Foundation Models are designed to serve as versatile embedding machines, with strong zero shot capabilities and superior generalization performance when fine-tuned on diverse downstream tasks. While this is largely true for language and vision foundation models, we argue that the inherent diversity of time series data makes them less suited for building effective foundation models. We demonstrate this using forecasting as our downstream task. We show that the zero-shot capabilities of a time series foundation model are significantly influenced and tied to the specific domains it has been pretrained on. Furthermore, when applied to unseen real-world time series data, fine-tuned foundation models do not consistently yield substantially better results, relative to their increased parameter count and memory footprint, than smaller, dedicated models tailored to the specific forecasting task at hand.

Time Series Foundation Models (TSFMs) have shown promising zero-shot generalization across diverse forecasting tasks. However, their robustness to continual adaptation remains underexplored. In this work, we investigate the extent to which TSFMs suffer from catastrophic forgetting when fine-tuned sequentially on multiple datasets. Using synthetic datasets designed with varying degrees of periodic structure, we measure the trade-off between adaptation to new data and retention of prior knowledge. Our experiments reveal that, while fine-tuning improves performance on new tasks, it often causes significant degradation on previously learned ones, illustrating a fundamental stability-plasticity dilemma.

This paper is dedicated to the analysis of forward backward stochastic differential equations driven by a L{é}vy process. We assume that the generator and the terminal condition are path-dependent and satisfy a local Lipschitz condition. We study solvability and Malliavin differentiability of such BSDEs. The proof of the existence and uniqueness is done in three steps. First of all, we truncate and localize the terminal condition and the generator. Then we use an iteration argument to get bounds for the solutions of the truncated BSDE (independent from the level of truncation). Finally, we let the level of truncation tend to infinity. A stability result ends the proof. The Malliavin differentiability result is based on a recent characterisation for the Malliavin Sobolev space D 1,2 by S. Geiss and Zhou.

Dense ground displacement measurements are crucial for geological studies but are impractical to collect directly. Traditionally, displacement fields are estimated using patch matching on optical satellite images from different acquisition times. While deep learning-based optical flow models are promising, their adoption in ground deformation analysis is hindered by challenges such as the absence of real ground truth, the need for sub-pixel precision, and temporal variations due to geological or anthropogenic changes. In particular, we identify that deep learning models relying on explicit correlation layers struggle at estimating small displacements in real-world conditions. Instead, we propose a model that employs iterative refinements with explicit warping layers and a correlation-independent backbone, enabling sub-pixel precision. Additionally, a non-convex variant of Total Variation regularization preserves fault-line sharpness while maintaining smoothness elsewhere. Our model significantly outperforms widely used geophysics methods on semi-synthetic benchmarks and generalizes well to challenging real-world scenarios captured by both medium- and high-resolution sensors. Project page: this https URL

Cherenkov Telescope Array Observatory (CTAO) is the next-generation ground-based gamma-ray observatory operating in the energy range from 20 GeV up to 300 TeV, with two sites in La Palma (Spain) and Paranal (Chile). It will consist of telescopes of three sizes, covering different parts of the large energy range. We report on the performance of Large-Sized Telescope prototype (LST-1) in the detection and characterization of extragalactic gamma-ray sources, with a focus on the reconstructed gamma-ray spectra and variability of classical bright BL Lacertae objects, which were observed during the early commissioning phase of the instrument. LST-1 data from known bright gamma-ray blazars - Markarian 421, Markarian 501, 1ES 1959+650, 1ES 0647+250, and PG 1553+113 - were collected between July 10, 2020, and May 23, 2022, covering a zenith angle range of 4 deg to 57 deg. The reconstructed light curves were analyzed using a Bayesian block algorithm to distinguish the different activity phases of each blazar. Simultaneous Fermi-LAT data were utilized to reconstruct the broadband $\gamma$-ray spectra for the sources during each activity phase. High-level reconstructed data in a format compatible with gammapy are provided together with measured light curves and spectral energy distributions (SEDs) for several bright blazars and an interpretation of the observed variability in long and short timescales. Simulations of historical flares are generated to evaluate the sensitivity of LST-1. This work represents the first milestone in monitoring bright BL Lacertae objects with a CTAO telescope.

