Denoising Medium Resolution Stellar Spectra with U-Net Convolutional Neural Networks

Fisher-matrix forecasts are presented for the cosmological surveys of the Javalambre Physics of the Accelerating Universe Astrophysical Survey (J-PAS) and the Subaru Prime Focus Spectrograph (PFS). The wide, low-redshift coverage of J-PAS and the high-density, high-redshift mapping of PFS are strongly complementary: combining the two reduces marginalized uncertainties on all primary parameters compared with either survey individually. Adding the joint J-PAS+PFS data to next-generation cosmic microwave background (CMB) measurements from the Simons Observatory and LiteBIRD yields an expected precision of σ (∑ m ν ) = 0.017 eV in the ΛCDM+∑ m ν + N eff framework, sufficient to disfavor the inverted neutrino hierarchy at 2.34 σ if the true mass sum equals the normal-ordering minimum. Motivated by recent DESI results, we also forecast within a w 0 w a CDM+∑ m ν + N eff cosmology, adopting the DESI DR2 best-fit values ( w 0 = −0.758, w a = −0.82) as fiducial. The combination CMB+J-PAS+PFS then delivers σ ( w 0 ) = 0.044 and σ ( w a ) = 0.18, corresponding to a 5.1 σ preference for a time-varying dark-energy equation of state. These findings show that J-PAS and PFS, especially when coupled with Stage IV CMB observations, will provide competitive tests of neutrino physics and the dynamics of cosmic acceleration.

This short article is about Prime Focus Spectrograph (PFS), a very wide-field, massively-multiplexed, and optical & near-infrared (NIR) spectrograph as a next generation facility instrument on Subaru Telescope. More details and updates are available on the PFS official website (http://pfs.ipmu.jp), blog (http://pfs.ipmu.jp/blog/), and references therein.The project, instrument, & timelinePFS will position 2400 fibers to science targets or blank sky in the 1.3 degree field on the Subaru prime focus. These fibers will be quickly (~60sec) reconfigurable and feed the photons during exposures to the Spectrograph System (SpS). SpS consists of 4 modules each of which accommodate ~600 fibers and deliver spectral images ranging from 380nm to 1260nm simultaneously at one exposure via the 3 arms of blue, red, and NIR cameras. The instrument development has been undertaken by the international collaboration at the initiative of Kavli IPMU. The project is now going into the construction phase aiming at system integration and on-sky engineering observations in 2017-2018, and science operation in 2019.The survey design has also been under development envisioning a survey spanning ~300 nights over ~5 years in the framework of Subaru Strategic Program (SSP). The key science areas are: Cosmology, galaxy/AGN evolution, and Galactic Archaeology (GA) (Takada et al. 2014). The cosmology program will be to constrain the nature of dark energy via a survey of emission line galaxies over a comoving volume of 10 Gpc3 at z=0.8-2.4. In the galaxy/AGN program, the wide wavelength coverage of PFS as well as the large field of view will be exploited to characterize the galaxy populations and its clustering properties over a wide redshift range. A survey of color-selected galaxies/AGN at z = 1-2 will be conducted over 20 square degrees yielding a fair sample of galaxies with stellar masses down to ~1010M⊙. In the GA program, radial velocities and chemical abundances of stars in the Milky Way, dwarf spheroids, and M31 will be used to understand the past assembly histories of those galaxies and the structures of their dark matter halos. Spectra will be taken for 1 million stars as faint as V = 22 mag therefore out to large distances from the Sun.PFS will provide powerful spectroscopic capabilities even in the era of Euclid, LSST, WFIRST and TMT, and the effective synergies are expected for further unique science outputs.

We describe the infrastructure being developed to align and characterize the detectors for the Subaru Measurement of Images and Redshifts (SuMIRe) Prime Focus Spectrograph (PFS). PFS will employ four three-channel spectrographs with an operating wavelength range of 3800 $\AA$ to 12600 $\AA$. Each spectrograph will be comprised of two visible channels and one near infrared (NIR) channel, where each channel will use a separate Schmidt camera to image the captured spectra onto their respective detectors. In the visible channels, Hamamatsu 2k x 4k CCDs will be mounted in pairs to create a single 4k x 4k detector, while the NIR channel will use a single Teledyne 4k x 4k H4RG HgCdTe device.

