[S-GL22] Geochronology and Isotope Geology

[SGL22-02] AI is just like a human: A multi-analyst and AI comparison study on Fission Track analyses

*Murat Taner Tamer¹, Samuel Boone^2,3, Ling Chung³ (1.China Earthquake Administration, 2.School of Geosciences, Faculty of Science, The University of Sydney, 3.School of Geography, Earth and Atmospheric Sciences, University of Melbourne)

Keywords:Artificial Intelligence, Geochronology

In this report, we present the findings of the inaugural image-based comparative study of human and artificial intelligence (AI) fission track analysis. The study encompasses two standards, three real-world samples, and one artificially mixed sample. The analysis was conducted by both a trained AI algorithm and a seasoned human analyst, who performed fission-track analysis on identical image datasets of grains of varying quality and track densities. The findings demonstrate that the AI model offers counts and density estimations for grains devoid of defects or inclusions, also referred to as "non-tracks," exhibiting within 80% the confidence lines of precision of an experienced human analyst. However, this precision undergoes a decline when analyzing grains containing non-tracks. A comparison of the AI model's outcomes with those attained by human analysts in a prior image-based inter-analyst study reveals the level of similarity. The present limitations of AI algorithms in fission-track studies pertain to their current lack of inclusion and defect pattern recognition capabilities, thereby hindering the automatic identification of non-tracks. However, with the ongoing development of defect and inclusion pattern recognition algorithms capable of automatically detecting non-track features and tracking-tip recognition algorithms that can automatically measure track length, there is a significant potential for the AI algorithm to function without variable bias and lead to a transformation in manual fission-track analysis, akin to "total annealing."