Maximilian Alber, Timo Milbich, Alexandra Carpen-Amarie
et al.
Pathology foundation models substantially advanced the possibilities in computational pathology -- yet tradeoffs in terms of performance, robustness, and computational requirements remained, which limited their clinical deployment. In this report, we present Atlas 2, Atlas 2-B, and Atlas 2-S, three pathology vision foundation models which bridge these shortcomings by showing state-of-the-art performance in prediction performance, robustness, and resource efficiency in a comprehensive evaluation across eighty public benchmarks. Our models were trained on the largest pathology foundation model dataset to date comprising 5.5 million histopathology whole slide images, collected from three medical institutions Charité - Universtätsmedizin Berlin, LMU Munich, and Mayo Clinic.
Foundation models learn transferable representations, motivating growing interest in their application to wireless systems. Existing wireless foundation models are predominantly based on transformer architectures, whose quadratic computational and memory complexity can hinder practical deployment for large-scale channels. In this work, we introduce WiMamba, a wireless foundation model built upon the recently proposed Mamba architecture, which replaces attention mechanisms with selective state-space models and enables linear-time sequence modeling. Leveraging this architectural advantage combined with adaptive preprocessing, WiMamba achieves scalable and low-latency inference while maintaining strong representational expressivity. We further develop a dedicated task-agnostic, self-supervised pre-training framework tailored to wireless channels, resulting in a genuine foundation model that learns transferable channel representations. Evaluations across four downstream tasks demonstrate that WiMamba matches or outperforms transformer-based wireless foundation models, while offering dramatic latency and memory reductions.
Audun L. Henriksen, Ole-Johan Skrede, Lisa van der Schee
et al.
Foundation models in histopathology are expected to facilitate the development of high-performing and generalisable deep learning systems. However, current models capture not only biologically relevant features, but also pre-analytic and scanner-specific variation that bias the predictions of task-specific models trained from the foundation model features. Here we show that introducing novel robustness losses during training of downstream task-specific models reduces sensitivity to technical variability. A purpose-designed comprehensive experimentation setup with 27,042 WSIs from 6155 patients is used to train thousands of models from the features of eight popular foundation models for computational pathology. In addition to a substantial improvement in robustness, we observe that prediction accuracy improves by focusing on biologically relevant features. Our approach successfully mitigates robustness issues of foundation models for computational pathology without retraining the foundation models themselves, enabling development of robust computational pathology models applicable to real-world data in routine clinical practice.
Carrollian manifolds offer an intrinsic geometric framework for the physics in the ultra-relativistic limit. The recently introduced Carrollian Lie algebroids are generalised to the setting of $ρ$-commutative geometry, (also known as almost commutative geometry), where the underlying algebras commute up to a numerical factor. Via $ρ$-Lie-Rinehart pairs, it is shown that the foundational tenets of Carrollian geometry have analogous statements in the almost commutative world. We explicitly build two toy examples: we equip the extended quantum plane and the noncommutative $2$-torus with Carrollian structures. This opens up the rigorous study of noncommutative Carrollian geometry via almost commutative geometry.
We revisit the classical, full-fledged Bayesian model averaging (BMA) paradigm to ensemble pre-trained and/or lightly-finetuned foundation models to enhance the classification performance on image and text data. To make BMA tractable under foundation models, we introduce trainable linear classifiers that take frozen features from the pre-trained foundation models as inputs. The model posteriors over the linear classifiers tell us which linear heads and frozen features are better suited for a given dataset, resulting in a principled model ensembling method. Furthermore, we propose a computationally cheaper, optimizable model averaging scheme (OMA). In OMA, we directly optimize the model ensemble weights, just like those weights based on model posterior distributions in BMA, by reducing the amount of surprise (expected entropy of the predictions) we get from predictions of ensembled models. With the rapid development of foundation models, these approaches will enable the incorporation of future, possibly significantly better foundation models to enhance the performance of challenging classification tasks.
