Coupling generative and predictive machine learning algorithms to enhance haloacetonitriles prediction in small water systems.

Estimating causal effects in quasi-experiments with spatio-temporal panel data often requires adjusting for unmeasured confounding that varies across space and time. Gaussian processes (GPs) offer a flexible, nonparametric modeling approach that can account for such complex dependencies through carefully chosen covariance kernels. In this paper, we provide a practical and interpretable framework for applying GPs to causal inference in panel data settings. We demonstrate how GPs generalize popular methods such as synthetic control and vertical regression, and we show that the GP posterior mean can be represented as a weighted average of observed outcomes, where the weights reflect spatial and temporal similarity. To support applied use, we explore how different kernel choices impact both estimation performance and interpretability, offering guidance for selecting between separable and nonseparable kernels. Through simulations and application to Hurricane Katrina mortality data, we illustrate how GP models can be used to estimate counterfactual outcomes and quantify treatment effects. All code and materials are made publicly available to support reproducibility and encourage adoption. Our results suggest that GPs are a promising and interpretable tool for addressing unmeasured spatio-temporal confounding in quasi-experimental studies.

Spatially correlated multi-state fragility via a warped Gaussian process

Hybrid and explainable machine learning for predicting water penetration depth in seawater based self-compacting concrete

Residual-enhanced Gaussian process network driving risk-aware multi-objective optimization for data-efficient truck frame design

Learning nonlinearity and measuring uncertainty--a multi-task neural network and additive gaussian process based travel choice model

State prediction from noisy time series measurements is a challenging task found in areas like intelligent transport, structural health monitoring, and environmental monitoring. This paper proposes a Spatio-Temporal Deep Kernel Gaussian Process (STDK GP) approach, which leverages the feature extraction capabilities of convolutional neural networks with the uncertainty quantification of Gaussian Process (GP) methods. The model features a composite spatio-temporal kernel that operates on learned deep features. A key aspect of our approach is that this kernel is learned end-to-end with the feature extractor, allowing it to effectively capture complex spatial and temporal patterns and enabling robust uncertainty quantification. Evaluated on a real-world vehicular traffic forecasting task, the proposed STDK GP demonstrates superior performance. Specifically, it achieves a root mean square error of 2.67105 km/h and improves prediction accuracy by approximately 15.28% over the standard GP approach, 21.23% over the state-of-the-art structural recurrent neural network and by more than 53% over stand-alone deep neural networks.

A GENERIC-guided active learning SPH method for viscoelastic fluids using Gaussian process regression

• Identification of load transfer mechanism of dry beam–column connection. • Development of collapse-resisting design of high-performance dry connection. • Development of macro joint model considering joint load transfer mechanisms. • Modification of reliability assessment method for strongly nonlinear problem. • Gaussian process regression model-based reliability sensitivity analysis. Beam–column joint connections are critical for precast concrete frames to resist collapse, yet their internal load transfer mechanisms and demand-side uncertainties remain underexplored. This study develops a preliminary reliability-aware collapse-resisting design framework for a high-performance dry connection, incorporating strengthened steel angles and high-strength bolts. The design part begins with the identification of load transfer mechanisms and configuration optimization to eliminate undesirable failure modes based on detailed finite element models. Thereby the design focuses on upper-bound estimations of column bolt axial force demand, derived based on rebar ultimate strengths. The reliability assessment of the derived demand relies on: (1) a newly developed and validated macro joint model, alleviating the computational burden of repeated model evaluations; and (2) a modified Active Learning Probabilistic Integration method ensuring highly efficient reliability assessment using a small number of model evaluations. The learning function and the point selection strategy are tailored to the highly nonlinear nature of the involved performance function. Material uncertainties in the structural components connected by the joint are explicitly considered because they significantly influence joint demands. Moreover, variance-based global sensitivity analysis and local reliability sensitivity analysis are performed by post-processing the Gaussian Process regression model obtained from the reliability assessment. The results indicate that the rebar ultimate strength, yield strength, and fracture strain are the most influential random variables affecting the design reliability. Their mean values are recommended to be explicitly considered in the design phase in future research; tighter quality control on rebar production, aimed at reducing material property variability, can further improve the design reliability.