The stochastic gravitational-wave background from compact binary coalescences is expected to be the first detectable stochastic signal via cross-correlation searches with terrestrial detectors. It encodes the cumulative merger history of stellar-mass binaries across cosmic time, offering a unique probe of the high-redshift Universe. However, predicting the background spectrum is challenging due to numerous modeling choices, each with distinct uncertainties. In this work, we present a comprehensive forecast of the astrophysical gravitational-wave background from binary black holes, binary neutron stars, and neutron star-black hole systems. We systematically assess the impact of uncertainties in population properties, waveform features, and the modeling of the merger rate evolution. By combining all uncertainties, we derive credible bands for the background spectrum, spanning approximately an order of magnitude in the fractional energy density. These results provide thorough predictions to facilitate the interpretation of current upper limits and future detections.

The EUNIS habitat classification is crucial for categorising European habitats, supporting European policy on nature conservation and implementing the Nature Restoration Law. To meet the growing demand for detailed and accurate habitat information, we provide spatial predictions for 260 EUNIS habitat types at hierarchical level 3, together with independent validation and uncertainty analyses. Using ensemble machine learning models, together with high-resolution satellite imagery and ecologically meaningful climatic, topographic and edaphic variables, we produced a European habitat map indicating the most probable EUNIS habitat at 100-m resolution across Europe. Additionally, we provide information on prediction uncertainty and the most probable habitats at level 3 within each EUNIS level 1 formation. This product is particularly useful for both conservation and restoration purposes. Predictions were cross-validated at European scale using a spatial block cross-validation and evaluated against independent data from France (forests only), the Netherlands and Austria. The habitat maps obtained strong predictive performances on the validation datasets with distinct trade-offs in terms of recall and precision across habitat formations.

Reliable seismicity catalogs are essential for seismology. Following phase picking, phase association groups arrivals into sets with consistent origins (i.e., events), determines event counts, and identifies outlier picks. To handle the substantial increase in the quantity of seismic phase picks from improved picking methods and larger deployments, several novel phase associators have recently been proposed. This study presents a detailed benchmark analysis of five seismic phase associators, including classical and machine learning-based approaches: PhaseLink, REAL, GaMMA, GENIE, and PyOcto. We use synthetic datasets mimicking real seismicity characteristics in crustal and subduction zone scenarios. We evaluate performance for different conditions, including low- and high-noise environments, out-of-network events, very high event rates, and variable station density. The results reveal notable differences in precision, recall, and computational efficiency. GENIE and PyOcto demonstrate robust performance, with almost perfect performance for most scenarios. Only for the most challenging conditions with high noise levels and event rates, performance drops but still maintains F1 scores above 0.8. PhaseLink's performance declines with noise and event density, particularly in subduction zones, dropping to near zero in the most complex cases. GaMMA outperforms PhaseLink but struggles with accuracy and scalability in high-noise, high-density scenarios. REAL performs reasonably but loses recall under extreme conditions. PyOcto and PhaseLink show the quickest runtimes for smaller-scale datasets, while REAL and GENIE are more than an order of magnitude slower for these cases. At the highest pick rates, GENIE's runtime disadvantage diminishes, matching PyOcto and scaling effectively. Our results can guide practitioners compiling seismicity catalogs and developers designing novel associators.

Earthworms are key drivers of soil function, influencing organic matter turnover, nutrient cycling, and soil structure. Understanding the environmental controls on their distribution is essential for predicting the impacts of land use and climate change on soil ecosystems. While local studies have identified abiotic drivers of earthworm communities, broad-scale spatial patterns remain underexplored. We developed a multi-species, multi-task deep learning model to jointly predict the distribution of 77 earthworm species across metropolitan France, using historical (1960-1970) and contemporary (1990-2020) records. The model integrates climate, soil, and land cover variables to estimate habitat suitability. We applied SHapley Additive exPlanations (SHAP) to identify key environmental drivers and used species clustering to reveal ecological response groups. The joint model achieved high predictive performance (TSS >= 0.7) and improved predictions for rare species compared to traditional species distribution models. Shared feature extraction across species allowed for more robust identification of common and contrasting environmental responses. Precipitation variability, temperature seasonality, and land cover emerged as dominant predictors of earthworm distribution. Species clustering revealed distinct ecological strategies tied to climatic and land use gradients. Our study advances both the methodological and ecological understanding of soil biodiversity. We demonstrate the utility of interpretable deep learning approaches for large-scale soil fauna modeling and provide new insights into earthworm habitat specialization. These findings support improved soil biodiversity monitoring and conservation planning in the face of global environmental change.