Recent advances in clinical AI have enabled remarkable progress across many clinical domains. However, existing benchmarks and models are primarily limited to a small set of modalities and tasks, which hinders the development of large-scale multimodal methods that can make holistic assessments of patient health and well-being. To bridge this gap, we introduce Clinical Large-Scale Integrative Multimodal Benchmark (CLIMB), a comprehensive clinical benchmark unifying diverse clinical data across imaging, language, temporal, and graph modalities. CLIMB comprises 4.51 million patient samples totaling 19.01 terabytes distributed across 2D imaging, 3D video, time series, graphs, and multimodal data. Through extensive empirical evaluation, we demonstrate that multitask pretraining significantly improves performance on understudied domains, achieving up to 29% improvement in ultrasound and 23% in ECG analysis over single-task learning. Pretraining on CLIMB also effectively improves models' generalization capability to new tasks, and strong unimodal encoder performance translates well to multimodal performance when paired with task-appropriate fusion strategies. Our findings provide a foundation for new architecture designs and pretraining strategies to advance clinical AI research. Code is released at https://github.com/DDVD233/climb.
This monograph presents a comprehensive analysis of the technological foundations of management decision-making in the reconstruction of complex gas pipeline systems. The study addresses the challenges posed by the aging infrastructure of gas supply networks and explores advanced strategies to improve their reliability, efficiency, and automation. Particular attention is given to the reconstruction of pipelines with various configurations linear, looped, and parallel systems under non-stationary gas flow conditions. The proposed models and methodologies offer solutions for optimizing operational parameters, improving emergency valve response, and ensuring uninterrupted gas supply through advanced management systems and data-driven decision support tools. Emphasis is placed on the integration of modern technologies, system theory, and feedback mechanisms in the design and operation of reconstructed pipeline systems. This work is intended for engineers, system designers, and researchers in the fields of gas supply, systems engineering, and energy infrastructure.
Alkis Koudounas, Moreno La Quatra, Marco Sabato Siniscalchi
et al.
Speech foundation models have demonstrated exceptional capabilities in speech-related tasks. Nevertheless, these models often struggle with non-verbal audio data, such as vocalizations, baby crying, etc., which are critical for various real-world applications. Audio foundation models well handle non-speech data but also fail to capture the nuanced features of non-verbal human sounds. In this work, we aim to overcome the above shortcoming and propose a novel foundation model, termed voc2vec, specifically designed for non-verbal human data leveraging exclusively open-source non-verbal audio datasets. We employ a collection of 10 datasets covering around 125 hours of non-verbal audio. Experimental results prove that voc2vec is effective in non-verbal vocalization classification, and it outperforms conventional speech and audio foundation models. Moreover, voc2vec consistently outperforms strong baselines, namely OpenSmile and emotion2vec, on six different benchmark datasets. To the best of the authors' knowledge, voc2vec is the first universal representation model for vocalization tasks.
V. Michaud-Belleau, M. Gaudreau, J. Lacoursière
et al.