Accurate and rapid characterization of lung mechanics remains a central challenge in respiratory disease management. Physics-informed poroelastic finite-element (FE) models resolve detailed tissue-airflow interactions but are computationally prohibitive for real-time or large-scale clinical applications, while lumped-parameter models sacrifice mechanistic fidelity for efficiency. In this work, we present a porcine-specific, multi-fidelity computational framework that integrates poroelastic FE modeling with machine learning to enable rapid, uncertainty-aware estimation of respiratory compliance ( ) and resistance ( ). High- and low-fidelity simulations are generated from CT-derived porcine lung geometries by sampling a physiologically relevant parameter space, and the resulting pressure-volume dynamics are used in an inverse modeling procedure to infer global respiratory mechanics. A key result is that multi-fidelity Gaussian process (MF-GP) surrogates achieve accurate predictions of and with errors below 5% relative to high-fidelity simulations, while providing computational speedups of over five orders of magnitude. In contrast, neural network (NN) surrogates exhibit relatively poor generalization in the data-scarce regime considered, highlighting the importance of model selection for scientific machine learning under limited high-fidelity data availability. Beyond predictive performance, global sensitivity analysis reveals a clear mechanistic separation in parameter influence: compliance is primarily governed by elastic stiffness and chest-wall coupling, whereas resistance is dominated by permeability. The weak interaction effects observed support an approximately additive response structure, enabling robust parameter identifiability and reduced-order representations of the inverse problem. The framework is validated against independent ventilator measurements from porcine lungs, showing strong agreement within clinically observed ranges. Overall, this study provides new insight into the structure of the inverse problem in poroelastic lung modeling and establishes a computationally efficient pathway for uncertainty-aware prediction and parameter estimation, with potential applications in personalized ventilation and preclinical study design.

To evaluate the methodological feasibility of a heterogeneous multi-output Gaussian process model for jointly handling a continuous birth outcome and its clinically used binary representation in routinely collected perinatal data, and to compare its predictive performance with that of conventional single-output models. Routinely collected live-birth certificate data from Medellín, Colombia, covering births from 2012 to 2021, were analyzed. After cleaning and class balancing, the analytic dataset included 32,110 records. A heterogeneous multi-output Gaussian process model was trained to jointly model birth weight in grams with a Gaussian likelihood and low birth weight status with a Bernoulli likelihood. Predictive performance was compared with that of conventional single-output regression and classification models. The heterogeneous multi-output Gaussian process model achieved acceptable predictive performance (R² = 0.67 for birth weight and accuracy = 0.845 for low birth weight classification), with results comparable to those of models fitted separately for each task. These findings support the practical feasibility of modeling heterogeneous outputs within a single probabilistic framework. In this application, the heterogeneous multi-output Gaussian process model was a viable methodological alternative for jointly modeling birth weight in grams and its binary low-birth-weight classification. This study should be interpreted primarily as a methodological demonstration of a flexible multi-output framework in perinatal data that may be extended in future studies to jointly model other outcomes of greater direct relevance to public health.

The government and investors have long put a tremendous lot of weight on commodities price projections. This study looks at the difficult problem of daily regional steel price index forecasting in the south Chinese market from January 1, 2010, to April 15, 2021. The literature has not given adequate attention to the forecast of this important commodity price indicator. We use Gaussian process regressions to validate our results after employing cross-validation and Bayesian optimizations to train our models. With an out-of-sample relative root mean square error of 0.5352%, the models that were constructed correctly forecasted the price indices between January 8, 2019, and April 15, 2021. The models developed can be used by policymakers and investors for analysis and decision-making. When reference data on the price patterns recommended by these models are employed, forecasting findings may be useful in creating comparable commodity price indices.