The Atacama segment in Northern Chile (24°S to 31°S) is a mature seismic gap with no major event (Mw>8) since 1922. In addition to regular seismicity, around the subducting Copiapó ridge, the region hosts seismic swarms, and shallow and deep slow slip events. To characterize the fine structure of this seismic gap and its seismic-aseismic interplay, we instrumented the region with almost 200 seismic and geodetic stations. Using machine learning, we derived a dense, high-resolution seismicity catalog, encompassing over 165,000 events with double-difference relocated hypocenters. Our catalog details the outer rise, interface, intraslab, crustal and mantle wedge seismicity. We infer a detailed slab geometry, showing that the flat slab is dipping towards the south with a narrower extent along dip. The slab geometry controls the intraslab seismicity, with cross-cutting activity in the region of highest bending and a downdip limit around 105 km slab depth. Our catalogue exhibits significant seismicity in the mantle wedge upper corner between 28°S and 31°S, highlighting the brittle behavior of the cold nose. On the subduction interface, interplate locking controls the updip end of the seismicity, with seismicity extending closer to the trench in low-locking areas. On fine scales, resolved by relative uncertainties below 50 m, the subduction interface has a complex 3D structure, showing a fractal distribution of seismic patches down to a scale of tens of meters. Our results provide a holistic view of this complex subduction zone, while at the same time giving insights into fine-scale structures and processes.

The dSphs around the Milky Way are commonly considered as systems that are supported by velocity dispersion against self-gravitation. They have been long accounted among the best targets to search for indirect DM signatures in the GeV-to-TeV gamma-rays due to absence of astrophysical gamma-ray foreground or background emission. We present forecasts on the sensitivity of the future CTAO for the search for annihilating or decaying DM in such targets. We perform an original selection of candidates out of the current catalog of known objects, including both classical and ultra-faint targets. For each of them, we calculate the expected amount of DM using the most updated and complete available samples of photometric and spectroscopic data of member stars, adopting a common framework of data treatment for both classes of objects. In this way, we are able to generate novel astrophysical factor profiles for general indirect DM searches that we compare with the current literature. Out of a starting sample of 64 dSphs, we highlight the 8 most promising targets - DraI, CBe, UMaII, UMi and Wil1 in the Northern hemisphere; RetII, Scl and SgrII in the Southern hemisphere - for which different DM density models (either cored or cuspy) lead to similar expectations, at variance with what happens for other DM targets - thus resulting in more robust predictions. We find that CTAO will provide the strongest limits above ~10 TeV, down to values of velocity-averaged annihilation cross section of ~5$\times 10^{-25}$ cm$^3$ s$^{-1}$ and up to decay lifetimes of ~10$^{26}$ s for combined limits on the best targets. We argue that the largest source of inaccuracy is due to the still imprecise determination of the DM content, especially for ultra-faint dSphs. We propose possible strategies of observation for CTAO, either optimized on a deep focus on the best known candidates, or on the diversification of targets.