<p>The Atmospheric Sounder Spectrometer by Infrared Spectral Technology (ASSIST) is a Fourier-transform spectrometer designed, fabricated, and sold by LR Tech Inc., which operates in the thermal infrared. When attached to its automated radiometric calibration module, it functions as an infrared spectroradiometer (IRS) that passively measures the absolute spectral radiance within a 46 mrad full field of view and over the 525 to 3300 cm<span class="inline-formula"><sup>−1</sup></span> (3 to 19 <span class="inline-formula">µm</span>) spectral range. For atmospheric studies, the ASSIST IRS is integrated into a mobile enclosure enabling autonomous and reliable operation across a range of environmental conditions. It is typically configured for downwelling radiance measurements (zenith view) at 0.5 cm<span class="inline-formula"><sup>−1</sup></span> bin spacing, 0.6 cm<span class="inline-formula"><sup>−1</sup></span> resolution, and 4 min<span class="inline-formula"><sup>−1</sup></span> sampling rate, closely replicating the behavior of the Atmospheric Emitted Radiance Interferometer (AERI, in rapid-sampling mode), a similar but older IRS. Atmospheric variables affecting the shape of the downwelling thermal infrared radiance spectrum at ground level can be retrieved from the ASSIST high-resolution measurements using dedicated inversion algorithms. This includes the properties of some aerosols and simple clouds, the mixing ratios of trace gases, and the vertical distribution of temperature and water vapor (thermodynamic profile) in the lower troposphere above the instrument. Due to the form of the radiative transfer equation, thermodynamic profiles can only be retrieved with low to moderate vertical resolution but with sufficient accuracy and temporal resolution to help fill the current boundary layer observational gap. This paper provides a detailed description of the ASSIST design and near-real-time processing algorithm producing the calibrated radiance spectra that are useful in a variety of applications.</p>
<p>Accurate determination of the planetary boundary layer (PBL) height, mixing-layer height (MLH), and aerosol layer top (ALT) is essential for air quality and climate studies, particularly in regions with complex aerosol dynamics such as Chiang Mai, northern Thailand. This study introduces a novel lidar-based retrieval method that integrates a temperature-dependent, dynamic maximum analysis altitude (MAA) into the traditional Haar wavelet covariance transform (WCT) framework. Unlike conventional fixed-altitude WCT approaches, which often misclassify the ALT as the PBL – especially under stable nighttime or transitional conditions – this dynamic approach adapts the vertical search range for PBL detection in real time using observed surface temperature variations. The method is physically grounded in boundary layer thermodynamics, allowing for more accurate identification of the true PBL top while reducing contamination from residual aerosol layers and low clouds. Validation against radiosonde observations and comparison with previously validated WRF-Chem simulations demonstrate strong agreement, with the lidar-derived PBL heights capturing diurnal variations more reliably than traditional methods. The findings also reveal model biases during high-aerosol events, highlighting the need for improved aerosol–meteorology coupling in mesoscale models. This integrated retrieval framework represents a significant advancement in lidar-based boundary layer detection and offers a robust tool for enhancing pollutant dispersion analysis, air quality forecasting, and climate modeling across aerosol-rich regions in Southeast Asia.</p>
<p>Aerosols have significant effects on Earth, which vary according to the type of these atmospheric particles. Different observing systems exist today to monitor aerosols, mainly through the retrieval of aerosol optical depth (AOD), among which meteorological satellites in geostationary orbit provide unique information thanks to their acquisition of several Earth's images per hour. The third generation of European geostationary satellites, Meteosat Third Generation-Imager with the onboard Flexible Combined Imager (FCI) operational since December 2024, brings new possibilities for aerosol remote sensing compared to its predecessor, Meteosat Second Generation, with the Spinning Enhanced Visible and Infrared Imager (SEVIRI) on board. This article assesses the improvements in aerosol characterization that will be made possible thanks to FCI, based on realistically generated synthetic data that are processed by optimal estimation methods to quantify aerosol information content and to retrieve relevant aerosol properties. Two case studies corresponding to challenging aerosol retrieval situations are simulated, a dust outbreak in North Africa and the wildfire season in South West Africa. First, synthetic data are used to study the potential for AOD retrieval of new FCI spectral channels in comparison to SEVIRI's. Results prove that channel VIS04 (centered at 444 nm) is the best suited for this task, with a significant decrease in retrieval error (root square mean error by 23 % and mean bias error by 65 %) in comparison to AOD estimated from the SEVIRI-heritage channel VIS06 (centered at 640 nm). Second, the FCI capabilities to further characterize aerosol particles are investigated, with the joint retrieval of AOD and fine mode fraction (FMF), which is linked to particle size distribution and therefore aerosol type. This is achieved by exploiting near-infrared channel NIR22 (centered at 2250 nm, and found to be sensitive to coarse particles only in the first part of the study) in addition to channel VIS04. Experiments show that, except under certain unfavorable conditions, the joint retrieval of AOD and FMF is possible, even if fast but less accurate radiative transfer models are used, which could be employed in an operational setting. This article demonstrates the possibility of obtaining advanced high temporal frequency aerosol observations from FCI and opens pathways for the future study of aerosol diurnal variations from space.</p>
Estimating the out-of-distribution performance in regimes where labels are scarce is critical to safely deploy foundation models. Recently, it was shown that ensembles of neural networks observe the phenomena "agreement-on-the-line", which can be leveraged to reliably predict OOD performance without labels. However, in contrast to classical neural networks that are trained on in-distribution data from scratch for numerous epochs, foundation models undergo minimal finetuning from heavily pretrained weights, which may reduce the ensemble diversity needed to observe agreement-on-the-line. In our work, we demonstrate that when lightly finetuning multiple runs from a single foundation model, the choice of randomness during training (linear head initialization, data ordering, and data subsetting) can lead to drastically different levels of agreement-on-the-line in the resulting ensemble. Surprisingly, only random head initialization is able to reliably induce agreement-on-the-line in finetuned foundation models across vision and language benchmarks. Second, we demonstrate that ensembles of multiple foundation models pretrained on different datasets but finetuned on the same task can also show agreement-on-the-line. In total, by careful construction of a diverse ensemble, we can utilize agreement-on-the-line-based methods to predict the OOD performance of foundation models with high precision.
