Search

Subduction megathrusts are the largest earthquakes occuring worldwide. Yet the generation of large subduction earthquakes is still poorly understood. Recent research revealed that aseismic deformation in the form of slow slip events (SSEs) might play a key rule in the build-up of these events. However, SSEs are hard to observe directly, due to there slow nature. One way to identify and study aseismic deformation is through co-occuring signals, for example, low-frequency earthquakes (LFEs). Yet these events are again difficult to observe due to their low signal-to-noise ratio and emergent onsets.In this project, we build machine learning models to identify low-frequency earthquakes. These models are more flexible and transferable than the commonly employed template matching techniques for LFE detection. We focus on deep learning based models, building upon their excellent performance for the picking and detection of regular seismicity. To train and evaluate these models we have compiled a collection of LFE datasets from multiple world region in a format tailored for machine learning. We integrate our LFE detector into the SeisBench library to allow easy application of the model in future studies.

Recent years have seen the development of several very powerful machine learning pickers for P and S waves. The recent development of the SeisBench platform (https://github.com/seisbench/seisbench) in combination with mixed regional teleseismic benchmark datasets published by the NEIC (USGS) and GEOFON (GFZ Potsdam) enabled the retraining of the most popular picker neural network models (PhaseNet and EQTransformer) optimised for global monitoring applications in the benchmark study of Münchmeyer et al (2021, J. Geophys. Res.). In this contribution we introduce a module scdlpicker, which connects SeisBench to SeisComP (https://www.seiscomp.de/) through a client submodule, which listens to new event detections from the regular SeisComP automatic detection system and triggers repicking of those events with any picker implemented in SeisBench, using the improved picks to trigger a relocation. The machine learning picks are subsequently available within the SeisComP GUI in case further manual refinement or checking is desired. We demonstrate application of this system with the GEOFON global earthquake monitoring service (https://geofon.gfz-potsdam.de/eqexplorer), evaluating the benefits of using the machine learning picker with respect to the conventional workflow relying on traditional pickers with respect to timeliness of reporting earthquakes and reduction of manual work load, and improvement in the number of high quality picks available for each event. The quantification of the uncertainty of machine learning picks is important when weighing the contribution of different picks in many location algorithms, yet this information is not readily available from machine learning pickers. They do return, however, a characteristic function (nominally the confidence in the pick), whose properties might correlate with the uncertainty of the pick. We will show whether and how the picking uncertainty correlates with properties of the characteristic function.

Detecting phase arrivals and pinpointing the arrival times of seismic phases in seismograms is crucial for many seismological analysis workflows. For land station data machine learning methods have already found widespread adoption. However, deep learning approaches are not yet commonly applied to ocean bottom data due to a lack of appropriate training data and models. Here, we compiled an extensive and labeled ocean bottom seismometer dataset from 15 deployments in different tectonic settings, comprising ~90,000 P and ~63,000 S manual picks from 13,190 events and 355 stations. We propose PickBlue, an adaptation ot the two popular deep learning networks EQTransformer and PhaseNet. PickBlue joint processes three seismometer recordings in conjunction with a hydrophone component and is trained with the waveforms in the new database. The performance is enhanced by employing transfer learning, where initial weights are derived from models trained with land earthquake data. PickBlue significantly outperforms neural networks trained with land stations and models trained without hydrophone data. The model achieves a mean absolute deviation (MAD) of 0.05 s for P waves and 0.12 s for S waves. We integrate our dataset and trained models into SeisBench to enable an easy and direct application in future deployments. KEY POINTS • We assembled a database of Ocean Bottom Seismometer waveforms and manual P and S picks, on which we train PickBlue, a deep learning picker. • Our picker significantly outperforms pickers trained with land-based data with confidence values reflecting the likelihood of outlier picks. • The picker and database are available in the SeisBench platform, allowing easy and direct application to OBS traces and hydrophone records.

