Extrapolating rock physicochemical properties from discrete sample data using computer vision

Quantifying the lithological, physical, and chemical properties of geological samples requires experts to generate records of visual material properties, followed by targeted sampling to build a comprehensive geological understanding of the material. This understanding, therefore, is grounded in a significant amount of time on the part of the annotator to reliably extrapolate physicochemical information gained from relatively few discrete samples. It is not feasible for an expert to visually describe the entire surface of a sample; leading to only a coarse level of information where large amounts of material are being scrutinized. To overcome these limitations, we leverage the descriptive power of deep learning to generate objective high spatial resolution classifications of a ∼100 m long drill core of in situ carbonate cemented ocean crust talus breccia. Our AI-generated lithological classifications are used to semiautonomously extrapolate bulk density and carbon concentration of discrete samples to the entire surface of the core at a 0.25 cm² resolution. Synthetic bulk density generated shows improvement to traditional GRA measurements, and our estimate of carbon content confirms the importance of oceanic talus breccia in global biogeochemical cycles.