<p>Aerosol particles undergo continuous changes in their chemical composition and physical properties throughout their lifecycles, leading to diverse climate and health impacts. In particular, organic nanoparticle’s surface equilibrium vapor pressure stands as a critical factor for gas–particle partitioning and is pivotal for understanding the evolution of aerosol properties. Herein, we present measurements of evaporation kinetics and surface equilibrium vapor pressures of a wide array of laboratory-generated organic nanoparticles, employing the dynamic-aerosol-size electrical mobility spectrometer (DEMS) methodology, a recent advancement in aerosol process characterization. The DEMS methodology is founded on the principle that the local velocity of a size-changing nanoparticle within a flow field has a one-to-one correspondence with its local size. Consequently, this approach can facilitate the in situ probing of rapid aerosol size-changing processes by analyzing the trajectories of size-changing nanoparticles within the classification region of a differential mobility analyzer (DMA). We employ the DEMS with a tandem DMA setup, where a heated sheath flow in the second DMA initiates particle evaporation in its classification region. Through analysis of the DEMS response and the underlying mechanism governing the evaporation process, we reconstruct temporal radius profiles of evaporating nanoparticles and derive their surface equilibrium vapor pressures across various temperatures. Our results demonstrate a good agreement between the vapor pressures deduced from DEMS measurements and those documented in literature. We discuss the measurable vapor pressure range achievable with DEMS and elucidate associated uncertainties. Furthermore, we outline prospective directions for refining this methodology and anticipate its potential to contribute to the characterization of aerosol-related kinetic processes with currently unknown mechanisms.</p>
<p>To improve the performance of the calibration model for the air quality monitoring, a low-cost multi-parameter air quality monitoring system (LCS) based on different machine learning algorithms is proposed. The LCS can measure particulate matter (PM<span class="inline-formula"><sub>2.5</sub></span> and PM<span class="inline-formula"><sub>10</sub></span>) and gas pollutants (SO<span class="inline-formula"><sub>2</sub></span>, NO<span class="inline-formula"><sub>2</sub></span>, CO and O<span class="inline-formula"><sub>3</sub></span>) simultaneously. The multi-input multi-output (MIMO) prediction model is developed based on the original signals of the sensors, ambient temperature (<span class="inline-formula"><i>T</i></span>) and relative humidity (RH), and the measurements of the reference instrumentations. The performance of the different algorithms (RF, MLR, KNN, BP, GA–BP) with parameters such as determination coefficient <span class="inline-formula"><i>R</i><sup>2</sup></span>, root mean square error (RMSE), and mean absolute error (MAE) are compared and discussed. Using these methods, the <span class="inline-formula"><i>R</i><sup>2</sup></span> of the algorithms (RF, MLR, KNN, BP, GA–BP) for the PM is in the range 0.68–0.99; the RMSE values of PM<span class="inline-formula"><sub>2.5</sub></span> and PM<span class="inline-formula"><sub>10</sub></span> are within 2.36–18.68 and 4.55–45.05 <span class="inline-formula">µ</span>g m<span class="inline-formula"><sup>−3</sup></span>, respectively; the MAE values of PM<span class="inline-formula"><sub>2.5</sub></span> and PM<span class="inline-formula"><sub>10</sub></span> are within 1.44–12.80 and 3.21–23.20 <span class="inline-formula">µ</span>g m<span class="inline-formula"><sup>−3</sup></span>, respectively. The <span class="inline-formula"><i>R</i><sup>2</sup></span> of the algorithms (RF, MLR, KNN, BP, GA–BP) for the gas pollutants (O<span class="inline-formula"><sub>3</sub></span>, CO and NO<span class="inline-formula"><sub>2</sub></span>) is within 0.70–0.99; the RMSE values for these pollutants are 4.05–17.79 <span class="inline-formula">µ</span>g m<span class="inline-formula"><sup>−3</sup></span>, 