Habitats integrate the abiotic conditions and biophysical structures that support biodiversity and sustain nature's contributions to people. As these ecosystems face mounting pressure from human activities, accurate, high-resolution habitat maps are essential for effective conservation and restoration. Yet current maps often fall short in thematic or spatial resolution because they must (1) model several mutually exclusive habitat types that co-occur across landscapes and (2) cope with severe class imbalance that complicate multi-class training. Here, we evaluated how high-resolution remote sensing (RS) data and Artificial Intelligence (AI) tools can improve habitat classification over large geographic extents at fine thematic resolution. Using vegetation plots from the European Vegetation Archive, we modelled Level 3 EUNIS habitats across Europe and assessed multiple modelling strategies against independent validation datasets. Strategies that exploited the hierarchical nature of habitat nomenclatures resolved classification ambiguities, especially in fragmented landscapes. Integrating multi-spectral (MSI) and synthetic aperture radar (SAR) imagery, particularly through Earth Observation Foundation models, enhanced within-formation discrimination and overall performance. Finally, ensemble machine learning that corrects class imbalance boosted accuracy further. Our methodological framework is transferable beyond Europe and adaptable to other classification systems. Future research should advance temporal modelling of dynamic habitats, extend to habitat segmentation and quality assessment, and exploit next-generation EO data paired with higher-quality in-situ observations.

Constraint-based methods and noise-based methods are two distinct families of methods proposed for uncovering causal graphs from observational data. However, both operate under strong assumptions that may be challenging to validate or could be violated in real-world scenarios. In response to these challenges, there is a growing interest in hybrid methods that amalgamate principles from both methods, showing robustness to assumption violations. This paper introduces a novel comprehensive framework for hybridizing constraint-based and noise-based methods designed to uncover causal graphs from observational time series. The framework is structured into two classes. The first class employs a noise-based strategy to identify a super graph, containing the true graph, followed by a constraint-based strategy to eliminate unnecessary edges. In the second class, a constraint-based strategy is applied to identify a skeleton, which is then oriented using a noise-based strategy. The paper provides theoretical guarantees for each class under the condition that all assumptions are satisfied, and it outlines some properties when assumptions are violated. To validate the efficacy of the framework, two algorithms from each class are experimentally tested on simulated data, realistic ecological data, and real datasets sourced from diverse applications. Notably, two novel datasets related to Information Technology monitoring are introduced within the set of considered real datasets. The experimental results underscore the robustness and effectiveness of the hybrid approaches across a broad spectrum of datasets.

The LHCb Upgrade II will fully exploit the flavour-physics opportunities of the HL-LHC, and study additional physics topics that take advantage of the forward acceptance of the LHCb spectrometer. The LHCb Upgrade I will begin operation in 2020. Consolidation will occur, and modest enhancements of the Upgrade I detector will be installed, in Long Shutdown 3 of the LHC (2025) and these are discussed here. The main Upgrade II detector will be installed in long shutdown 4 of the LHC (2030) and will build on the strengths of the current LHCb experiment and the Upgrade I. It will operate at a luminosity up to $2 \times 10^{34} \rm cm^{-2}s^{-1}$, ten times that of the Upgrade I detector. New detector components will improve the intrinsic performance of the experiment in certain key areas. An Expression Of Interest proposing Upgrade II was submitted in February 2017. The physics case for the Upgrade II is presented here in more depth. $CP$-violating phases will be measured with precisions unattainable at any other envisaged facility. The experiment will probe $b\to s \ell^+\ell^-$ and $b\to d \ell^+\ell^-$ transitions in both muon and electron decays in modes not accessible at Upgrade I. Minimal flavour violation will be tested with a precision measurement of the ratio of $B(B^0\to\mu^+\mu^-)/B(B_s^0\to \mu^+\mu^-)$. Probing charm $CP$ violation at the $10^{-5}$ level may result in its long sought discovery. Major advances in hadron spectroscopy will be possible, which will be powerful probes of low energy QCD. Upgrade II potentially will have the highest sensitivity of all the LHC experiments on the Higgs to charm-quark couplings. Generically, the new physics mass scale probed, for fixed couplings, will almost double compared with the pre-HL-LHC era; this extended reach for flavour physics is similar to that which would be achieved by the HE-LHC proposal for the energy frontier.

The Cherenkov Telescope Array Observatory (CTAO) is the next generation of ground-based observatories employing the imaging air Cherenkov technique for the study of very high energy gamma rays. The software Gammalearn proposes to apply Deep Learning as a part of the CTAO data analysis to reconstruct event parameters directly from images captured by the telescopes with minimal pre-processing to maximize the information conserved. In CTAO, the data analysis will include a data volume reduction that will definitely remove pixels. This step is necessary for data transfer and storage but could also involve information loss that could be used by sensitive algorithms such as neural networks (NN). In this work, we evaluate the performance of the gamma-PhysNet when applying different cleaning masks on images from Monte-Carlo simulations from the first Large-Sized Telescope. This study is critical to assess the impact of pixel removal in the data processing, mainly motivated by data compression.

