Conference Agenda

The structural integrity of geological materials are closely tied to their porosity. Accurate knowledge of microspores in potential host rocks such as Opalinus Clay is essential for assessing their physical properties, including permeability and strength. Traditional methods for porosity analysis, such as mercury intrusion porosimetry (MIP) and gas pycnometry, provide valuable quantitative data on porosity and pore size distribution but do not offer insights into pore morphology or spatial distribution.

A methodological advancement comes with the combination of broad ion beam (BIB) milling and scanning electron microscopy (SEM), which allows for the visualization of pores at the nanoscale and facilitates detailed analysis of pore structures. However, segmenting pores from BIB-SEM images poses challenges due to the complexity of the images and the variability in pore shapes and sizes. This task is further complicated by the limited resolution of SEM and the subjective nature of manual pore identification.

To address these challenges, machine learning (ML) has emerged as a useful tool for automating the segmentation of pores from BIB-SEM images. We explore the use of conditional random fields (CRFs) as an ensemble method that improves segmentation by utilizing spatial and contextual information within the images. CRFs enhance segmentation accuracy and offer a robust framework for integrating results from multiple ML-classifiers. This probabilistic approach not only refines the segmentation accuracy but also enables the assessment of uncertainty levels in segmented pores, which is beneficial for accurately interpreting the microstructural properties.

In the past years numerous machine learning based applications have been introduced to the field of seismology. These applications for example address issues such as earthquake detection, event classification, feature extraction and waveform data analysis.

In this study we focus on the denoising of waveform data by separating the seismic signal from different noise sources. Machine learning models are able to recognize noise patterns and can effectively suppress unwanted noise, enhancing the quality of the waveform signals. A deep learning based denoising autoencoder algorithm is tested on regional and teleseismic seismological and hydroacoustic datasets, which are compiled from the International Monitoring System of the Comprehensive Nuclear-Test-Ban Treaty Organisation. We focus on seismic and hydroacoustic stations which can be relevant to investigate North Korean nuclear tests.

We investigate the performance of different denoising autoencoder models, for short- and long waveform periods, trained on the complete station network as well as on individual stations. We investigate if the denoised waveform signals are useful for seismic source analysis and if the denoised waveforms can reliably be used in further analysis steps, such as the comparison of computed array beams, seismic phase picking and or amplitude estimation.

The declared North Korean nuclear tests are a suitable benchmark test set, as they have been extensively researched and their source type and location can be assumed known. Further the verification of the source type is of particular interest for potential nuclear tests under international law.

Monitoring local seismic events is among the responsibilities of the German Federal Seismic Survey. This entity comprises a seismological subdepartment responsible for overseeing the operations of the German Regional Seismic Network and a data center tasked with collecting, archiving, and distributing continuous seismological and infrasound waveform data. As the amount of recorded seismic data dramatically increases every year, the imperative for an appropriate automatic real-time monitoring system becomes apparent. Leveraging advances in deep-learning methods in seismology, we develop a Python wrapper for the automatic estimation of hypocenter, magnitude and first-motion polarity in real time. To assess the performance of our algorithm, we compare the resulting event locations with catalogs of manually located events, with promising outcomes.

Aleatoric uncertainty, inherent in the variability of data itself, presents a significant challenge in predictive modeling, especially in scenarios with intrinsic randomness and noise. Traditionally viewed as irreducible, this type of uncertainty fundamentally limits the precision of predictions, as it is directly tied to the stochastic nature of the underlying data. However, this research proposes a methodology that combines clustering with subsequent predictive modeling to mitigate the effects of aleatoric uncertainty, thereby enhancing the transparency and reliability of model outputs. Our approach begins with a clustering process, where data points are grouped based on similarity in features to form homogeneous subsets. Following clustering, we employ quantile random forests on each subset rendering the modeling tailored to each cluster's specific characteristics. This strategy allows for the models to not only be more sensitive to the subtle nuances within a group but also more robust against the noise inherent in the dataset. Finally, we estimate heat flow over continental Africa. Through extensive quantitative analysis, this study demonstrates that while aleatoric uncertainty is indeed irreducible from a theoretical standpoint, practical interventions like quality data acquisition combined with clustering can effectively diminish its impact on predictive accuracy.

Geological mapping is essential for understanding the Earth's surface and subsurface structures, aiding resource exploration, environmental monitoring, and hazard assessment. Semantic segmentation, a computer vision technique, has shown promise in automating geological mapping processes by classifying image pixels into meaningful categories. This study explores the integration of semantic segmentation into geological mapping workflows by leveraging an artificially blended texture dataset.

Traditional geological mapping relies on extensive fieldwork in combination with manual aerial or satellite imagery interpretation, which can be time-consuming and subjective. Semantic segmentation can efficiently classify geological features by learning distinctive patterns and textures from data. However, obtaining high-quality datasets for this purpose is challenging due to the heterogeneous nature of geological formations and limited ground truth data.

We address this challenge by employing an artificially blended texture dataset that combines real-world geological textures. This blended dataset aims to enrich the training data with diverse texture and geological feature combinations, potentially enhancing the model's ability to generalize to unseen terrain conditions. Moreover, it reduces the potential for label bias by eliminating the need for manual delineation of label classes in the image, instead relying on generated borders.

Through experimental evaluation, we explore the effectiveness of semantic segmentation with the blended texture dataset in accurately delineating geological units and structures. We also discuss the implications of incorporating semantic segmentation into geological mapping workflows at the Baltic cliff coast, including its potential for improving mapping efficiency, reducing human bias, and facilitating remote sensing data integration with geological interpretations.

Machine learning approaches are increasingly used to predict groundwater levels, with local models for single monitoring wells currently being state of the art. Global models enable training and forecasting at multiple monitoring wells simultaneously, incorporating dynamic (e.g., meteorological) and static (e.g., hydro(geo)logical) features. These models can generalize predictions to wells with similar site characteristics and offer computational scaling benefits by requiring only one model for a larger area.

This study presents two global machine-learning models for short-term groundwater level prediction (up to 12 weeks): the Temporal Fusion Transformer (TFT) and Neural Hierarchical Interpolation for Time Series Forecasting (N-HiTS). The TFT combines recurrent neural networks with an attention mechanism and can determine the significance of individual input variables (feature importance). The N-HiTS model uses a fork architecture with multiple stacks to model different data frequencies, enhancing prediction accuracy.

We used a dataset of approximately 5300 monitoring wells across Germany, with groundwater levels from 1990 to 2016 (around 4.5 million values). Input features included groundwater levels, meteorological parameters, and site-specific environmental features such as hydro(geo)logical, soil, and spatial characteristics.

The TFT model showed a median NSE of 0.34, while the N-HiTS model performed better with a median NSE of 0.5 for the 12-week forecast. Around 25% of the test sites achieved an NSE over 0.68. Key features for forecast quality included historical groundwater levels, precipitation, the standard deviation of groundwater levels, and major hydrogeological districts. The topographical wetness index was the most important static feature, though its impact on model performance was minimal.