A Deep Learning Enabled Approach for Igneous Textural Timescales

Abstract

<div> <div> <p><span data-contrast="auto">Textural information, such as crystal size distributions (CSD’s) or crystal aspect ratios are powerful tools in igneous petrography for interrogating the thermal history of rocks and the timescales of processes affecting them [1-3]. Plagioclase feldspar especially has found extensive use as a reliable tracer for igneous thermal history and processes with both the apparent 2D [4] and 3D [5] morphologies shown to vary predictably with crystallization time. However, most textural studies, especially relating to 3D morphologies, require extensive data collection which can be cumbersome and time consuming when performed manually. The aim of this work is to present a holistic and automated workflow to enable the rapid extraction of igneous timescales from plagioclase textures through an automated approach. These developments are vital to better allow petrologists to make timescale estimates that can be used in conjunction with diffusion chronometry and more fully characterise the temperature-time paths of igneous rocks.</span><span data-ccp-props="{"201341983":0,"335551550":6,"335551620":6,"335559739":0,"335559740":240}"> </span></p> </div> <div> <p><span data-ccp-props="{"201341983":0,"335551550":6,"335551620":6,"335559739":0,"335559740":240}"> </span></p> </div> <div> <p><span data-contrast="auto">We propose the use of a deep learning-based computer vision technique, termed instance segmentation [6-7], to automatically detect the exact pixel-by-pixel location of each plagioclase crystal (crystal masks) in thin section images. By re-training the models using a custom set of segmented geological thin section images, one can re-purpose these models for petrographic use, limitations notwithstanding based on the training data. The model outputs can then be used to measure the physical properties of the detected crystals, such as size and aspect ratio</span><span data-contrast="auto">,</span><span data-contrast="auto"> to automate the production of CSDs and aspect ratio distributions which are routinely used to interrogate the timescales of igneous processes.</span><span data-ccp-props="{"201341983":0,"335551550":6,"335551620":6,"335559739":0,"335559740":240}"> </span></p> </div> <div> <p><span data-ccp-props="{"201341983":0,"335551550":6,"335551620":6,"335559739":0,"335559740":240}"> </span></p> </div> <div> <p><span data-contrast="auto">The validity of our method will be showcased using a range of established volcanic and plutonic sample sets that have been previously well-characterised [4,8] through manual segmentation; these will include subglacial pillow basalts from Skuggafjoll and basaltic intrusions such as the Basement Sill in Antarctica and the Karlshamn dyke from Sweden. For sills, we make use of the correlation between plagioclase shape and crystallisation time [4] for rapid timescale determination straight from thin section photomicrographs to complement the information acquired from CSD’s. The vast amounts of data available from the automated segmentation of thin section scans are ripe for 3D shape studies over extensive sample suites to complement traditional textural approaches to timescales. These timescales will be linked to those obtained from diffusion chronometry such as Mg-in-plagioclase diffusion.</span><span data-ccp-props="{"201341983":0,"335551550":6,"335551620":6,"335559739":0,"335559740":240}"> </span></p> </div> <div> <p><span data-ccp-props="{"201341983":0,"335551550":6,"335551620":6,"335559739":0,"335559740":240}"> </span></p> </div> <div> <p><em><span data-contrast="auto">References:</span></em><span data-ccp-props="{"201341983":0,"335559739":0,"335559740":276}"> </span></p> </div> <div> <p><span data-contrast="auto">[1] Cashman KV and Marsh BD (1988) Contrib Mineral Petrol 99, 277–291 </span><span data-ccp-props="{"201341983":0,"335559739":0,"335559740":276}"> </span></p> </div> <div> <p><span data-contrast="auto">[2] Higgins MD (2000) American Mineralogist, 85, 1105-1116</span><span data-ccp-props="{"201341983":0,"335559739":0,"335559740":276}"> </span></p> </div> <div> <p><span data-contrast="auto">[3] Armienti P (2008) Reviews in Mineralogy and Geochemistry. 69. 623-649</span><span data-ccp-props="{"201341983":0,"335559739":0,"335559740":276}"> </span></p> </div> <div> <p><span data-contrast="auto">[4] Holness MB (2014) Contrib Mineral Petrol 168, 1076</span><span data-ccp-props="{"201341983":0,"335559739":0,"335559740":276}"> </span></p> </div> <div> <p><span data-contrast="auto">[5] Mangler MF et al. (2022) Contrib Mineral Petrol 177, 64</span><span data-ccp-props="{"201341983":0,"335559739":0,"335559740":276}"> </span></p> </div> <div> <p><span data-contrast="auto">[6] He K et al. (2017) IEEE International Conference on Computer Vision (ICCV) pp. 2980-2988</span><span data-ccp-props="{"201341983":0,"335559739":0,"335559740":276}"> </span></p> </div> <div> <p><span data-contrast="auto">[7] Qiao S et al (2021) Proc. IEEE/CVF Conf. CVPR pp. 10213-1022</span><span data-ccp-props="{"201341983":0,"335559739":0,"335559740":276}"> </span></p> </div> <div> <p><span data-contrast="auto">[8] Neave DA et al. (2014) Crystal Storage and Transfer in Basaltic Systems: The Skuggafjöll Eruption, Iceland, Journal of Petrology, Volume 55 pp.2311–2346</span><span data-ccp-props="{"201341983":0,"335559739":0,"335559740":276}"> </span></p> </div> </div>