We conducted a series of triaxial creep experiments on shale specimens coming from Tournemire, France, using the stress‐stepping method up to failure, at a confining pressure of 80 MPa, on two orientations (parallel and perpendicular to bedding), and at temperatures of 26 and 75 °C. In these week‐long experiments, stress, strains, and P wave ultrasonic velocities were recorded (quasi‐) continuously. The strength at creep failure of Tournemire shale was ~70% higher than the peak strength measured during constant strain rate (~10⁻⁷/s) experiments, and failure was reached at larger strains. An overall transition from P wave velocity increase at moderate differential stress to P wave velocity decrease closer to brittle failure was also observed. At a smaller timescale, P wave velocities initially decreased and then increased gradually during each step of creep deformation. The magnitude of these variations showed important (i) stress, (ii) orientation, and (iii) temperature dependences: larger increase was observed for P wave propagating along the main compressive stress orientation, larger decrease for P wave propagating perpendicular to it, and a changing behavior enhanced at a higher temperature. Scanning electron microscopy performed postmortem revealed evidence of time‐dependent pressure solution, localized compaction, crack growth, and sealing/healing. Our data reveal that shale deformation is highly stress sensitive only in a narrow stress domain where stress corrosion cracking‐induced brittle dilatant creep deformation is dominant. At stresses below, pressure solution compaction creep dominates the deformation and shales compact, consolidate, and heal. This has important implications for the mechanics of shallow fault zones and accretionary prisms.

Geoscientists mainly identify subsurface geologic features using exploration-derived seismic data. Classification or segmentation of 2D/3D seismic images commonly relies on conventional deep learning methods for image recognition. However, complex reflections of seismic waves tend to form high-dimensional and multi-scale signals, making traditional convolutional neural networks (CNNs) computationally costly. Here we propose a highly efficient and resource-saving CNN architecture (SeismicPatchNet) with topological modules and multi-scale-feature fusion units for classifying seismic data, which was discovered by an automated data-driven search strategy. The storage volume of the architecture parameters (0.73 M) is only ~2.7 MB, ~0.5% of the well-known VGG-16 architecture. SeismicPatchNet predicts nearly 18 times faster than ResNet-50 and shows an overwhelming advantage in identifying Bottom Simulating Reflection (BSR), an indicator of marine gas-hydrate resources. Saliency mapping demonstrated that our architecture captured key features well. These results suggest the prospect of end-to-end interpretation of multiple seismic datasets at extremely low computational cost.

We conducted two brittle creep experiments on shale samples under upper crustal conditions (confining pressure of 80 MPa at 26°C and 75°C). We deformed the samples to failure, with bedding oriented perpendicular to the maximum compressive stress direction, using the stress‐stepping methodology. In both experiments, the failure stress was ~64% higher than the short‐term peak strength. Throughout each differential stress step, ultrasonic wave velocities initially decreased and then gradually increased with deformation/time. The magnitude of these variations depends both on the direction of measurement with respect to the bedding and the temperature, and it is largest for velocities measured parallel to the bedding and at high temperature. Elastic wave anisotropy was completely reversed at 75°C, following a limited amount of axial strain (~0.6%). Scanning electron microscope investigation confirmed evidence of a time‐dependent pressure solution, localized compaction, crack sealing/healing, and mineral rotation. Our observations reveal that elastic anisotropy can evolve rapidly in both time and space, which has implications on the stress state and its rotation near fault zones.

The real-time detection and analysis of seismic signals is crucial in geophysics research, especially when it comes to monitoring catastrophic events. We present an evolutionary deep learning method that yields a model named MCU-Quake. This model encodes the discrimination process as a single numerical value, offering interpretability with only 2693 parameters. Trained on raw seismic waveforms from Utah, USA, MCU-Quake demonstrates its generalization capability across a global natural earthquake dataset. Notably, the model effectively identifies typical explosions during the Russia-Ukraine war in Europe. The knowledge to discriminate between ambient noise, explosions and natural earthquakes can be represented by values of −5.01 (std: 1.14), 1.96 (std: 0.36), 1.01 (std: 0.49), respectively. The model can be deployed on Internet of Things (IoT) devices, including most microcontrollers, which are constrained by limited computational resources (kilo-bytes of memory) and energy consumption (micro-Watts). The results indicate the prospect of on-site missions of artificial intelligent sensors.