Metin Zontul
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Metin Zontul has a total of 32 co-authors. Here are the top co-authors by number of shared publications:
- Seda Karateke
Department of Software Engineering, Faculty of Engineering and Natural Sciences, Istanbul Atlas University, Istanbul 34408, Türkiye
12 shared publications - George A. Anastassiou
Department of Mathematical Sciences, University of Memphis, Memphis, TN 38152, USA
7 shared publications - Ilkay Yelmen
Department of Software Engineering, Faculty of Engineering and Natural Sciences, Istinye University, Istanbul 34396, Türkiye
6 shared publications - Ali Güneş
Software Development Department, İstanbul Aydın University, Istanbul 34295, Türkiye
3 shared publications - Fazli Yildirim
Faculty of Economics, Administrative and Social Sciences, Department of Management Information Systems, Istanbul Topkapı University, 34087 Istanbul, Türkiye
3 shared publications - Tolga Karaköy
Department of Plant Protection, Faculty of Agricultural Sciences and Technologies, Sivas University of Science and Technology, Sivas 58000, Türkiye
3 shared publications
Latest Publications
- Open Access
- 26 October 2026
- Published by MDPI AG in Mathematics
- Vol. 14 (21) , 3911
Moment Structure and Approximation Properties of Univariate and Bivariate Durrmeyer–Beta-Type Operators
We study a Durrmeyer–Beta-type positive linear operator on the half-line, constructed from rising-factorial weights and a normalized beta-prime kernel. Exact raw moments through degree four and the corresponding central moments are derived under the required moment-existence conditions. These identities yield local approximation estimates, weighted Korovkin convergence, and a weighted rate for functions with bounded first and second derivatives. A tensor-product extension provides bivariate convergence and error bounds involving partial moduli of continuity and coordinatewise regularity. Comparisons with related Durrmeyer constructions clarify the operator’s structure, while a matched-index comparison with a Dunkl–Beta operator examines its numerical accuracy. The numerical examples include smooth, oscillatory, and nonsmooth bounded functions, with an additional experiment on . We also investigate reconstruction from noisy observations using 30 signal–noise realizations at each of five Gaussian noise levels. The comparison includes a Gaussian kernel smoother, an ARIMA model, and an LSTM reconstruction network. The Gaussian smoother gives the best mean accuracy at the lowest noise level, whereas gives the lowest mean normalized errors and the highest mean at the four higher levels. These experiments illustrate the approximation and smoothing behavior of the operator under the stated computational settings.- 26 October 2026
- Published by MDPI AG in Mathematics
- Vol. 14 (21) , 3911
Open AccessMoment Structure and Approximation Properties of Univariate and Bivariate Durrmeyer–Beta-Type Operators
We study a Durrmeyer–Beta-type positive linear operator on the half-line, constructed from rising-factorial weights and a normalized beta-prime kernel. Exact raw moments through degree four and the corresponding central moments are derived under the required moment-existence conditions. These identities yield local approximation estimates, weighted Korovkin convergence, and a weighted rate for functions with bounded first and second derivatives. A tensor-product extension provides bivariate convergence and error bounds involving partial moduli of continuity and coordinatewise regularity. Comparisons with related Durrmeyer constructions clarify the operator’s structure, while a matched-index comparison with a Dunkl–Beta operator examines its numerical accuracy. The numerical examples include smooth, oscillatory, and nonsmooth bounded functions, with an additional experiment on . We also investigate reconstruction from noisy observations using 30 signal–noise realizations at each of five Gaussian noise levels. The comparison includes a Gaussian kernel smoother, an ARIMA model, and an LSTM reconstruction network. The Gaussian smoother gives the best mean accuracy at the lowest noise level, whereas gives the lowest mean normalized errors and the highest mean at the four higher levels. These experiments illustrate the approximation and smoothing behavior of the operator under the stated computational settings. - Open Access
- 10 October 2026
- Published by MDPI AG in Agriculture
- Vol. 16 (20) , 2197
Predicting Lentil Seed Yield from Agronomic Traits Using a Foundation-Augmented Machine Learning Ensemble