To improve bridge deck icing warning, this study investigates two related tasks, namely bridge deck temperature prediction and bridge deck icing-state prediction, within a unified surrogate-modeling framework. Three representative surrogate models, including deep neural network (DNN), Gaussian process (GP), and polynomial chaos expansion (PCE), are established and compared in COMSOL under consistent workflows, datasets, and evaluation metrics. For bridge deck temperature prediction, real monitoring data collected from a bridge in Shandong Province are used. For bridge deck icing-state prediction, both a MATLAB-generated synthetic dataset and a real dataset derived from monitored samples through physical labeling and class balancing are considered. The results show that all three surrogate models can effectively predict bridge deck temperature, among which GP achieves the best overall regression performance, while DNN performs comparably well. For icing-state prediction, DNN exhibits the best overall classification performance on both the synthetic and real datasets. GP and PCE show relatively stronger ability to identify icy samples on the real dataset, but this advantage is accompanied by an increased probability of false alarms. The results indicate that the suitability of surrogate models is task-dependent: GP is more advantageous for bridge deck temperature regression, whereas DNN provides a better balance of accuracy, robustness, and generalization for bridge deck icing-state prediction. This study provides a useful reference for surrogate-model selection in bridge deck icing warning.

Abstract Background Ecosystems tend to fluctuate around stable equilibria in response to internal dynamics and environmental factors. Occasionally, they enter an unstable tipping region and collapse into an alternative stable state. Being able to quantify and predict these dynamics is key to our understanding of how microbial communities vary over time and respond to perturbations. Results Mechanistic models of microbial community dynamics often fail to characterise observed fluctuations in naturally occurring microbiomes and inform us about key dynamical properties such as stability and resilience. An alternative approach is to characterise the dynamical landscape using non-parametric models. However, the scarcity of long, dense time series data poses a severe bottleneck for characterising community dynamics using existing methods. We overcome this limitation by combining information across multiple short time series using Bayesian inference. By decomposing dynamics into deterministic and stochastic components using Gaussian process priors, we are able to predict stable and tipping regions along a unidimensional stability landscape while simultaneously addressing the associated uncertainty. In particular, we estimate a recently proposed probabilistic metric for resilience in multistable systems: the expected “exit time” out of the current stable state under stochastic fluctuations. We validate our approach on simulated data and highlight in particular that our model is able to distinguish bistability from bimodality, which are often conflated in classical potential analyses. We further demonstrate our approach by re-analysing ecological time series data of lake cyanobacteria abundance, for which we recover similar results as a previous study using three orders of magnitude fewer data points. Finally, we use our model to re-evaluate the stability of previously proposed “tipping elements” in the human gut microbiota. Conclusions We introduce a probabilistic non-parametric approach to characterise stationary community dynamics from short time series, which is potentially applicable to a broad range of systems in microbial ecology and beyond. We use this model to clarify the distinction between bistable and bimodal dynamics and to contribute to contemporary debates on the stability and resilience of ecological communities, in particular the human gut microbiota.

