AI re-examines dinosaur footprints

For decades, paleontologists have pondered over mysterious three-toed dinosaur footprints. Were they left by fierce carnivores, gentle plant-eaters, or even early birds? Now, an international team has used artificial intelligence to tackle the problem—creating a free app that readily lets anyone decipher the past.

Dinosaur footprints are iconic trace fossils, but they’re notoriously hard to interpret. Traditional machine learning methods require huge datasets and manual labeling, which can introduce bias—especially since the true maker of a footprint is rarely known. To overcome this, a team lead by Gregor Hartmann of Helmholtz-Zentrum Berlin and Stephen Brusatte of the University of Edinburgh applied an unsupervised neural network called a “disentangled variational autoencoder.”

Beyond Dinosaurs

Hartmann applies similar AI techniques to analyze electron orbits of the synchrotron BESSY II, characterize X‑ray pulses at DESY's FLASH, study brain scans for early signs of dementia, identify nuclide contribution of gamma spectra and evaluate chemical reactions in battery and catalyst materials. “It’s exciting to see how these tools can advance both cutting-edge physics and our understanding of ancient life,” he says.

The team trained the model on nearly 2,000 fossil footprints—plus millions of augmented variations to mimic realistic changes such as compression and edge displacement. After testing almost 1,000 neural architectures, they found a compact, robust network that on its own identified eight key features of footprint variation: amount of ground contact; digit spread; digit attachment; heel load; digit and heel emphasis; loading position; heel position; and left-right load. When compared with expert classifications, the algorithm achieved 80–93% agreement, even for controversial specimens.

Science for Everyone: The DinoTracker App

To make their research accessible, the team developed DinoTracker—a free app which allows scientists and enthusiasts to upload or sketch a footprint and receive instant analysis. “Our method provides an unbiased way to recognize variation in footprints and test hypotheses about their makers,” says Hartmann. “It’s a tool for research, education, and even fieldwork.”

Expert Praise

“This study is an exciting contribution for paleontology,” says Brusatte. “This is an objective, data-driven way to classify dinosaur footprints—something that has stumped experts for over a century. It opens up exciting new possibilities for understanding how these incredible animals lived and moved.”

From Curiosity to Collaboration

The idea began at home: Hartmann’s young son, Julius, became fascinated with dinosaurs, inspiring Hartmann to explore paleontology literature. After reading the Rise and Fall of the Dinosaurs, he reached out to Brusatte and two of his students in Edinburgh, footprint experts Paige dePolo and Tone Blakesley, who had studied numerous Jurassic-aged dinosaur footprints on the Isle of Skye, Scotland, during their graduate studies. That spark grew into a collaboration and ultimately a PNAS publication.

Caption: A Jurassic-aged dinosaur footprint from the Isle of Skye, Scotland, rendered in 5 mm contours from a photogrammetric model. The schematic above and below the footprint represents the machine learning neural network of Hartmann et al., which untangles the key aspects of variation in a large sample of dinosaur tracks and uses that information to help researchers identify plausible trackmakers.

YouTube: The scientists and developer of the DinoTracker app explain how the app is structured and how it works.