Accurate agricultural yield prediction is crucial for crop management and data-driven decision-making. This study proposes a machine learning framework for predicting total lentil seed yield using 264 observations collected during a single growing season at one location in Sivas, Türkiye. Seven regression models and the tabular foundation model TabPFN were evaluated through nested cross-validation. The fixed post hoc Foundation-Augmented Averaging Ensemble (FAAE) combined TabPFN, Extra Trees, CatBoost, and XGBoost using equal weights, while the primary fully nested evaluation selected the four ensemble members within each outer training fold. The fully nested FAAE achieved R2 = 0.891 ± 0.045, comparable to the strongest individual models, while the fixed post hoc FAAE yielded R2 = 0.897 ± 0.038. Split conformal prediction produced well-calibrated uncertainty intervals at 80%, 90%, and 95% coverage. SHAP and LOCO identified pods per plant and biological yield as the dominant predictors of lentil yield. Overall, these findings indicate that FAAE can provide competitive predictive performance within a leakage-controlled framework while also supporting uncertainty quantification and biologically meaningful interpretation. Because the dominant predictors, pods per plant and biological yield, are measured at or near physiological maturity, the proposed framework is more suitable for end-of-season yield estimation than for early-season forecasting. The framework therefore shows potential for within-trial yield assessment, although validation using replicated material across additional years and locations is required before broader generalization.
- 10 October 2026
- Published by MDPI AG in Agriculture
- Vol. 16 (20) , 2197
Open AccessPredicting Lentil Seed Yield from Agronomic Traits Using a Foundation-Augmented Machine Learning Ensemble
Accurate agricultural yield prediction is crucial for crop management and data-driven decision-making. This study proposes a machine learning framework for predicting total lentil seed yield using 264 observations collected during a single growing season at one location in Sivas, Türkiye. Seven regression models and the tabular foundation model TabPFN were evaluated through nested cross-validation. The fixed post hoc Foundation-Augmented Averaging Ensemble (FAAE) combined TabPFN, Extra Trees, CatBoost, and XGBoost using equal weights, while the primary fully nested evaluation selected the four ensemble members within each outer training fold. The fully nested FAAE achieved R2 = 0.891 ± 0.045, comparable to the strongest individual models, while the fixed post hoc FAAE yielded R2 = 0.897 ± 0.038. Split conformal prediction produced well-calibrated uncertainty intervals at 80%, 90%, and 95% coverage. SHAP and LOCO identified pods per plant and biological yield as the dominant predictors of lentil yield. Overall, these findings indicate that FAAE can provide competitive predictive performance within a leakage-controlled framework while also supporting uncertainty quantification and biologically meaningful interpretation. Because the dominant predictors, pods per plant and biological yield, are measured at or near physiological maturity, the proposed framework is more suitable for end-of-season yield estimation than for early-season forecasting. The framework therefore shows potential for within-trial yield assessment, although validation using replicated material across additional years and locations is required before broader generalization.
- Open Access
Explainable Machine Learning for Methylene Blue Removal Under Irradiation: Assessing Nominal Cu-Loading in AC@NiO
The time-dependent removal performance under irradiation of activated-carbon-supported nickel oxide (AC@NiO) samples containing nominal Cu-loadings of 0%, 3%, 5%, 7.5%, and 10% was examined using experimental concentration data and an explainable machine-learning (ML) framework. Reaction time and nominal Cu-loading were used as predictors. Preliminary screening with as the common target showed that Random Forest (RF) outperformed K-Nearest Neighbors, Multi-Layer Perceptron, and Support Vector Regression. The target-specific regularized RF models were fitted for and ; normalized apparent removal efficiency was derived deterministically as . Across 10 controlled random seeds, the mean leave-one-out cross-validation values were for and the derived , and for . A fixed chronological evaluation, trained at min and evaluated at min, yielded mean values of and , respectively. Because the composition-specific RF predictions were constant across this boundary interval, these scores are interpreted as later-time boundary diagnostics rather than evidence of temporal extrapolation. Point-estimate SHAP, impurity-based, and permutation-based analyses assigned greater predictive importance to reaction time, whereas moving-block bootstrap results showed that this ranking was sensitive to temporal resampling. Among the five tested compositions, the nominal 5% Cu-containing sample exhibited the most favorable apparent removal profile. This finding is specific to the investigated conditions and does not establish a universal optimum or a causal physicochemical mechanism.Open AccessExplainable Machine Learning for Methylene Blue Removal Under Irradiation: Assessing Nominal Cu-Loading in AC@NiO
