Periodic Reporting for period 1 - FTAIGE (Towards automated fission-track age determination via artificial intelligence)
Période du rapport: 2021-10-01 au 2022-10-31
Understanding and predicting the continuous change of the environment is crucial for scientists, policy makers and ultimately for the entire society. Geochronology is the science of measuring the timing of processes on Earth and thus the key for understanding the past and making accurate predictions for the future. Techniques to model past and future events have evolved to an advanced state and geochronology carries the responsibility for providing accurate, precise and statistically robust age data for such models. Fission-track dating is a well-established geochronological method, which is based on the manual counting and length measurement of nuclear damage tracks (i.e. fission tracks) in minerals by means of optical microscopy. Due to the complexity of microscopic images and objects to be studied, the operator-based optical counting remains the most widely applied approach until these days. However, the manual approach has serious limitations especially with respect to the number of grains being dated as well as the comparability and reproducibility of the results. The main objective of the project is to automatize large parts of the slow and tedious manual procedure via cutting-edge image analysis techniques. It combines fission-track dating with artificial-intelligence (AI)-assisted image analysis exploiting the capability of convoluted neural nets (the AI) to be ‘taught’ to detect user defined objects in an image. In this specific application, the objects of interest are mineral grains and fission tracks on microphotographs. However, the overall result is not a solution to this sole scientific problem, but a flexible framework that can be freely used and refined by all geochronology laboratories to produce age data meeting the high requirements of cutting-edge research. Due to the flexibility of the framework – consisting of a set of scripts and a graphical user interface – it is intended to be applicable by researchers of any other scientific subdiscipline working on images. The successful objective regarding the communication of the results has been to reach a broad non-scientific audience, mainly in the disadvantaged groups of the society as well as to provide a stable basis for future research inside and far outside geosciences.
The general objective of the project was achieved on the basis of an extensive microscopic image database created by the researcher. In order to facilitate the marking of a sufficient amount of objects for training the neural network, large image stitches were needed, rather than individual microphotographs. Over 40 high-resolution image stitches were prepared, each consisting of individual images in the range of 100-s and a few 1000-s of images. A series of Python scripts were written that enable the following: (i) sort and rename microphotographs, regardless of settings and the the hardware used for acquisition (ii) stack and enhance a series of photographs and (iii) stitch and error-correct a series or multiple series of photographs. Furthermore, in order to enable the automatic detection of mineral grains, fission tracks and their surface marks (i.e. etch pits), a Python machine learning script was constructed, which provides the opportunity to use about a dozen of network architectures for automated learning. Several thousand objects were taught to the neural networks on a set of image stitches. In most grain mounts the trained network could detect over 90% of the objects of interest. Owing to our stitching technique, objects to be detected at different magnifications could be trained at different zoom levels. A further Python script was developed on the top of the neural network, which extracts the outlines of detected objects. This script does not only enable measuring lengths, but also other geometric features such as the orientation and the roundness of the detected polygons. Finally, as overlapping tracks were detected as one polygon in transmitted light images, these results were overlaid by the detected etch pits on reflected light images. Overall, owing to three separate neural networks, track contours can be counted within grain contours and track openings can be counted within the detected tracks, facilitating automated track detection.
The described results were presented by the researcher at the Meeting of the International Mineralogical Association (Lyon, France) and at the Conference of the International Association for Mathematical Geosciences (Nancy, France). He has also co-organized and actively participated in 10th Sedimentary Provenance Analysis short course at the Department of Sedimentology and Environmental Geology, Georg-August Universität Göttingen with one talk, one practical class and a laboratory tour.
Besides scientific presentatitons, the researcher has given three popular scientific talks to Hungarian minorities living in Romania and Slovakia, namely school classes and interested teachers. The main focus of these talks was geological time and they were organized in an interactive way – children had to form groups and solve task themselves with the assistance of the researcher related to the research project. A fourth talk was given to two school classes in Spain (age group 8-9) in English language. Over 30 questions were asked by young pupils, which extended the session to over two hours, and as such can be considered as a particular success. Furthermore, the researcher has given a presentation to the Society of Kurdish Enlightened Women, a group of ca. 15 women living in the Kurdish part of Iraq, dedicated to the role of geosciences in understanding the current global environmental challenges.
Besides popular scientific presentations, the researcher also participated in the communication campaign of the „Science is Wonderful” event organized by the European Commission and maintained a blog on his personal webpage and posted 7 posts in a popular scientific language.
The described results were presented by the researcher at the Meeting of the International Mineralogical Association (Lyon, France) and at the Conference of the International Association for Mathematical Geosciences (Nancy, France). He has also co-organized and actively participated in 10th Sedimentary Provenance Analysis short course at the Department of Sedimentology and Environmental Geology, Georg-August Universität Göttingen with one talk, one practical class and a laboratory tour.
Besides scientific presentatitons, the researcher has given three popular scientific talks to Hungarian minorities living in Romania and Slovakia, namely school classes and interested teachers. The main focus of these talks was geological time and they were organized in an interactive way – children had to form groups and solve task themselves with the assistance of the researcher related to the research project. A fourth talk was given to two school classes in Spain (age group 8-9) in English language. Over 30 questions were asked by young pupils, which extended the session to over two hours, and as such can be considered as a particular success. Furthermore, the researcher has given a presentation to the Society of Kurdish Enlightened Women, a group of ca. 15 women living in the Kurdish part of Iraq, dedicated to the role of geosciences in understanding the current global environmental challenges.
Besides popular scientific presentations, the researcher also participated in the communication campaign of the „Science is Wonderful” event organized by the European Commission and maintained a blog on his personal webpage and posted 7 posts in a popular scientific language.
In this research field, this is the first study that deals with the difficulties related to stitching high-resolution photographs of entire grain mounts, meaning precise stitching of cm-scale samples captured at a sub-micrometer resolution. The error-correction function within the image stitching script is a further innovative result of our efforts, as our scripts provide a flexible framework to work with any kind of photographs, independently of the research field. Furthermore, the automated detection of apatite grains, fission tracks and etch pits on any kind of grain mount via convolutional neural networks is now possible. The universal selectability of various neural networks is particularly innovative, as most applications of convolutional neural networks stick to one specific architecture, which is a serious limiting factor in applying the same scripts for different kinds of machine learning tasks. Regardless of the network type used, future users utilizing the FTAIGE framework will have to opportunity to select between architectures flexibly for their particular image series. Overall, the project results will lead to more reliable geological ages, which are key to understand the long term changes of the environment.