Address
81-83 Campbell Street, Surry Hills NSW 2010
Work Hours
Monday to Friday: 7AM - 7PM AEST
The availability of high-accuracy and detailed geological data is rapidly increasing, however our ability to effectively use this data in predictive modelling remains limited due to challenges such as the large variety of facies present in a rock formation. Current methods for using geologic information can be divided into two categories: supervised and unsupervised learning. Supervised learning involves estimating one variable from multiple variables, while unsupervised learning involves learning a variable without being given any other variables. In contrast, generative adversarial networks (GANs) are an unsupervised learning technique that involves training two neural networks with different architectures against different datasets so that the resulting network learns to produce samples from the dataset by mimicking how an unbiased network would generate the same samples. This paper will explore how geologic facies can be classified using generative adversarial networks with a focus on exploring 3D sedimentary environments.
Supervised Learning
Supervised learning algorithms are typically used to distinguish between lithologies and facies. The key difference between the two techniques is that in supervised learning algorithms, one variable is estimated from correlated variables. For example, a supervised algorithm can be used to estimate the depth of an object from its width and height.
Unsupervised learning
GANs have been recently used in the field of computer vision to generate new images from a set of training data. Supervised learning has mostly been applied to large datasets with high dimensionality, while unsupervised learning has been primarily used for classification and regression tasks such as image segmentation and model fitting. In this paper, generative adversarial networks will be used to classify sedimentary facies within 3D geologic environments. As a baseline, GANs will be trained using synthetic data generated from a single input dimension. The resulting network will then be evaluated on real world data.
Generative adversarial network
Classification of geologic facies 5D and 3D geological interpretations are increasingly important in understanding rock formation and prediction. Geologic facies present a challenge in organizing data to achieve the highest level of accuracy for predictive modelling. Current methods for using geologic information can be divided into two categories: supervised and unsupervised learning. Supervised learning involves estimating one variable from multiple variables, while unsupervised learning involves learning a variable without being given any other variables. In contrast, generative adversarial networks (GANs) are an unsupervised learning technique that involves training two neural networks with different architectures against different datasets so that the resulting network learns to produce samples from the dataset by mimicking how an unbiased network would generate the same samples. This paper will explore how geologic facies can be classified using generative adversarial networks with a focus on exploring 3D sedimentary environments.
The Role of Neural Networks in Classification
A neural network is a computer program that is trained to classify data by convolving it with a non-linear activation function and providing a bias parameter. In contrast, an artificial intelligence algorithm that uses generative adversarial networks (GANs) requires two separate neural networks with different architectures to train the model on two datasets in order to generate samples from both datasets. The first neural network trains the model on one dataset, while generating samples of the second dataset. The resulting network learns to produce samples from this second dataset as though it were producing real samples.
The availability of high-accuracy and detailed geological data increases our ability to use geologic information effectively and efficiently, however current methods for using geologic information can be limited due to challenges such as the large variety of facies present in a rock formation. This paper will explore how generative adversarial networks can help identify sedimentary environments through classification of 3D facies.
GAN Architecture for Classification
The generative adversarial training begins with an input layer that is composed of the X-ray images for all of the samples. This generates an artificial data representation that can be used to train a classification model. The convolutional neural network (CNN) is then trained with this generated image representation as its input and outputs a probability distribution of the locations in 3D space where rocks and sand are present. A GAN architecture is then trained so that it learns to produce samples from the dataset by mimicking how an unbiased network would generate the same samples. The resulting neural network creates a new image, which is then fed into the CNN for further processing.
Training the GAN Model
GANs are trained by having one neural network produce samples that are input into the other neural network to be classified. The first network is referred to as the discriminator and the second neural network is called the generator. In this work, we used images from a dataset of sedimentary facies provided by the USGS that were converted into 2D grayscale images using an RGB colour space. After classification, samples were sorted based on their facies class label and a 3D image was created.
