, we develop machine learning models for reservoir fluid density, similar to the GOR prediction approach in the previous paper. Based on an extensive fluid database, we establish machine learning models to predict reservoir fluid density from C1 to C5 compositions (same format as AMG data). As the second approach, instead of developing a machine learning model for individual fluid property from C1 to C5 compositions, we investigate the most important properties for EOS fluid characterization: C6 and C7+ compositions, and molecular weight and density of C7+ fraction. Once these properties are in place, the entire spectrum of fluid properties can be calculated with the equation of state (EOS) model. The results of fluid property prediction are satisfactory with both approaches. The reservoir oil density prediction has a mean average error (MAE) of 0.039 g/cm3. The accuracy is superior to typical density derived from the pressure gradient from wireline logging data. For the essential fluid properties required for EOS model prediction, the overall accuracy is acceptable compared with laboratory measurements. The reservoir fluid properties predicted from the EOS model are similar or better than the predictions from individual machine learning models. We applied the field measured AMG data into the fluid property models and achieved reasonable results, as illustrated by the reservoir fluid density example. The paper completed the methodology to predict all fluid properties based on AMG data. This work paves the way to generate a complete reservoir fluid log for all relevant fluid properties. The method has a significant business impact due to the full coverage of reservoir fluid properties along the well path compared with the availability of discrete PVT samples. The advantage of providing fluid properties in all reservoir zones along the well path far outweighs the limitation of somewhat reduced fluid property accuracy.