Accurate prediction of the freezing level (FZL) is essential for hydrometeorological forecasting systems, with direct implications for runoff generation and reservoir management. In this study, we develop a deep learning–based postprocessing framework using the Unet convolutional neural network (CNN) architecture to refine the FZL forecasts from the West Weather Research and Forecasting (West-WRF) model. The proposed framework leverages reforecast data from West-WRF and FZL estimates from the California-Nevada River Forecast Center (CNRFC) to develop Unet models over the northern Sierra Nevada watersheds, such as the hydrologically critical Yuba-Feather watershed. We introduce two Unet model variants, Unet-log and Unet-GMM, that employ specialized loss functions beyond the standard benchmarks to enhance forecast skill. Unet-log utilizes the Log-Cosh cosine of Error, and Unet-GMM uses Gaussian Mixture Model loss functions, to enhance FZL forecasts. Results show that Unet-based postprocessing reduces centered root-mean-squared errors by up to 20% and increases forecast–observation correlation by about 10% compared to raw West-WRF. Evaluation using the continuous ranked probability score (CRPS) for Unet-GMM further demonstrates consistent improvements across lead times. While performance fluctuates with forecast horizon, storm variability, and diurnal forcing, Unet-GMM and Unet-log consistently outperform the baseline. The models capture the spatiotemporal variability of the FZL across different elevations, mitigating biases from the West-WRF model. This novel deep learning–based postprocessing approach demonstrates a promising pathway for integrating machine learning into hydrometeorological forecasting and decision support within the Forecast Informed Reservoir Operations (FIRO) framework.

In the context of upcoming sixth-generation (6G) wireless communication systems, the use of millimeter wave (mmWave) frequencies is a key technology for achieving high-throughput communications. Accurate parametric estimation of mmWave channels is critical for effective beamforming design and configuration, requiring sophisticated models to capture the directional characteristics of these channels. This work considers an innovative artificial intelligence (AI) approach for accurate estimation of angle-of-arrival (AoA) and angle-of-departure (AoD) parameters from frequency-domain channel observations. Our approach is based on the implementation of two convolutional neural networks (CNNs): a residual CNN (ResNet) and a U-Net CNN. Specifically, this work focuses on the efficient implementation of both schemes in an embedded system suitable for edge AI. We performed the experiments in a low-power NVIDIA Jetson Orin Nano platform and evaluated the effect of modifying the frequencies of its CPU and GPU on the performance of the inference process, both in terms of execution time and energy consumption. Experimental results showed that the U-Net model is more power consuming, but as it is faster, it consumes less energy per channel.

We present details of the design for the CCD readout electronics for the Subaru Telescope Prime Focus Spectrograph (PFS). The spectrograph is comprised of four identical spectrograph modules, each collecting roughly 600 spectra. The spectrograph modules provide simultaneous wavelength coverage over the entire band from 380 nm to 1260 nm through the use of three separate optical channels: blue, red, and near infrared (NIR). A camera in each channel images the multi-object spectra onto a 4k x 4k, 15 um pixel, detector format. The two visible cameras use a pair of Hamamatsu 2k x 4k CCDs with readout provided by custom electronics, while the NIR camera uses a single Teledyne HgCdTe 4k x 4k detector and ASIC Sidecar to read the device. The CCD readout system is a custom design comprised of three electrical subsystems: the Back End Electronics (BEE), the Front End Electronics (FEE), and a Pre-amplifier. The BEE is an off-the-shelf PC104 computer, with an auxiliary Xilinx FPGA module. The computer serves as the main interface to the Subaru messaging hub and controls other peripheral devices associated with the camera, while the FPGA is used to generate the necessary clocks and transfer image data from the CCDs. The FEE board sets clock biases, substrate bias, and CDS offsets. It also monitors bias voltages, offset voltages, power rail voltage, substrate voltage and CCD temperature. The board translates LVDS clock signals to biased clocks and returns digitized analog data via LVDS. Monitoring and control messages are sent from the BEE to the FEE using a standard serial interface. The Pre-amplifier board resides behind the detectors and acts as an interface to the two Hamamatsu CCDs. The Pre-amplifier passes clocks and biases to the CCDs, and analog CCD data is buffered and amplified prior to being returned to the FEE.