We describe a search for gravitational waves from compact binaries with at least one component with mass 0.2 $M_\odot$ -- $1.0 M_\odot$ and mass ratio $q \geq 0.1$ in Advanced LIGO and Advanced Virgo data collected between 1 November 2019, 15:00 UTC and 27 March 2020, 17:00 UTC. No signals were detected. The most significant candidate has a false alarm rate of 0.2 $\mathrm{yr}^{-1}$. We estimate the sensitivity of our search over the entirety of Advanced LIGO's and Advanced Virgo's third observing run, and present the most stringent limits to date on the merger rate of binary black holes with at least one subsolar-mass component. We use the upper limits to constrain two fiducial scenarios that could produce subsolar-mass black holes: primordial black holes (PBH) and a model of dissipative dark matter. The PBH model uses recent prescriptions for the merger rate of PBH binaries that include a rate suppression factor to effectively account for PBH early binary disruptions. If the PBHs are monochromatically distributed, we can exclude a dark matter fraction in PBHs $f_\mathrm{PBH} \gtrsim 0.6$ (at 90% confidence) in the probed subsolar-mass range. However, if we allow for broad PBH mass distributions we are unable to rule out $f_\mathrm{PBH} = 1$. For the dissipative model, where the dark matter has chemistry that allows a small fraction to cool and collapse into black holes, we find an upper bound f_{\mathrm{DBH}} < 10^{-5} on the fraction of atomic dark matter collapsed into black holes.

Time Series Foundation Models (TSFMs) have shown promising zero-shot generalization across diverse forecasting tasks. However, their robustness to continual adaptation remains underexplored. In this work, we investigate the extent to which TSFMs suffer from catastrophic forgetting when fine-tuned sequentially on multiple datasets. Using synthetic datasets designed with varying degrees of periodic structure, we measure the trade-off between adaptation to new data and retention of prior knowledge. Our experiments reveal that, while fine-tuning improves performance on new tasks, it often causes significant degradation on previously learned ones, illustrating a fundamental stability-plasticity dilemma.

The Cherenkov Telescope Array Observatory (CTAO) is the next generation of observatories employing the imaging air Cherenkov technique for the study of very high energy gamma rays. The deployment of deep learning methods for the reconstruction of physical attributes of incident particles has evinced promising outcomes when conducted on simulations. However, the transition of this approach to observational data is accompanied by challenges, as deep learning-based models are susceptible to domain shifts. In this paper, we integrate domain adaptation in the physics-based context of the CTAO and shed light on the gain in performance that these techniques bring using LST-1 real acquisitions.

Seismology has entered the petabyte era, driven by decades of continuous recordings of broadband networks, the increase in nodal seismic experiments, and the recent emergence of Distributed Acoustic Sensing (DAS). This review explains how commercial clouds - AWS, Google Cloud, and Azure - by providing object storage, elastic compute, and managed databases, enable researchers to "bring the code to the data," thereby overcoming traditional HPC solutions' bandwidth and capacity limitations. After literature reviews of cloud concepts and their research applications in seismology, we illustrate the capacities of cloud-native workflows using two canonical end-to-end demonstrations: 1) ambient noise seismology and cross-correlation, and 2) earthquake detection, discrimination, and phase picking. Both workflows utilized S3 for streaming I/O and DocumentDB for provenance, demonstrating that cloud throughput can rival on-premises HPC at comparable costs, scanning 100 TBs to 1.3 PBs of seismic data in a few hours or days of processing. The review also discusses research and education initiatives, the reproducibility benefits of containers, and cost pitfalls (e.g., egress, I/O fees) of energy-intensive seismological research computing. While designing cloud pipelines remains non-trivial, partnerships with research software engineers enable converting domain code into scalable, automated, and environmentally conscious solutions for next-generation seismology.