Earthquakes are major hazards to humans, buildings and infrastructure. Early warning methods aim to provide advance notice of incoming strong shaking to enable preventive action and mitigate seismic risk. Their usefulness depends on accuracy, the relation between true, missed and false alerts, and timeliness, the time between a warning and the arrival of strong shaking. Current approaches suffer from apparent aleatoric uncertainties due to simplified modelling or short warning times. Here we propose a novel early warning method, the deep-learning based transformer earthquake alerting model (TEAM), to mitigate these limitations. TEAM analyzes raw, strong motion waveforms of an arbitrary number of stations at arbitrary locations in real-time, making it easily adaptable to changing seismic networks and warning targets. We evaluate TEAM on two regions with high seismic hazard, Japan and Italy, that are complementary in their seismicity. On both datasets TEAM outperforms existing early warning methods considerably, offering accurate and timely warnings. Using domain adaptation, TEAM even provides reliable alerts for events larger than any in the training data, a property of highest importance as records from very large events are rare in many regions., Comment: 13 pages main text, 16 pages appendix. This article has been accepted for publication in Geophysical Journal International Published by Oxford University Press on behalf of the Royal Astronomical Society

Machine learning (ML) approaches have demonstrated the ability to predict molecular spectra at a fraction of the computational cost of traditional theoretical chemistry methods while maintaining high accuracy. Graph neural networks (GNNs) are particularly promising in this regard, but different types of GNNs have not yet been systematically compared. In this work, we benchmark and analyze five different GNNs for the prediction of excitation spectra from the QM9 dataset of organic molecules. We compare the GNN performance in the obvious runtime measurements, prediction accuracy, and analysis of outliers in the test set. Moreover, through TMAP clustering and statistical analysis, we are able to highlight clear hotspots of high prediction errors as well as optimal spectra prediction for molecules with certain functional groups. This in-depth benchmarking and subsequent analysis protocol lays down a recipe for comparing different ML methods and evaluating dataset quality.

Seismic event detection and phase picking are the base of many seismological workflows. In recent years, several publications demonstrated that deep learning approaches significantly outperform classical approaches, achieving human-like performance under certain circumstances. However, as studies differ in the datasets and evaluation tasks, it is unclear how the different approaches compare to each other. Furthermore, there are no systematic studies about model performance in cross-domain scenarios, that is, when applied to data with different characteristics. Here, we address these questions by conducting a large-scale benchmark. We compare six previously published deep learning models on eight data sets covering local to teleseismic distances and on three tasks: event detection, phase identification and onset time picking. Furthermore, we compare the results to a classical Baer-Kradolfer picker. Overall, we observe the best performance for EQTransformer, GPD and PhaseNet, with a small advantage for EQTransformer on teleseismic data. Furthermore, we conduct a cross-domain study, analyzing model performance on data sets they were not trained on. We show that trained models can be transferred between regions with only mild performance degradation, but models trained on regional data do not transfer well to teleseismic data. As deep learning for detection and picking is a rapidly evolving field, we ensured extensibility of our benchmark by building our code on standardized frameworks and making it openly accessible. This allows model developers to easily evaluate new models or performance on new data sets. Furthermore, we make all trained models available through the SeisBench framework, giving end-users an easy way to apply these models., Journal of Geophysical Research: Solid Earth, 127 (1), ISSN:2169-9313, ISSN:0148-0227, ISSN:2169-9356

Ruptures of the largest earthquakes can last between a few seconds and several minutes. An early assessment of the final earthquake size is essential for early warning systems. However, it is still unclear when in the rupture history this final size can be predicted. Here we introduce a probabilistic view of rupture evolution - how likely is the event to become large - allowing for a clear and well-founded answer with implications for earthquake physics and early warning. We apply our approach to real time magnitude estimation based on either moment rate functions or broadband teleseismic P arrivals. In both cases, we find strong and principled evidence against early rupture predictability because differentiation between differently sized ruptures only occurs once half of the rupture has been observed. Even then, it is impossible to foresee future asperities. Our results hint towards a universal initiation behavior for small and large ruptures., 7 pages main text, 15 pages supplement

The end-Permian mass extinction occurred alongside a large swath of environmental changes that are often invoked as extinction mechanisms, even when a direct link is lacking. One way to elucidate the cause(s) of a mass extinction is to investigate extinction selectivity, as it can reveal critical information on organismic traits as key determinants of extinction and survival. Here we show that machine learning algorithms, specifically gradient boosted decision trees, can be used to identify determinants of extinction as well as to predict extinction risk. To understand which factors led to the end-Permian mass extinction during an extreme global warming event, we quantified the ecological selectivity of marine extinctions in the well-studied South China region. We find that extinction selectivity varies between different groups of organisms and that a synergy of multiple environmental stressors best explains the overall end-Permian extinction selectivity pattern. Extinction risk was greater for genera that had a low species richness, narrow bathymetric ranges limited to deep-water habitats, a stationary mode of life, a siliceous skeleton, or, less critically, calcitic skeletons. These selective losses directly link the extinctions to the environmental effects of rapid injections of carbon dioxide into the ocean–atmosphere system, specifically the combined effects of expanded oxygen minimum zones, rapid warming, and potentially ocean acidification.

