Back to search

PETROMAKS2-Stort program petroleum

AI Augmented Analysis in digital biostratigraphy - palynology

Alternative title: KI-forsterket analyse i digital biostratigrafi - palynolgi

Awarded: NOK 5.5 mill.

Biostratigraphy is the dating of rocks with the help of fossils from plants and animals. The fossil composition of any given biostratigraphic sample depends on the age, the geographic position and the environmental conditions at the time the sample was formed. In investigations that target the subsurface geology, (e.g. petroleum exploration) samples are dated by examining microfossils such as palynomorphs - pollen, spores, microalgae, and marine plankton - through a microscope, and manually identifying and counting these in each sample. This is time-consuming work, as a single sample can contain several thousands of palynomorphs, where experience is required in order to identify the fossils correctly. An experienced palynologist can therefore only analyze a handful of samples per day. Inspired by recent advances in digital scanning developed for pathology where tissue samples are scanned into a digital image and analyzed by Artificial Intelligence (AI), this project has utilized the digital scanning technique and developed an AI for detection and identification of fossil palynomorphs. The project had three phases. In-house standard palynological microscope slides at APT were selected for study; representing a variety of sedimentary samples ranging in age from the Early Jurassic to the Neogene, as well as a variety of depositional environments, depth below sea-level, and degree of fossil preservation. In the first phase the team at APT performed and compared manual microscopy analysis of these palynological slides with on-screen analysis of high-resolution digital images of the same slides, in order to evaluate if the results from the scanned slides were comparable to the results from the microscope. The results show that on-screen analysis of a sample on average takes 40–50 minutes longer than a microscope analysis which normally takes between 1–3 hours depending on the sample content and preservation. Overall, the analytical results did not differ significantly. The samples were assigned the same age regardless of the analytical method, and they were mostly classified as belonging to the same depositional environment. In those cases where samples were classified as differently in the on-screen analysis, the microscope analysis had placed the samples close to the statistical border between two depositional environments. Slides produced by conventional palynological preparation technique for routine analysis in the microscope often contained unevenly distributed and/or overlapping palynological material resulting in identification issues in the digital slides; issues which would be an even bigger problem for an AI than for a biostratigrapher. Based on the results from the first phase, in the second phase the team at APT tested several adjustments to the preparation method and developed a protocol for optimized palynological preparation procedures to ensure high quality of the digital slides. The palynological slides from phase two were also scanned digitally and analyzed the same way and compared to the results obtained in the first phase. The results show that digitally scanned high-resolution palynological slides can be used for image recognition, and that adjustments in the preparation method used to produce the palynological microscope slides are needed to ensure optimal quality for scanning and for image recognition and machine learning (ML). In the third phase, we used image recognition and ML algorithms to develop an artificial intelligence (AI) software for analysis of digital palynological slides. The AI-software, developed by Simula Research laboratory, consists of several modules; data processing and ML, detection and identification, as well as clustering after selected parameters. For training purposes of the AI, the team at APT built an extensive database of labelled digital examples of different palynomorphs. With the developed user interface (UI) it is possible for the user to upload digital palynological slides and if there are annotations linked to the slide, inspect these. The AI can detect and identified palynomorphs that it has been trained to recognize, and the user can choose to accept or reject the AI’s identification, and further has the option of providing a new identification. The UI also provides quantitative information and degree of certainty regarding its identifications. In addition, it is possible to do clustering of palynomorphs after selected parameters. The developed AI software for palynology, together with the improvements to the preparation technique that ensures higher quality digitally scanned images of palynological slides, have the potential to change how biostratigraphy is used in geological investigations such as that of the oil industry. The new technique is expected to have an impact on planning, drilling and evaluation of wells and formations and will be an important aspect of risk management and a valuable contribution to sustainable development goals.

The results of the project have shown that digitally scanned high-resolution palynological slides can be used for image recognition, and that adjustments in the preparation method used to produce the palynological microscope slides are needed to ensure optimal quality for scanning and for image recognition and machine learning. The developed artificial intelligence (AI) software is capable of detecting, recognizing and identifying palynomorphs, as well as clustering palynomorphs after selected parameters. The developed AI software has the potential to change how commercial and academic palynologist plan, execute and interpret their data. The digitalization of samples and analysis has an impact on how palynologist share data, cooperate and exchange knowledge and experience. The AI software has the possibility of being deployed as a cloud based software, potentially allowing multiple users to contribute to its database. Larger and more statistically accurate data sets further provide new possibilities in areas such as paleoenvironmental and paleoclimate research. The developed AI software also has the potential of being used in biostratigraphic training of future palynologists. Together, the improvements to the palynological preparation technique that ensures higher quality digitally scanned images of palynological slides, and the developed AI software which can detect, identify and quantify palynomorphs in the scanned slides, have the potential to change how biostratigraphy is used by oil companies. The new technique is expected to have an impact on planning, drilling and evaluation of wells and formations and will be an important aspect of risk management and a valuable contribution to sustainable development goals.

Biostratigraphy, using fossils to date rocks, is obligatory for all exploration wells drilled in the offshore Norwegian Sector. It is used to help understand the sub-surface geology and to correlate sections in wells both within fields and on a larger regional scale. Although biostratigraphic data was part of the digital revolution in the industry of the 1980’s, the actual study of the microfossils with a microscope and prepared glass slides has changed very little in the past decades. By using image digitalisation and Artificial Intelligence a technique can be developed to scan microscope slides into a digital high-resolution image and an AI software specifically developed for this usage will find, identify and quantify the fossil content. With this new analysis method, we believe that we can not only decrease the time used for analyses but also obtaining data at a more statistically accurate level and mitigate human inconsistencies and biases. With 3-dimentional microfossils that need to be identified in a 2-dimentional view, based on several morphological parameters, the image recognition software needs to identify fossils from different angles, poorly preserved and fragmented fossils, as well as partially hidden or folded fossils. A species can also have a morphological range and evolution. This morphological and evolutionary complexity is the base for the research needed to develop the AI software. Within this project we intend to establish if the digital scans of the biostratigraphical samples have the resolution needed to be used for quantitative analyses, we will also develop a sample preparation process to reliably produce the best digital samples possible. We will then design and develop an Artificial intelligence software for clustering and classification of microfossils based on deep-learning based algorithms for detection and segmentation of microfossils and algorithms based on autoencoders for extracting features predictive of different fossil types.

Funding scheme:

PETROMAKS2-Stort program petroleum