Condition level deteriorations modeling of RC beam bridges with U-Net convolutional neural networks

<p dir="ltr">Visuospatial processing is a fundamental cognitive capacity that underlies functions ranging from eye movements to working memory and even mathematical reasoning. Training visuospatial working memory can enhance performance across many cognitive domains — a phenomenon known as cognitive transfer. Although this effect is well established and holds therapeutic promise, the mechanisms driving transfer remain poorly understood. Measuring transfer is also challenging, as improvements during training may arise from multiple intertwined cognitive processes, as well as task-specific learning that does not generalise.</p><p dir="ltr">The aim of this thesis was to understand how and why visuospatial working memory training produces cognitive transfer. To this end, we developed methods for quantifying transferable improvements and related these behavioural changes to underlying neural dynamics.</p><p dir="ltr">To address <i>how</i>, we first combined a large behavioural dataset from an online cognitive training platform with dense longitudinal Magnetoencephalography (MEG) recordings in four participants. In Study 1, behavioural analyses revealed that task-specific gains appeared early in training, whereas transferable improvements accumulated gradually and consistently over time. In Study 2, MEG data showed that transfer was accompanied by a progressive, linear increase in alpha (~ 10 Hz) synchronisation within the dorsal attention network — a spatially selective network central to top-down control.</p><p dir="ltr">To address <i>why</i>, we analysed the longitudinal MEG cohort alongside data from 83 participants in the Human Connectome Project (HCP). In Study 3, we found that dorsal alpha oscillations were tightly linked to the maintenance of visuospatial information. Using a whole-brain computational model in which low-frequency oscillations modulated high-frequency neural activity, we specifically demonstrated that alpha synchronisation could facilitate information routing in the dorsal attention network.</p><p dir="ltr">Interestingly, the dorsal alpha network was active during the maintenance of non-spatial information as well. This suggests that the network supports a more general neural state of cognitive stability. Such mechanisms offer a compelling explanation for how training-induced neural plasticity can extend beyond visuospatial abilities, leading to broader improvements in attention and general cognitive function — even over relatively short training periods.</p><h3>List of scientific papers</h3><p dir="ltr">I. <b>Ericson, J.</b>, & Klingberg, T. (2023). A dual-process model for cognitive training. Npj science of learning, 8(1), 12. <a href="https://doi.org/10.1038/s41539-023-00161-2" rel="noreferrer" target="_blank">https://doi.org/10.1038/s41539-023-00161-2</a></p><p dir="ltr">II. <b>Ericson, J.</b>, Palva, S., Palva, M., & Klingberg, T. (2024). Strengthening of alpha synchronization is a neural correlate of cognitive transfer. Cerebral Cortex, 34(2), bhad527. <a href="https://doi.org/10.1093/cercor/bhad527" rel="noreferrer" target="_blank">https://doi.org/10.1093/cercor/bhad527</a></p><p dir="ltr">III. <b>Ericson, J.</b>, Ruiz Ibáñez, N., Lundqvist, M., & Klingberg, T. (2025). Low frequency oscillations - neural correlates of stability and flexibility in cognition. Nature Communications, 16(1), 5381. <a href="https://doi.org/10.1038/s41467-025-60821-2" rel="noreferrer" target="_blank">https://doi.org/10.1038/s41467-025-60821-2</a></p>

