The b-value from the Gutenberg-Richter relation captures the ratio between small and large earthquakes triggering for a region. Several researchers have observed fluctuation in b-values prior to an earthquake, especially during the aftershock sequence analysis. By leveraging the ever-increasing earthquake database and the development in the field of deep learning computations, this study focuses on analyzing the spatial-temporal b-value series for Japan.

We follow a two step approach, in the first step a spatial-temporal b-value series is calculated from the earthquake catalog, which takes the shape of a series of [32 × 32] pixel images of spatial b-value distributions. On this we train an autoencoder, which compresses and then decompresses the input to learn the normal behavior and relationships within the data. In the second step, we then take the pixel by pixel reconstruction error as input for a Convolutional Dilated Network, whose model output can be interpreted as a quantity related to the earthquake probability. For this study we developed tow new architectures and compared them with common newtorks to be used on this kind of data. Furthermore we developed a novel training method to mimic real life use.
We expect that this study will improve the understanding of earthquake occurrences which could further be beneficial for Early Warning, rapid response and mitigation plans, especially for sustainable human habitats in earthquake-prone regions.

How neural activity in cortex is shaped by the underlying neural circuitry remains poorly understood. Recent experiments in ferrets have shown that at an early stage in development, spontaneous activity exhibits a modular correlation structure that is similar to a quantitative degree across multiple cortical areas (including both sensory and higher association areas) [1].

In this work, we investigate how this correlation structure evolves over the course of development in different cortical areas. We recorded the spontaneous activity with two-photon calcium imaging before, around and after the time of eye and ear canal opening, which takes place in ferrets around 30 days after birth, and computed the correlation structure of spontaneous activity. In all areas we observed a robust decrease in the correlation between nearby neurons and a pronounced increase in the dimensionality of spontaneous activity patterns over this developmental period, indicating a transition from an ordered, modular organization to a more fine-scaled, disordered organization. To explain these results, we study a linear recurrent neural network model. Assuming the recurrent interactions follow a local excitation and lateral inhibition (LELI) scheme, the model is able to reproduce the modular structure of spontaneous activity we observe in the early cortex [2]. We then analyse four different scenarios of possible network changes and analyse whether any of these can explain the developmental changes in nearest-neighbour correlations and dimensionality we observed in the data: i) a transition towards local inhibition and lateral excitation, ii) an increase in the heterogeneity of recurrent connections, iii) an increase in the sparsity of connections, and iv) a decrease in the effective strength of recurrent connections. We find that the changes predicted by scenario iv) provide the closest match to the changes we observe across development, suggesting that an effective weakening of recurrent connections over development is a major factor affecting the degree of modularity and how it changes across development.

The Compressed Baryonic Matter experiment at FAIR will conduct a systematic research program to explore the phase diagram of strongly interacting matter at highest net baryon densities and moderate temperatures. These conditions are to be created in collisions of heavy-ion beams with nuclear targets in the projectile beam energy range of 2 to 45 GeV/nucleon ( √s NN = 1.9 − 9 GeV), initially coming from the SIS 100 synchrotron (up to 14 GeV/nucleon) and in a next phase from SIS 300 enabling studies at the highest net baryon densities. It takes heavy-ion collisions between 105^ and 10^{^7} collisions per second to create extremely rare probes with previously unheard-of statistics in this energy range. Their complicated signatures include phase space distributions of strangeness, charm, di-leptons, and strange matter, as well as excitation functions of yields, for which there are currently little or no data available at FAIR energies. As the leptons escape the contact zone undistorted, vector-mesons decaying into dileptons provide an excellent probe of the hot and dense fireball. In the momentum range below 10 GeV/c, the RICH detector is intended to identify electrons and suppress pions for the study of these dielectronic decay channels of vector mesons. RICH detector build is in a projective geometry with focusing mirror elements and a photo detector. With a pion threshold for Cherenkov radiation of 4.65 GeV/c, CO2 is the radiator gas. For RICH, pion suppression of the order of 500–1000 is needed, and when combined with TRD, the order of 10^{^4}. The RICH detector goes through multiple steps in the event reconstruction process. Initially, the rebuilt STS tracks are projected to the mirrors. Track positions in the photodetector plane are then obtained by reflecting these tracks onto it. Subsequently, ring-shaped structures are identified among the hits in the RICH detector. The primary obstacle to ring detection in the CBM RICH detector stems from the high charged particle multiplicity of heavy ions. We use a novel neural network based approach to estimate the possible ring centers. Later a modified standard chi-square circle fitting method with reduced parameters is used to obtain the full set of ring parameters.