ABSTRACT Cloud base height (CBH) is a fundamental atmospheric parameter for weather forecasting, aviation safety, and climate research, as it provides key information on boundary layer structure, atmospheric stability, and cloud–radiation interactions. In this study, a two-step hybrid framework combining statistical cloud detection with machine learning (ML) regression for CBH estimation is developed and evaluated. In the first step, cloud presence is identified using a Variability Index (VI) defined as the ratio of the standard deviation of the backscatter profile to its peak value. This physically interpretable index shows strong class separability, with a large effect size (Cohen’s d ≈ 1.96), a maximum F1 score of about 0.83 at a VI of approximately 0.24, and an area under the ROC curve of 0.88, indicating effective cloud detection performance. In the second step, Multiple Linear Regression, Fine Tree Regression, Random Forest, and Gaussian Process Regression (GPR) models are applied to cloud-present profiles to estimate CBH. Among these, Random Forest, a tree based nonlinear ensemble model perform best, achieving a high correlation coefficients of about 0.94 ± 0.04. GPR, a kernel-based model, demonstrated slightly lower performance compared to Random Forest, achieving correlation coefficients of R = 0.91 ± 0.05, while the Fine Tree model showed the weakest performance among the nonlinear models tested in this study, achieving R = 0.89 ± 0.07. In contrast, Multiple Linear Regression model showed lowest accuracy with R = 0.58 ± 0.13. The results demonstrate that combining a simple, explainable statistical classification approach with advanced machine learning regression significantly improves the reliability and accuracy of CBH retrieval from Lidar backscatter data. The proposed framework is computationally efficient and shows strong potential for operational implementation in real-time atmospheric monitoring networks.

High-throughput screening workflows often rank materials with a sequence of filters, but a single sequence can hide how strongly the final ranking depends on the chosen physical approximations. Here a multiscale cascade means an ordered workflow in which outputs from electronic, lattice, microstructural, uncertainty, and device-level models are passed from one layer to the next; the layers are theory or surrogate-model steps, not layers of LLMs. We use a large language model (LLM) as a workflow-design assistant to assemble candidate model stacks from the literature, after which the equations, code, and physical handoffs are inspected and implemented by the author. The test case is thermoelectric screening, where the dimensionless figure of merit ZT = S2σT/(κe + κL) combines the Seebeck coefficient S, electrical conductivity σ, electronic thermal conductivity κe, and lattice thermal conductivity κL. Two independently assembled eight-layer cascades are applied to the same 314-compound vacancy-containing chalcogenide library. LLM1 uses Fan-Migdal band gap renormalization, a Kubo-DMFT transport surrogate, and a literature-trained Gaussian process for κL. LLM2 replaces those three early layers with Bose-Einstein band gap renormalization, acoustic-phonon Boltzmann transport, and a Debye-Callaway integral. The common four-stage screen sends 15 compounds to full evaluation in each cascade: LLM1 selects tellurides headed by CuAlTe2 (ZTpeak = 8.61), whereas LLM2 selects nonoverlapping sulfides headed by CuPb2S4 (ZTpeak = 0.325). The absolute LLM1 values are not claimed as validated performance forecasts: CuAlTe2 is literature-supported as a promising bulk thermoelectric, but reported and expected values are closer to ZT ≲ 2 than to 8.6. This study underscores the importance of transparently comparing multiple workflow designs to understand the sensitivity and reliability of high-throughput screening outcomes in materials discovery.

A Gaussian Process Regression (GPR) model was developed to predict the shear capacity of exterior reinforced concrete (RC) beam-column joints subjected to seismic loading. The model accounts for key parameters, including beam and column geometry, reinforcement detailing, axial column load, and concrete compressive strength. A database of 273 experimentally tested specimens was used, with emphasis on horizontal joint shear strength. Three kernel structures within the GPR framework-Primary, Rational Second-Order, and Combined kernels-were examined. Model predictions were evaluated against existing shear-strength formulations using deterministic metrics (MAE, RMSE, and R2) and probabilistic measures (NLPD and MSLL). The results show that kernel effectiveness depends on model formulation; however, the Combined kernel exhibited more stable predictions and improved uncertainty calibration for models with higher-dimensional input sets. Sensitivity analysis identified concrete compressive strength, beam depth, and joint transverse reinforcement as dominant variables, followed by column height, axial load ratio, and reinforcement configuration. Overall, the proposed framework enables consistent shear-strength prediction and quantifies the relative influence of geometric and material parameters, contributing to more informed assessment and design of RC beam-column joints.

