Myungsun Kim

This study addresses two questions: what limits the training efficiency of deep seismic phase pickers on archive-scale data, and why does compressing training waveforms improve picking accuracy? Profiling a 1.45-million-parameter convolutional picker on STEAD (1.27 million traces, 98 GB HDF5) shows that storage input/output, rather than computation, accounts for 89% of each training step. Converting the archive to memory-mapped arrays eliminates this bottleneck (9.6× per-step speedup). Band-limited pre-encoding at a compression ratio (CR) of 4 offers an additional gain of 1.9×, yielding a 19.0× per-step acceleration over naive HDF5 access, 2.1× reduction in wall-clock training time relative to the optimized memmap baseline, and 76% reduction in storage. Compression improves accuracy: P-phase F1 rises by 7.1 percentage points to 0.976, whereas S-phase F1 decreases slightly (-0.8 percentage points, eliminated at CR = 2). Controlled ablations—a length-preserving low-pass filter with no improvement and a full-length dilated-convolution control reproducing the gain—identify effective receptive-field enlargement, not spectral denoising, as the dominant mechanism (gradient-based receptive fields: baseline, 5.3 s; dilated, 19.1 s; compressed, 18.0 s). The gain is constrained to receptive-field-limited convolutional pickers. With an adequate receptive field (verified by retraining PhaseNet), compression becomes approximately accuracy-neutral while retaining the speedup. On a second dataset (INSTANCE), the compression speedup is reproduced (the separate format-optimization gain is specific to STEAD’s per-trace layout), whereas the accuracy gain is dataset-specific. Against 1,827 blind manual P picks from a Korean deep-borehole network, the compressed model proves brittle under zero-shot transfer (P recall: 0.08 vs. 0.69 for the full-band model). Self-distillation from the full-band teacher using local waveforms—label-free in the sense that no pick labels are required—improves P-F1 from 0.14 to 0.59 and S-F1 from 0.42 to 0.81, reaching within 0.02 of pick-supervised fine-tuning at a GPU cost of only a few seconds.

We present an evaluation of the grain bulk modulus of rocks through experimental and theoretical methods. Unjacketed tests were performed on three rock types—Berea sandstone, Idaho sandstone, and Hwangdeung granite—to measure grain-scale compressibility. Theoretical estimates were obtained using mineral composition data and the Voigt–Reuss–Hill averaging method. To examine the role of pore structure, X-ray computed tomography (CT) was employed to visualize and quantify pore geometry, including isolated pores. Experimental results indicated material-dependent variation in grain bulk modulus not directly correlated with conventional mechanical properties. Theoretical values were consistently higher than those of experimental measurements, with deviations attributed to pore-scale characteristics unaccounted for in the theoretical model. X-ray CT analysis confirmed that isolated pores contribute to additional volumetric deformation, which is unaccounted for in volume fraction-based models. A correlation between theoretical estimates and experimental values was derived to improve the applicability of the model. Conclusively, we present a systematic methodology for accurately evaluating grain bulk modulus by integrating experimental and theoretical approaches, offering foundational data and insight critical for geomechanical modeling in deep subsurface applications such as radioactive waste disposal or underground liquid hydrogen storage.

Among many parameters needed to characterize subsurface rock mass related to underground storage and disposal projects, an important one is the heat production rate of the bedrock itself.The rock produces heat from the decay of radioactive isotopes consisting of minerals and its gamma-ray emissions, of which the magnitude is dominated by contents of major radioactive isotopes, e.g., uranium, thorium, and potassium. The heat production rate is generally calculated from the rock density and the radiogenic isotope contents, which can be measured from the spectrometry of drilled cores or rock fragments. However, such approaches are rarely available in deep boreholes because recovering rock samples from several hundred meters to a few kilometers is quite difficult.A recent geophysical logging technique for deep boreholes is available where the uranium, thorium, and potassium contents are measured from the gamma-ray spectrum. However, this technique requires the density to be measured separately, and the measurement depth of the equipment is still limited. As an alternative method, a natural gamma-ray logging tool was adopted to estimate briefly the heat production from the total gamma activity, which is relatively easy to measure. This study introduces the development of the proposed method for evaluating the heat production of a granitic rock mass with domestic commercial borehole logging tools, as well as its application and verifications in deep boreholes.

It is necessary to monitor, acquire, preprocess, and classify microseismic data to understand active faults or other causes of earthquakes, thereby facilitating the preparation of early-warning earthquake systems. Accordingly, this study proposes the application of machine learning for signal–noise classification of microseismic data from Pohang, South Korea. For the first time, unique microseismic data were obtained from the monitoring system of the borehole station PHBS8 located in Yongcheon-ri, Pohang region, while hydraulic stimulation was being conducted. The collected data were properly preprocessed and utilized as training and test data for supervised and unsupervised learning methods: random forest, convolutional neural network, and K-medoids clustering with fast Fourier transform. The supervised learning methods showed 100% and 97.4% of accuracy for the training and test data, respectively. The unsupervised method showed 97.0% accuracy. Consequently, the results from machine learning validated that automation based on the proposed supervised and unsupervised learning applications can classify the acquired microseismic data in real time.