| Popis: |
Summary In the last decade, Deep Learning applications to unstructured data, such as images and texts, has known significant technical progress and democratization. However successfully adapting these technologies to geological data and activities is far from straightforward. As a contribution to the digital transformation of the subsurface industries, in this study we present three promising Deep Learning applications to unstructured geological data. The first use case is an automated classification of macroscopic rock samples pictures with convolutional neural networks. The second use case is an accelerated delineation of foraminifera micro-fossils on thin sections scans using segmentation algorithms. The third use case is an assisted mining of scientific texts to characterize hydrocarbon source rock formations, based on an entity extraction engine. From these use cases, we highlight the main challenges to expect in similar projects and share some good practices. Notably, we describe innovative methods to embed prior geological knowledge in the algorithms, to handle situations where only little training data is available, and to distribute the corresponding codes to geologists in user-friendly ways. |