0.02–0.18 mg m<span class="inline-formula"><sup>−3</sup></span>, 2.88–14.54 <span class="inline-formula">µ</span>g m<span class="inline-formula"><sup>−3</sup></span>, respectively; the MAE values for these pollutants are 2.76–13.46 <span class="inline-formula">µ</span>g m<span class="inline-formula"><sup>−3</sup></span>, 0.02–0.19 mg m<span class="inline-formula"><sup>−3</sup></span>, 1.84–11.08 <span class="inline-formula">µ</span>g m<span class="inline-formula"><sup>−3</sup></span>, respectively. The <span class="inline-formula"><i>R</i><sup>2</sup></span> of the algorithms (RF, KNN, BP, GA–BP, except for MLR) for SO<span class="inline-formula"><sub>2</sub></span> is within 0.27–0.97, the RMSE value is in the range 0.64–5.37 <span class="inline-formula">µ</span>g m<span class="inline-formula"><sup>−3</sup></span>, and the MAE value is in the range 0.39–4.24 <span class="inline-formula">µ</span>g m<span class="inline-formula"><sup>−3</sup></span>. These measurements are consistent with the national environmental protection standard requirement of China, and the LCS based on the machine learning algorithms can be used to predict the concentrations of PM and gas pollution.</p>
<p>The first closure study involving passive microwave and submillimetre measurements of ice clouds with the consideration of oriented particles is presented, using a unique combination of polarised observations from the ISMAR spectral-like radiometer, two radars with frequencies of 35 and 95 <span class="inline-formula">GHz</span>, and a variety of in situ instruments. Of particular interest to this study are the large V–H polarised brightness temperature differences measured from ISMAR above a thick frontal ice cloud. Previous studies combining radar and passive submillimetre measurements have not considered polarisation differences. Moreover, they have assumed particle habits a priori. We aim to test whether the large V–H measurements can be simulated successfully by using an atmospheric model consistent with in situ microphysics.</p>
<p>An atmospheric model is constructed using information from the in situ measurements, such as the ice water content, the particle size distribution, and the mass and shape of particles, as well as background information obtained from dropsonde profiles. Columnar and dendritic aggregate particle models are generated specifically for this case, and their scattering properties are calculated using the independent monomer approximation under the assumption of horizontal orientation. The scattering properties are used to perform polarised radiative transfer simulations using ARTS to test whether we can successfully simulate the measured large V–H differences. Radar measurements are used to extrapolate the 1-D microphysical profile to derive a time series of particle size distributions which are used to simulate ISMAR brightness temperatures. These simulations are compared to the observations.</p>
<p>It is found that particle models that are consistent with in situ microphysics observations are capable of reproducing the brightness temperature depression and polarisation signature measured from ISMAR at the dual-polarised channel of 243 <span class="inline-formula">GHz</span>. However, it was required that a proportion of the particles were changed in order to increase the V–H polarised brightness temperature differences. Thus, we incorporated millimetre-sized dendritic crystals, as these particles were observed in the probe imagery. At the second dual-polarised channel of 664 <span class="inline-formula">GHz</span>, the brightness temperature depressions were generally simulated at the correct locations; however, the simulated V–H was too large. This work shows that multi-frequency polarisation information could be used to infer realistic particle shapes, orientations, and representations of the split between single crystals and aggregates within the cloud.</p>