Machine Learning (ML) methods have seen widespread adoption in seismology in recent years. The ability of these techniques to efficiently infer the statistical properties of large datasets often provides significant improvements over traditional techniques when the number of data are large (>millions of examples). With the entire spectrum of seismological tasks, e.g., seismic picking and detection, magnitude and source property estimation, ground motion prediction, hypocentre determination; among others, now incorporating ML approaches, numerous models are emerging as these techniques are further adopted within seismology.To evaluate these algorithms, quality controlled benchmark datasets that contain representative class distributions are vital. In addition to this, models require implementation through a common framework to facilitate comparison. Accessing both benchmark datasets, and integrating models built in such varying frameworks is currently a time-consuming process, hindering further advancement of ML techniques within seismology. These development bottlenecks also affect 'practitioners' seeking to deploy the latest models on seismic data, who may not want to necessarily learn entirely new ML frameworks to perform this task.We present SeisBench as a software package to tackle these issues. SeisBench is an open-source framework for deploying ML in seismology. SeisBench standardises access to both models and datasets, whilst also integrating a range of common processing and data augmentation operations through the API. Through SeisBench, users can access several seismological ML models and benchmark datasets available in the literature via a single interface. SeisBench is built to be extensible, with community involvement encouraged to expand the package. Having such frameworks available for accessing leading ML models forms an essential tool for seismologists seeking to iterate and apply the next generation of ML techniques to seismic data.

SUMMARYMagnitude estimation is a central task in seismology needed for a wide spectrum of applications ranging from seismicity analysis to rapid assessment of earthquakes. However, magnitude estimates at individual stations show significant variability, mostly due to propagation effects, radiation pattern and ambient noise. To obtain reliable and precise magnitude estimates, measurements from multiple stations are therefore usually averaged. This strategy requires good data availability, which is not always given, for example for near real time applications or for small events. We developed a method to achieve precise magnitude estimations even in the presence of only few stations. We achieve this by reducing the variability between single station estimates through a combination of optimization and machine learning techniques on a large catalogue. We evaluate our method on the large scale IPOC catalogue with >100 000 events, covering seismicity in the northern Chile subduction zone between 2007 and 2014. Our aim is to create a method that provides low uncertainty magnitude estimates based on physically meaningful features. Therefore we combine physics based correction functions with boosting tree regression. In a first step, we extract 110 features from each waveform, including displacement, velocity, acceleration and cumulative energy features. We correct those features for source, station and path effects by imposing a linear relation between magnitude and the logarithm of the features. For the correction terms, we define a non-parametric correction function dependent on epicentral distance and event depth and a station specific, adaptive 3-D source and path correction function. In a final step, we use boosting tree regression to further reduce interstation variance by combining multiple features. Compared to a standard, non-parametric, 1-D correction function, our method reduces the standard deviation of single station estimates by up to $57\, {\rm per\, cent}$, of which $17\, {\rm per\, cent}$ can be attributed to the improved correction functions, while boosting tree regression gives a further reduction of $40\, {\rm per\, cent}$. We analyse the resulting magnitude estimates regarding their residuals and relation to each other. The definition of a physics-based correction function enables us to inspect the path corrections and compare them to structural features. By analysing feature importance, we show that envelope and P wave derived features are key parameters for reducing uncertainties. Nonetheless the variety of features is essential for the effectiveness of the boosting tree regression. To further elucidate the information extractable from a single station trace, we train another boosting tree on the uncorrected features. This regression yields magnitude estimates with uncertainties similar to the single features after correction, but without using the earthquake location as required for applying the correction terms. Finally, we use our results to provide high precision magnitudes and their uncertainties for the IPOC catalogue.