Bone material contains a hierarchical network of micro- and nano-cavities and channels, known as the lacuna-canalicular network (LCN), that is thought to play an important role in mechanobiology and turnover. The LCN comprises micrometer-sized lacunae, voids that house osteocytes, and submicrometer-sized canaliculi that connect bone cells. Characterization of this network in three dimensions is crucial for many bone studies. To quantify X-ray Zernike phase-contrast nanotomography data, deep learning is used to isolate and assess porosity in artifact-laden tomographies of zebrafish bones. A technical solution is proposed to overcome the halo and shade-off domains in order to reliably obtain the distribution and morphology of the LCN in the tomographic data. Convolutional neural network (CNN) models are utilized with increasing numbers of images, repeatedly validated by `error loss' and `accuracy' metrics. U-Net and Sensor3D CNN models were trained on data obtained from two different synchrotron Zernike phase-contrast transmission X-ray microscopes, the ANATOMIX beamline at SOLEIL (Paris, France) and the P05 beamline at PETRA III (Hamburg, Germany). The Sensor3D CNN model with a smaller batch size of 32 and a training data size of 70 images showed the best performance (accuracy 0.983 and error loss 0.032). The analysis procedures, validated by comparison with human-identified ground-truth images, correctly identified the voids within the bone matrix. This proposed approach may have further application to classify structures in volumetric images that contain non-linear artifacts that degrade image quality and hinder feature identification.

Recognizing objects from subcategories with very subtle differences remains a challenging task due to the large intra-class and small inter-class variation. Recent work tackles this problem in a weakly-supervised manner: object parts are first detected and the corresponding part-specific features are extracted for fine-grained classification. However, these methods typically treat the part-specific features of each image in isolation while neglecting their relationships between different images. In this paper, we propose Cross-X learning, a simple yet effective approach that exploits the relationships between different images and between different network layers for robust multi-scale feature learning. Our approach involves two novel components: (i) a cross-category cross-semantic regularizer that guides the extracted features to represent semantic parts and, (ii) a cross-layer regularizer that improves the robustness of multi-scale features by matching the prediction distribution across multiple layers. Our approach can be easily trained end-to-end and is scalable to large datasets like NABirds. We empirically analyze the contributions of different components of our approach and demonstrate its robustness, effectiveness and state-of-the-art performance on five benchmark datasets. Code is available at \url{https://github.com/cswluo/CrossX}.

Abstract. Aerosol effects on cloud properties and the atmospheric energy and radiation budgets are studied through ensemble simulations over two month-long periods during the NARVAL campaigns (Next-generation Aircraft Remote-Sensing for Validation Studies, December 2013 and August 2016). For each day, two simulations are conducted with low and high cloud droplet number concentrations (CDNCs), representing low and high aerosol concentrations, respectively. This large data set, which is based on a large spread of co-varying realistic initial conditions, enables robust identification of the effect of CDNC changes on cloud properties. We show that increases in CDNC drive a reduction in the top-of-atmosphere (TOA) net shortwave flux (more reflection) and a decrease in the lower-tropospheric stability for all cases examined, while the TOA longwave flux and the liquid and ice water path changes are generally positive. However, changes in cloud fraction or precipitation, that could appear significant for a given day, are not as robustly affected, and, at least for the summer month, are not statistically distinguishable from zero. These results highlight the need for using a large sample of initial conditions for cloud–aerosol studies for identifying the significance of the response. In addition, we demonstrate the dependence of the aerosol effects on the season, as it is shown that the TOA net radiative effect is doubled during the winter month as compared to the summer month. By separating the simulations into different dominant cloud regimes, we show that the difference between the different months emerges due to the compensation of the longwave effect induced by an increase in ice content as compared to the shortwave effect of the liquid clouds. The CDNC effect on the longwave flux is stronger in the summer as the clouds are deeper and the atmosphere is more unstable.

The vertebral levels of the spine provide a useful coordinate system when making measurements of plaque, muscle, fat, and bone mineral density. Correctly classifying vertebral levels with high accuracy is challenging due to the similar appearance of each vertebra, the curvature of the spine, and the possibility of anomalies such as fractured vertebrae, implants, lumbarization of the sacrum, and sacralization of L5. The goal of this work is to develop a system that can accurately and robustly identify the L1 level in large heterogeneous datasets. The first approach we study is using a 3D U-Net to segment the L1 vertebra directly using the entire scan volume to provide context. We also tested models for two class segmentation of L1 and T12 and a three class segmentation of L1, T12 and the rib attached to T12. By increasing the number of training examples to 249 scans using pseudo-segmentations from an in-house segmentation tool we were able to achieve 98% accuracy with respect to identifying the L1 vertebra, with an average error of 4.5 mm in the craniocaudal level. We next developed an algorithm which performs iterative instance segmentation and classification of the entire spine with a 3D U-Net. We found the instance based approach was able to yield better segmentations of nearly the entire spine, but had lower classification accuracy for L1.