Computed Tomography (CT) imaging has expanded possibilities for biological profile estimation in forensic contexts. This study examined whether two-dimensional (2D) morphometric measurements and Hounsfield Unit (HU) values derived from CT scans of dry proximal femora could reliably estimate sex, stature, and age, and whether machine learning (ML) could meaningfully improve on traditional methods. Three hundred left femora from Thai individuals were scanned, and mid-coronal sections were used to extract measurements from defined anatomical regions. For sex estimation, conventional estimation equations reached 93.2% accuracy, while Naïve Bayes classification achieved 96.5% as the best performance among the ML models tested. Stature estimation using sex-specific 2D parameters yielded a Standard Error of Estimate (SEE) of 4.43cm, which dropped to 3.96cm when Support Vector Machines (SVM) and Gaussian Process Regression (GPR) were applied. Age estimation relied on HU values, which showed a consistent negative relationship with age. The lowest SEE for age was 9.67 years from measurements at the Primary Tensile Line (PTL) and Ward's Triangle in females. Models also performed better when applied to older age groups. Although sex-specific equations outperformed mixed-sex ones, the latter were kept in the analysis as a practical alternative when sex cannot be established prior to analysis. Overall, 2D morphometrics proved most useful for sex and stature estimation, while HU values emerged as a reliable, quantitative approach to age estimation. ML consistently improved model performance across all three estimation tasks, supporting its role in modern forensic anthropological practice.

Raman spectrometry, with its capability to noninvasively characterize the molecular composition of microscopic subcellular volumes, including single organelles in live cells, has revolutionized cell biology research. Being introduced as a label-free approach for biochemical imaging, the practical applications of Raman spectrometry still often include the fluorescence probes for the localization of organelles and other subcellular domains of interest. Aiming to overcome this limitation, we report on the development of an artificial intelligence/machine learning approach for true label-free identification of different types of subcellular structures. Here, we explore the application of machine learning (ML) to learn the relationship between a set of biochemical parameters in single organelles of live cells. The biochemical parameters are extracted by Ramanomics, an optical Omics technology, from Raman spectra of single organelles of live cells of different cell lines. Several classification algorithms, such as neural networks, Random Forests, support vector machines, logistic regression, and Gaussian process classification, are evaluated. We report the performance of the best classifier, a shallow neural network, to classify the type of organelle using the biochemical parameters. Evaluation is done using k-fold cross-validation (k = 10), and the final output classification is compared against the ground truth. The k-fold cross-validation shows that the NN-based classifier has significant accuracy (∼90%) to distinguish between different organelles using Ramanomics measurements. Our approach allows us to identify the precise location of separate organelles by local Raman measurement without labeling.

High-temperature structural components are susceptible to creep deformation, which can ultimately lead to failure. In this work, an AI-based framework was developed capable of predicting the creep time of 316 austenitic stainless steel. Here, creep time refers to both the time to reach specific strain levels and the time to rupture. However, the scope of the present work is limited to rupture-time prediction, while the application of the framework to strain-level prediction will be reported in future work. The dataset consisted of creep strain curves from four heats, including both rupture and non-rupture curves. Random Forest (RF), Gradient Boosting (GB), Extreme Gradient Boosting (XGB), Support Vector Regressor (SVR), Gaussian Process Regressor (GPR), and Neural Network (NN) were employed. The effects of square-root and cube-root transformations on data distribution and model learning behaviour were analysed using model learning curves. An Optuna (version 4.3.0)-based hyperparameter tuning strategy was employed. The cube-root transformation improved the learning performance of SVR, GPR, and NN, whereas RF, GB, and XGB remained unaffected. Learning curves revealed mild overfitting for RF, GB, and XGB, and very minimal overfitting for SVR, GPR, and NN. NN achieved the best predictive performance (R2=0.92,RMSE=0.195, deviation factor of 1.57). The findings demonstrated that the combined useof creep strain curves, data transformation, learning curve guided model selection, and rigorous hyperparameter tuning can improve the prediction accuracy under a limited dataset.