Summary: Named Entity Recognition (NER) is an important step in biomedical information extraction pipelines. Tools for NER should be easy to use, cover multiple entity types, highly accurate, and robust towards variations in text genre and style. To this end, we propose HunFlair, an NER tagger covering multiple entity types integrated into the widely used NLP framework Flair. HunFlair outperforms other state-of-the-art standalone NER tools with an average gain of 7.26 pp over the next best tool, can be installed with a single command and is applied with only four lines of code. Availability: HunFlair is freely available through the Flair framework under an MIT license: https://github.com/flairNLP/flair and is compatible with all major operating systems. Contact:{weberple,saengema,alan.akbik}@informatik.hu-berlin.de, Comment: - Corrected author list - Updated project link

The estimation of earthquake source parameters, in particular magnitude and location, in real time is one of the key tasks for earthquake early warning and rapid response. In recent years, several publications introduced deep learning approaches for these fast assessment tasks. Deep learning is well suited for these tasks, as it can work directly on waveforms and can learn features and their relation from data.A drawback of deep learning models is their lack of interpretability, i.e., it is usually unknown what reasoning the network uses. Due to this issue, it is also hard to estimate how the model will handle new data whose properties differ in some aspects from the training set, for example earthquakes in previously seismically quite regions. The discussions of previous studies usually focused on the average performance of models and did not consider this point in any detail.Here we analyze a deep learning model for real time magnitude and location estimation through targeted experiments and a qualitative error analysis. We conduct our analysis on three large scale regional data sets from regions with diverse seismotectonic settings and network properties: Italy and Japan with dense networks (station spacing down to 10 km) of strong motion sensors, and North Chile with a sparser network (station spacing around 40 km) of broadband stations. We obtained several key insights. First, the deep learning model does not seem to follow the classical approaches for magnitude and location estimation. For magnitude, one would classically expect the model to estimate attenuation, but the network rather seems to focus its attention on the spectral composition of the waveforms. For location, one would expect a triangulation approach, but our experiments instead show indications of a fingerprinting approach. Second, we can pinpoint the effect of training data size on model performance. For example, a four times larger training set reduces average errors for both magnitude and location prediction by more than half, and reduces the required time for real time assessment by a factor of four. Third, the model fails for events with few similar training examples. For magnitude, this means that the largest events are systematically underestimated. For location, events in regions with few events in the training set tend to get mislocated to regions with more training events. These characteristics can have severe consequences in downstream tasks like early warning and need to be taken into account for future model development and evaluation.

The key task of earthquake early warning is to provide timely and accurate estimates of the ground shaking at target sites. Current approaches use either source or propagation based methods. Source based methods calculate fast estimates of the earthquake source parameters and apply ground motion prediction equations to estimate shaking. They suffer from saturation effects for large events, simplified assumptions and the need for a well known hypocentral location, which usually requires arrivals at multiple stations. Propagation based methods estimate levels of shaking from the shaking at neighboring stations and therefore have short warning times and possibly large blind zones. Both methods only use specific features from the waveform. In contrast, we present a multi-station neural network method to estimate horizontal peak ground acceleration (PGA) anywhere in the target region directly from raw accelerometer waveforms in real time.The three main components of our model are a convolutional neural network (CNN) for extracting features from the single-station three-component accelerograms, a transformer network for combining features from multiple stations and for transferring them to the target site features and a mixture density network to generate probabilistic PGA estimates. By using a transformer network, our model is able to handle a varying set and number of stations as well as target sites. We train our model end-to-end using recorded waveforms and PGAs. We use data augmentation to enable the model to provide estimations at targets without waveform recordings. Starting with the arrival of a P wave at any station of the network, our model issues real-time predictions at each new sample. The predictions are Gaussian mixtures, giving estimates of both expected value and uncertainties. The model can be used to predict PGA at specific target sites, as well as to generate ground motion maps.We analyze the model on two strong motion data sets from Japan and Italy in terms of standard deviation and lead times. Through the probabilistic predictions we are able to give lead times for different levels of uncertainty and ground shaking. This allows to control the ratio of missed detections to false alerts. Preliminary analysis suggest that for levels between 1%g and 10%g our model achieves multi-second lead times even for the closest stations at a false-positive rate below 25%. For an example event at 50 km depth, lead times at the closest stations with epicentral distances below 20 km are 6 s and 7.5 s. This suggests that our model is able to effectively use the difference between P and S travel time and accurately assess the future level of ground shaking from the first parts of the P wave. It additionally makes effective use of the information contained in the absence of signal at other stations.