Vision measuring, involving the measurement of the three-dimensional shape and color of large objects using stereo vision-based methods, is a challenging task. In this study, a freely adjustable vision measurement system for large objects is developed by connecting a camera to a total station telescope via a fixed mount. In this system, a camera is focused on the target object to be measured in far-range scenarios. A high-precision calibration method is proposed based on non-parametric camera models. In addition, a point cloud-based three-dimensional reconstruction method is developed by integrating with the non-parametric camera models. The experimental results showed that the proposed system has high precision and robustness. When measuring objects from a distance of 8m in an indoor scenario, the error in the maximum spatial distance measurement between two corner points on the calibration board is less than 0.24mm. The maximum relative measurement error is less than 0.09%. When measuring an object 20m away in an outdoor scenario, the average error in the spatial distance measurement between two arbitrary points in an 11 m×4 m testing region is 5.6mm. The average relative error in the distance measurement is 0.12%.

Most Chinese archaeologists assume that the scapular implements used in the Hemudu culture in eastern China (7000–5000 BP) were the si agricultural implements (tools for breaking ground and turning soils over to assist in seeding) recorded in ancient Chinese literatures and, accordingly, assume the Hemudu culture was a farming society. However, ethnographic and historical literatures worldwide have suggested inconclusive functions for similar implements. We conducted a range of experiments under realistic conditions, including hide and plant processing and earth-working, followed by use-wear analysis, to identify the functions of the Hemudu scapular implements. The results suggest that no more than half of the implements were employed as si and that their penetrability and durability were rather limited. These findings help explain why Hemudu should not be labeled as a farming society. Through experimentation and use-wear analysis, we produced relatively large datasets that make a significant contribution to the identification of soil-derived wear patterns on bone tools. We also included quantitative measurements of soil properties to ensure similarities in use contexts between our experimental and archaeological analogies in order to reach reliable functional identifications. Our approaches and results, therefore, provided a solid base for re-evaluating previous research as well as building a standardized database of scientific value for future evaluation and adjustment, even if that future research is done in isolation and in different soil contexts.

We investigate how cosmic web structures affect galaxy quenching in the IllustrisTNG (TNG100) cosmological simulations by reconstructing the cosmic web within each snapshot using the DisPerSE framework. We measure the comoving distance from each galaxy with stellar mass to the nearest node (d node) and the nearest filament spine (d fil) to study the dependence of both the median specific star formation rate (〈sSFR〉) and the median gas fraction (〈f gas〉) on these distances. We find that the 〈sSFR〉 of galaxies is only dependent on the cosmic web environment at z < 2, with the dependence increasing with time. At z ≤ 0.5, galaxies are quenched at d node ≲ 1 Mpc, and have significantly suppressed star formation at d fil ≲ 1 Mpc, trends driven mostly by satellite galaxies. At z ≤ 1, in contrast to the monotonic drop in 〈sSFR〉 of galaxies with decreasing d node and d fil, galaxies—both centrals and satellites—experience an upturn in 〈sSFR〉 at d node ≲ 0.2 Mpc. Much of this cosmic web dependence of star formation activity can be explained by an evolution in 〈f gas〉. Our results suggest that in the past ∼10 Gyr, low-mass satellites are quenched by rapid gas stripping in dense environments near nodes and gradual gas starvation in intermediate-density environments near filaments. At earlier times, cosmic web structures efficiently channeled cold gas into most galaxies. State-of-the-art ongoing spectroscopic surveys such as the Sloan Digital Sky Survey and DESI, as well as those planned with the Subaru Prime Focus Spectrograph, JWST, and Roman, are required to test our predictions against observations.