New machine learning technique could revolutionize research on 500-million-year-old microfossils

Have you ever heard of palynomorphs, the “microfossils” that are found in large numbers almost everywhere? These microscopic fossils appear in sedimentary rocks all over the world and are invaluable to geologists and paleontologists studying the evolutionary history of the planet. But their small size and sheer numbers can make them a challenge to work with, so researchers have created a new machine-learning technique that makes this otherwise difficult task easier.

Palynomorphs are really small; their size can range from 5 to 500 micrometers. If we consider the diameter of a human hair, which ranges from 17 to 181 micrometers, then we can imagine how small they can be. Even pollen grains tend to be larger than the smallest Palynomorphs.

These tiny fragments are composed of compounds that are remarkably resistant to most forms of decomposition, often consisting of sporopollenin, dinosporin, or similar compounds. They formed anywhere from a few million years ago to more than 500 million years ago. As such, they are valuable to researchers who want to age a rock layer or reconstruct a long-lost environment—for example, whether a layer formed underwater or was a terrestrial feature.

Analyzing these changes allows us to learn a lot about how the Earth has changed, and can also shed light on past climatic conditions and geological events.

Previously, scientists spent tedious hours manually classifying these microfossils, staring into microscopes where they could see billions of samples on multiple slides. It’s a tedious and frustrating process, but new advances in AI-assisted techniques could make the process much easier.

A researcher led by a team from the University of Tromsø in Norway has introduced a two-step AI-based system that detects and classifies microfossils based on microscopic images.

“We propose an automated pipeline for the extraction and classification of microfossils from raw microscopic images. The method is fast and efficient and does not require intensive computational power,” the team wrote.

“We show that our approach improves the state of the art in fossil extraction. Identifying individual species using machine learning is new and promising.”

The team achieved this in stages. First, they used a pre-trained object detection model—YOLOv5—to examine, identify, and extract individual palynomorphs from slide images. This process creates bounding boxes that appear around each microfossil, saving dozens of hours of work.

Then, in the second stage, the team used self-supervised learning (SSL), which is a relatively new and increasingly popular learning paradigm. This technique can essentially be trained to extract specific features from processed samples. It relies on self-supervised models to generate hidden labels from unstructured data.

In this study, the team compared two SSL platforms – SimCLR and DINO – both of which proved to be invaluable in accelerating the classification process.

“This work shows that the use of AI in this field has huge potential,” Iver Martinsen, the study’s first and co-author, said in a statement. “By using AI to automatically detect and recognize fossils, geologists can have a tool that helps them better use the vast amount of information provided by drill samples.”

The team used AI to detect Palynomorphs using data obtained by the Norwegian Offshore Directorate from the Norwegian continental shelf. To test its accuracy, the team tested the model by classifying several hundred previously labeled fossils from the same drill hole.

“We are very pleased with our results. Our model outperforms previous available standards. We hope that this work will be beneficial to geoscientists in both industry and academia,” Martinsen adds.

The article was published in the journal Artificial Intelligence in Geosciences.