Motivation Several recent studies showed that the application of deep neural networks advanced the state-of-the-art in named entity recognition (NER), including biomedical NER. However, the impact on performance and the robustness of improvements crucially depends on the availability of sufficiently large training corpora, which is a problem in the biomedical domain with its often rather small gold standard corpora. Results We evaluate different methods for alleviating the data sparsity problem by pretraining a deep neural network (LSTM-CRF), followed by a rather short fine-tuning phase focusing on a particular corpus. Experiments were performed using 34 different corpora covering five different biomedical entity types, yielding an average increase in F1-score of ∼2 pp compared to learning without pretraining. We experimented both with supervised and semi-supervised pretraining, leading to interesting insights into the precision/recall trade-off. Based on our results, we created the stand-alone NER tool HUNER incorporating fully trained models for five entity types. On the independent CRAFT corpus, which was not used for creating HUNER, it outperforms the state-of-the-art tools GNormPlus and tmChem by 5–13 pp on the entity types chemicals, species and genes. Availability and implementation HUNER is freely available at https://hu-ner.github.io. HUNER comes in containers, making it easy to install and use, and it can be applied off-the-shelf to arbitrary texts. We also provide an integrated tool for obtaining and converting all 34 corpora used in our evaluation, including fixed training, development and test splits to enable fair comparisons in the future. Supplementary information Supplementary data are available at Bioinformatics online.

Recent research showed that machine learning, in particular deep learning, can be applied with great success to a multitude of seismological tasks, e.g. phase picking and earthquake localization. One reason is that neural networks can be used as feature extractors, generating generally applicable representations of complex data. We employ a convolutional network to condense earthquake waveforms from a varying set of stations into a high dimensional vector, which we call event embedding. For each event the embedding is calculated from instrument-corrected waveforms beginning at the first P pick and updated continuously with incoming data. We employ event embeddings for real time magnitude estimation, earthquake localization and ground motion prediction, which are central tasks for early warning and for guiding rapid disaster response. We evaluate our model on the IPOC catalog for Northern Chile, containing ∼100,000 events with low uncertainty hypocenters and magnitude estimates. We split the catalog sequentially into a training and a test set, with the 2014 Iquique event (Mw 8.1) and its fore- and aftershocks contained in the test set. Following preliminary results the system achieves a test RMSE of 0.28 magnitude units (m.u.) and 35 km hypocentral distance 1 s after the first P arrival at the nearest station, which improves to 0.17 m.u. and 22 km after 5 s and 0.11 m.u. and 15 km after 25 s. As applications in the hazard domain require proper uncertainty estimates, we propose a probabilistic model using Gaussian mixture density networks. By analyzing the predictions in terms of their calibration, we show that the model exhibits overconfidence i.e. overly optimistic confidence intervals. We show that deep ensembles substantially improve calibration. To assess the limitations of our model and elucidate the pitfalls of machine learning for early warning in general, we conduct an error analysis and discuss mitigation strategies. Despite the size of our catalog, we observe issues with two kinds of data sparsity. First, we observe increased residuals for the source parameters of the largest events, as training data for these events is scarce. Second, similar inaccuracies occur in areas without events of a certain size in the training catalog. We investigate the impact of these limitations on the Iquique fore- and aftershocks.

Rule-based models are attractive for various tasks because they inherently lead to interpretable and explainable decisions and can easily incorporate prior knowledge. However, such systems are difficult to apply to problems involving natural language, due to its linguistic variability. In contrast, neural models can cope very well with ambiguity by learning distributed representations of words and their composition from data, but lead to models that are difficult to interpret. In this paper, we describe a model combining neural networks with logic programming in a novel manner for solving multi-hop reasoning tasks over natural language. Specifically, we propose to use a Prolog prover which we extend to utilize a similarity function over pretrained sentence encoders. We fine-tune the representations for the similarity function via backpropagation. This leads to a system that can apply rule-based reasoning to natural language, and induce domain-specific rules from training data. We evaluate the proposed system on two different question answering tasks, showing that it outperforms two baselines -- BIDAF (Seo et al., 2016a) and FAST QA (Weissenborn et al., 2017b) on a subset of the WikiHop corpus and achieves competitive results on the MedHop data set (Welbl et al., 2017)., Comment: ACL 2019