Decades of seismological observations have highlighted the variability of foreshock occurrence prior to natural earthquakes, making thus difficult to track how earthquakes start. Here, we report on three stick-slip experiments performed on cylindrical samples of Indian metagabbro under upper crustal stress conditions (30-60 MPa). Acoustic emissions (AEs) were continuously recorded by 8 calibrated acoustic sensors during the experiments. Seismological parameters of the detected AEs (-8.8 <= Mw <= -7 ) follow the scaling law between moment magnitude and corner frequency that characterizes natural earthquakes. AE activity always increases towards failure and is found to be driven by along fault slip velocity. The stacked AE foreshock sequences follow an inverse power-law of the time to failure (inverse Omori), with a characteristic Omori time c inversely proportional to normal stress and nucleation length. AEs moment magnitudes increase towards failure, as manifested by a decrease in b-value from ~ 1 to ~ 0.5 at the end of the nucleation process. During nucleation, the averaged distance of foreshocks to mainshock continuously decreases, highlighting the fast migration of foreshocks towards the mainshock epicenter location, and stabilizing at a distance from the latter compatible with the predicted Rate-and-State nucleation size. Finally, the seismic component of the nucleation phase is orders of magnitude smaller than that of its aseismic component, which suggests that foreshocks are the byproducts of a process almost fully aseismic. Seismic/aseismic energy release ratio continuously increases during nucleation, which starts as a fully aseismic process and evolves towards a cascading process.

Recent experiments systematically explore rock friction under crustal earthquake conditions revealing that faults undergo abrupt dynamic weakening. Processes related to heating and weakening of fault surface have been invoked to explain pronounced velocity weakening. Both contact asperity temperature $T_a$ and background temperature $T$ of the slip zone evolve significantly during high velocity slip due to heat sources (frictional work), heat sinks (e.g. latent heat of decomposition processes) and diffusion. Using carefully calibrated High Velocity Rotary Friction experiments, we test the compatibility of thermal weakening models: (1) a model of friction based only on $T$ in an extremely simplified, Arrhenius-like thermal dependence; (2) a flash heating model which accounts for evolution of both $V$ and $T$; (3) same but including heat sinks in the thermal balance; (4) same but including the thermal dependence of diffusivity and heat capacity. All models reflect the experimental results but model (1) results in unrealistically low temperatures and models (2) reproduces the restrengthening phase only by modifying the parameters for each experimental condition. The presence of dissipative heat sinks in (3) significantly affects $T$ and reflects on the friction, allowing a better joint fit of the initial weakening and final strength recovery across a range of experiments. Temperature is significantly altered by thermal dependence of (4). However, similar results can be obtained by (3) and (4) by adjusting the energy sinks. To compute temperature in this type of problem we compare the efficiency of three different numerical solutions (Finite differences, wavenumber summation, and discrete integral).

Convolutional Neural Networks (CNNs) can detect patterns that are otherwise difficult to identify and have been shown to excel in predicting fault characteristics in laboratory shear experiments and slow slip \emph{in situ}. Here we show that during the precursory phase of some natural earthquakes, a subtle change in the seismic background signal occurs that can be identified by a suitably designed CNN, and used as a probabilistic forecasting tool. We use 31 earthquakes of $\textrm{M}_\textrm{w}\geq 6$ in Japan and vicinity, between March 2012 and February 2020, all recorded by station IU MAJO (Japan main island) except one recorded by station IU MA2 (Kamtchatka). The CNN is trained on 24 events, where a 16 mn time window preceding each earthquake is labelled as precursory (presumably containing a strong precursory signal), and another 16 mn time window far from the time of earthquake occurrence is labelled as noise (presumably containing weak or no precursory signal). The 7 remaining events were used for testing. The CNN achieves 98\% training accuracy and a 96\% testing accuracy in discriminating noise and precursor windows. Time windows in the $\sim 3$ hours preceding the earthquakes are progressively interpreted by the model as precursors as earthquake time approaches. To characterize the signal detected by the CNN, we analyse spectra from noise and from precursory windows. Discriminative features appear most dominant over a frequency range of $\approx$ 0.1-0.9 Hz (in particular $\approx$0.16 and $\approx$0.21 Hz) coinciding with microseismic noise and recent observations of broadband slow earthquake signal \cite{masuda_bridging_2020}.