Determining reservoir fluid phase envelope from downhole fluid analysis data using physics-informed machine learning techniques

A computer-implemented method of training an artificial neural network for processing data related to a downhole fluid, the method comprising:

• collecting, using a downhole fluid analysis module, data related to the downhole fluid, the data comprising at least a measured mole percent of each of methane (C1), ethane (C2), propane (C3)-n-pentane (C5), hexane plus fraction (C6+), carbon dioxide (CO2), hydrogen sulfide (H2S), and nitrogen (N2) of the downhole fluid;
• processing, by a computer, the collected data related to the downhole fluid to produce a dataset;
• partitioning the dataset of the collected data into a testing data portion and a training data portion;
• performing an output validation on the testing data portion;
• training an artificial neural network model to produce a training data output in real-time of:
• a fluid property estimation at each respective temperature among a plurality of temperatures, the fluid property estimation being based on at least the measured mole percent of each of methane (C1), ethane (C2), propane (C3)-n-pentane (C5), hexane plus fraction (C6+), carbon dioxide (CO2), hydrogen sulfide (H2S), and nitrogen (N2) of the downhole fluid and the plurality of temperatures as inputs to the artificial neural network to estimate saturation pressures at each respective temperature among the plurality of temperatures;
• and
• a phase envelope for the downhole fluid based on the estimated saturation pressures,
• wherein the phase envelope is produced in real-time at a wellsite during an exploration phase without direct pressure-volume-temperature testing,
• wherein a number of training cycles performed during the training of the artificial neural network is regulated to reduce overfitting by a cost function comprising a mean-squared error function,
• wherein the number of training cycles is less than 50,000,
• wherein the artificial neural network comprises:
• an input layer having 8 inputs respectively comprising the measured mole percent of each of C1, C2, C3-C5, C6+, CO2, H2S, N2, and a discretized temperature value,
• exactly two hidden layers, each hidden layer having exactly 5 neurons, and
• an output layer having 2 outputs respectively corresponding to a high saturation pressure value and a low saturation pressure value, and
• wherein the phase envelope is produced by combining the output high saturation pressure value and the output low saturation pressure value across a plurality of discretized temperature values;
• and
• performing an output validation on the training data output.