<p>We present a new method for estimating <span class="inline-formula">NO<sub><i>x</i></sub></span> emissions and effective lifetimes from large cities based on <span class="inline-formula">NO<sub>2</sub></span> measurements from the TROPOspheric Monitoring Instrument (TROPOMI) (PAL dataset, May 2018–November 2021). As in previous studies, the estimate is based on the downwind plume evolution for different wind directions separately. The novelty of the presented approach lies in the simultaneous fit of downwind patterns for opposing wind directions, which makes the method far more robust (i.e., less prone to local minima with nonphysically high or low lifetimes) than a single exponential decay fit. In addition, the new method does not require the assumption of a city being a “point source” but also derives the spatial distribution of emissions.</p>
<p>The method was successfully applied to 100 cities worldwide on a seasonal scale. Fitted emissions generally agree reasonably with the Emissions Database for Global Atmospheric Research (EDGAR) v6 (<span class="inline-formula"><i>R</i>=0.72</span>) and are on average 14 % lower, while estimated uncertainties are still rather large (<span class="inline-formula">≈</span> 30 %–50 %). Lifetimes were found to be rather short (2.44 <span class="inline-formula">±</span> 0.68 h) and show no distinct dependency on season or latitude, which might be a consequence of discarding observations at high solar zenith angles (<span class="inline-formula">>65</span>°).</p>
<p>The main limitations of this and similar methods are the underlying assumptions of steady state (meaning constant emissions, wind fields and chemical conditions) within about 100 km downwind from a city, which is probably a simplification that is too strong in order to reach higher accuracies.</p>
<p>Satellite-based estimations of dry-air column-averaged mixing ratios of methane (<span class="inline-formula">XCH<sub>4</sub></span>) contribute to a better understanding of changes in <span class="inline-formula">CH<sub>4</sub></span> emission sources and variations in its atmospheric growth rates. High accuracy of the satellite measurements is required, and therefore, extensive validation is performed, mainly against the Total Carbon Column Observing Network (TCCON). However, validation opportunities at open-ocean areas outside the coastal regions are sparse. We propose a new approach to assess the accuracy of satellite-derived <span class="inline-formula">XCH<sub>4</sub></span> trends and variations. We combine various ship and aircraft observations with the help of atmospheric chemistry models, mainly used for the stratospheric column, to derive observation-based <span class="inline-formula">XCH<sub>4</sub></span> (obs. <span class="inline-formula">XCH<sub>4</sub></span>). Based on our previously developed approach for the application to <span class="inline-formula">XCO<sub>2</sub></span>, we investigated three different advancements, from a simple approach to more elaborate approaches (approaches 1, 2, and 3), to account for the higher tropospheric and stratospheric variability in <span class="inline-formula">CH<sub>4</sub></span> as compared to <span class="inline-formula">CO<sub>2</sub></span>. Between 2014 and 2018, at 20–40° N of the western Pacific, we discuss the uncertainties in the approaches and the derived obs. <span class="inline-formula">XCH<sub>4</sub></span> within 10° by 20° latitude–longitude boxes. Uncertainties were 22 <span class="inline-formula">ppb</span> (parts per billion) for approach 1, 20 <span class="inline-formula">ppb</span> for approach 2, and 16 <span class="inline-formula">ppb</span> for approach 3. We analyzed the consistency with the nearest TCCON stations and found agreement of approach 3 with Saga of <span class="inline-formula">1±12</span> <span class="inline-formula">ppb</span> and <span class="inline-formula"><math xmlns="http://www.w3.org/1998/Math/MathML" id="M17" display="inline" overflow="scroll" dspmath="mathml"><mrow><mo>-</mo><mn mathvariant="normal">1</mn><mo>±</mo><mn mathvariant="normal">11</mn></mrow></math><span><svg:svg xmlns:svg="http://www.w3.org/2000/svg" width="39pt" height="10pt" class="svg-formula" dspmath="mathimg" md5hash="6a90032efed1f8b55483afe692e8b4b8"><svg:image xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="amt-17-1297-2024-ie00001.svg" width="39pt" height="10pt" src="amt-17-1297-2024-ie00001.png"/></svg:svg></span></span> <span class="inline-formula">ppb</span> with Tsukuba for the northern and southern latitude box, respectively. Furthermore, we discuss the impact of the modeled stratospheric column on the derived obs. <span class="inline-formula">XCH<sub>4</sub></span> by applying three different models in our approaches. Depending on the models, the difference can be more than 12 <span class="inline-formula">ppb</span> (0.6 %), showing the importance for the appropriate choice. We show that our obs. <span class="inline-formula">XCH<sub>4</sub></span> dataset accurately captures seasonal variations in <span class="inline-formula">CH<sub>4</sub></span> over the ocean. Using different retrievals of the Greenhouse Gases Observing Satellite (GOSAT) from the National Institute for Environmental Studies (NIES), the RemoTeC full-physics retrieval operated at the Netherlands Institute for Space Research (SRON), and the full-physics retrieval of the University of Leicester (UoL-OCFP), we demonstrate the applicability of the dataset for satellite evaluation. The comparison with results of approach 3 revealed that NIES showed a difference of <span class="inline-formula">−</span>0.04 <span class="inline-formula">±</span> 13 <span class="inline-formula">ppb</span> and strong scatter at 20–30° N, while RemoTeC and OCFP have a rather systematic negative bias of <span class="inline-formula">−</span>12.1 <span class="inline-formula">±</span> 8.1 and <span class="inline-formula">−</span>10.3 <span class="inline-formula">±</span> 9.6 <span class="inline-formula">ppb</span>. Our new approach to derive <span class="inline-formula">XCH<sub>4</sub></span> reference datasets over the ocean can contribute to the validation of existing and upcoming satellite missions in future.</p>
The occurrence of over-excavation or under-excavation in tunnel construction poses significant safety risks. Moreover, there is currently no automatic estimation method available for real-time estimation of earthwork excavation, particularly in the case of shield tunnels. In this study, we tracked the excavation process of Chengdu Metro Line 19, acquired tunneling parameters and earthwork excavation data using various sensors, and subsequently proposed an automatic estimation method that combines Bayesian optimization (BO) and gradient boosting regression tree (GBRT) algorithm. The results of our case study indicate that the BO-GBRT model improves the performance of earthwork excavation estimation, reducing the residual after each calculation with a root mean square error (RMSE) of 1.712 and mean absolute error (MAE) of 1.331. Furthermore, compared to other machine learning methods, the proposed BO-GBRT model demonstrates superior estimation performance. Additionally, the importance distribution of input parameters reveals that propulsion pressure, foam pressure, and rotation speed are the most critical factors affecting earthwork excavation. Overall, the proposed automatic estimation method shows great promise as a tool for efficiently estimating earthwork excavation in shield tunnel construction.
Pietro Barbiero, Stefano Fioravanti, Francesco Giannini
et al.
Explainable AI (XAI) aims to address the human need for safe and reliable AI systems. However, numerous surveys emphasize the absence of a sound mathematical formalization of key XAI notions -- remarkably including the term "explanation" which still lacks a precise definition. To bridge this gap, this paper presents the first mathematically rigorous definitions of key XAI notions and processes, using the well-funded formalism of Category theory. We show that our categorical framework allows to: (i) model existing learning schemes and architectures, (ii) formally define the term "explanation", (iii) establish a theoretical basis for XAI taxonomies, and (iv) analyze commonly overlooked aspects of explaining methods. As a consequence, our categorical framework promotes the ethical and secure deployment of AI technologies as it represents a significant step towards a sound theoretical foundation of explainable AI.