| Being involved in seismic data identification,faults and porosity are crucial for modelling prospect,characterizing reservoirs and placing the wells.Faults show lateral discontinuities of reflections in seismic images and are interpreted using the so-called seismic attributes.Porosity can be obtained from well logs and core samples.However,the relationship between porosity and seismic attributes,core samples and well logs is often challenging,time-consuming and the efficiency of their prediction may be negatively affected by a lot of factors.Therefore,we investigated the approach of artificial intelligence to solve the issues above using two types of models namely the multiple regression neural network(MRNN)model and the convolutional neural network(CNN)model.Faults detection was performed based on pixels of seismic images to classify which of them belong to fault class or non-fault class like in binary image segmentation problem.In turn,porosity prediction was solved as a supervised multiple regression problem.Herein,lithology,acoustic impedance,longitude and latitude were defined as the independent variables.Since the well-trained neural networks must result in small losses,errors,and keep the high accuracy,we trained several networks with different optimization functions to choose the best one.As results,the network trained with Adam optimizer showed the best performance in both cases and learned to bank on relevant features which are important for the:(1)recognition of faults with 97.2% of accuracy and 6% of loss;(2)prediction of porosity with 0.04% of MSE in loss,1.4% of MAE and 14% of RMSE in testing error.All the results were compared with those of Random Forest and Decision Tree models which are commonly used in the field.CNN results displayed the cleanest and accurate fault detection in milliseconds and MRNN predicted porosity with up to 99% of Pearson correlation to the measured one.In order to reduce the computational cost while preserving the prediction efficiency,the structure of the implemented CNN Unet-based contains 4 layers instead of 23 in the original one.Each layer was suggested to a number of operations: after one convolution we set the dropout and back to the convolution and operate a pooling on the generated features.In the 4th layer the pooling operation is missed since we had to up sample all the features to back to the original size of the input sample step by step.This allowed working with few data sets in contrast of what is common in deep learning and increasing the speed of prediction.To run the code,a 2 cores computer where each core is 3.30 GHz with 8 GB of RAM is only required.The program language in which this code is written is Python version 3.7 and only the following few packages are needed: tensorflow,os,sys,random,numpy,opencv,matplotlib,tqdm,skimage,sklearn and keras. |