The High-rate Global Navigation Satellite System (HR-GNSS) instruments are devices that can measure ground displacement generated by an earthquake with high precision and detect the first seismic wave arrivals. By integrating HR-GNSS data with other sensors and models, we can improve the accuracy of earthquake assessments and provide valuable information for early warning and disaster preparedness. Our focus lies in developing deep-learning models leveraging HR-GNSS waveform data. These models significantly empower our capacity to detect, evaluate, and respond to large earthquakes. Yet, the rapid analysis of HR-GNSS data using deep learning algorithms remains a current challenge. To overcome this challenge, it is crucial to have access to large and high-quality datasets. Unfortunately, GNSS stations are not distributed enough in all the regions, which can lead to data gaps. Additionally, the presence of noise in GNSS recordings particularly impacts data quality, especially for earthquakes measuring below magnitude 7. As a consequence, our training of Deep Learning (DL) models primarily relies on the data available from the largest earthquakes—events that occur less frequently and provide a limited dataset, making it less representative for model training. Therefore, we have faced a lack of data and have used both synthetic and real HR-GNSS data for model training, validation, and testing. Our investigation explores how diverse factors—such as noise, earthquake magnitude, station density, distance from the epicenter, and duration of the signal—affect the performance of our models. Our ultimate aim is to generalize this methodology for real-time monitoring of large earthquakes across diverse tectonic regions.

In ferret visual cortex, spontaneous activity prior to eye-opening is organized into large-scale, modular patterns in the absence of long-range horizontal projections. This correlated activity reveals endogenous networks that predict aspects of future orientation selectivity [1].Previous modelling works have shown that the long-range correlations observed in these networks can arise purely from locally connected neurons through multi-synaptic interactions [1,2]. Here we seek to test the extent to which cortical activity is organized through lateral interactions using localized optogenetic perturbations in vivo.

We first constructed a recurrent neural network model of rectified excitatory and inhibitory units, with effective local heterogeneous Mexican hat connections. Our model predicts that perturbing a small region of inhibitory neurons leads to long-range reorganization in the spatial patterns of ongoing activity as well as network correlation structure. Notably, the degree of disruption to correlation structure depends on the perturbation location and can be predicted from the stimulation site’s overlap with the leading principal components of baseline spontaneous activity. Only a fraction of variance of perturbed activity patterns overlaps with the leading variance components of spontaneous activity, suggesting a shift in the activity manifold towards novel patterns after local disruption.

To test these predictions, we virally expressed GCaMP6s in excitatory neurons and Chrimson-ST in inhibitory neurons in layer 2/3 of young ferret visual cortex, allowing us to optogenetically activate small regions (~500μm diameter) of inhibitory neurons and simultaneously record widefield calcium activity. 

In line with model predictions, local optogenetic inhibitory perturbations induce a large-scale reorganization of activity, even in areas up to 2mm away from stimulation site. Perturbing locations that overlap better with prominent spontaneous neural modes leads to a larger degree of disruption. Furthermore, the variance in perturbed activity patterns can only partially be explained by spontaneous components, confirming that local perturbation indeed introduces new patterns.

Our results are consistent with the presence of strongly coupled E and I networks in early cortex, and demonstrate that network behaviour is an emergent property with local activity exerting specific and large-scale influences. 

Addressing the Core Challenge:

In developmental biology, particularly within biophysics-driven mechanistic modeling, a key challenge is parameter inference due to limited quantitative data.
Qualitative experimental observations in systems like the mouse blastocyst inner cell mass (ICM) need transformation into relevant quantitative data for effective modeling.
Innovative Methodological Contributions:
Our study contrasts two novel approaches for parameter inference in modeling the mouse ICM,
central to understanding early embryonic cell-fate specification:
    Simulated Annealing-Inspired Optimization Technique:
    Incorporates a simulated annealing-like method infused with reinforcement learning principles,
    representing an innovative blend of classical and computational optimization.
    AI-Driven Simulation-Based Inference (SBI):
    Leverages an AI-powered SBI framework, inspired by the Sequential Neural Posterior Estimation (SNPE) algorithm,
    to align simulations with empirical data for refined parameter estimation.
Comparative Insights:
By comparing these methods, we shed light on their unique strengths and applicability in developmental biology,
aiming to enhance the precision and predictive capability of biophysical models.