The time-dependent removal performance under irradiation of activated-carbon-supported nickel oxide (AC@NiO) samples containing nominal Cu-loadings of 0%, 3%, 5%, 7.5%, and 10% was examined using experimental concentration data and an explainable machine-learning (ML) framework. Reaction time and nominal Cu-loading were used as predictors. Preliminary screening with as the common target showed that Random Forest (RF) outperformed K-Nearest Neighbors, Multi-Layer Perceptron, and Support Vector Regression. The target-specific regularized RF models were fitted for and ; normalized apparent removal efficiency was derived deterministically as . Across 10 controlled random seeds, the mean leave-one-out cross-validation values were for and the derived , and for . A fixed chronological evaluation, trained at min and evaluated at min, yielded mean values of and , respectively. Because the composition-specific RF predictions were constant across this boundary interval, these scores are interpreted as later-time boundary diagnostics rather than evidence of temporal extrapolation. Point-estimate SHAP, impurity-based, and permutation-based analyses assigned greater predictive importance to reaction time, whereas moving-block bootstrap results showed that this ranking was sensitive to temporal resampling. Among the five tested compositions, the nominal 5% Cu-containing sample exhibited the most favorable apparent removal profile. This finding is specific to the investigated conditions and does not establish a universal optimum or a causal physicochemical mechanism. - Open Access
Explainable TabPFN-Based Machine Learning for Single-Plant Yield Estimation and Trait Prioritization in Faba Bean (Vicia faba L.)
Faba bean yield reflects complex relationships among genotype, environment, and agronomic traits. This study evaluated an explainable Tabular Prior-data Fitted Network (TabPFN) framework for estimating plot-mean single-plant yield and prioritizing traits using 398 plot-level observations, 13 measured agronomic predictors, and six derived features. On the reference 80/20 split, TabPFN achieved the best values for all four test metrics (R2 = 0.8746, RMSE = 1.9132 g plant−1, MAE = 1.0819 g plant−1, and MAPE = 8.16%). The Friedman test detected differences among the six models (χ2(5) = 16.75, p = 0.005); Nemenyi comparisons distinguished TabPFN from HistGradientBoosting and SVR, whereas the Holm-corrected Wilcoxon analysis confirmed only the TabPFN–SVR difference. Across 10 repeated 80/20 splits, TabPFN obtained the highest mean test R2 (0.8614 ± 0.0691), ranked first in eight splits, and produced a higher R2 than every tuned baseline in at least eight splits. SHAP, permutation importance, and LOCO analyses emphasized pod-, seed-, and biomass-related predictors. Repeated-split ablation showed that derived features improved TabPFN consistently, whereas removing selected target-proximal yield variables reduced performance for every model. The framework is therefore a harvest-time trait-estimation and trait-prioritization tool rather than an early-season forecasting system. Notably, TabPFN achieved this performance without the 100-trial Optuna search used for each baseline; only n_estimators was screened over four prespecified values.
Open AccessExplainable TabPFN-Based Machine Learning for Single-Plant Yield Estimation and Trait Prioritization in Faba Bean (Vicia faba L.)
Faba bean yield reflects complex relationships among genotype, environment, and agronomic traits. This study evaluated an explainable Tabular Prior-data Fitted Network (TabPFN) framework for estimating plot-mean single-plant yield and prioritizing traits using 398 plot-level observations, 13 measured agronomic predictors, and six derived features. On the reference 80/20 split, TabPFN achieved the best values for all four test metrics (R2 = 0.8746, RMSE = 1.9132 g plant−1, MAE = 1.0819 g plant−1, and MAPE = 8.16%). The Friedman test detected differences among the six models (χ2(5) = 16.75, p = 0.005); Nemenyi comparisons distinguished TabPFN from HistGradientBoosting and SVR, whereas the Holm-corrected Wilcoxon analysis confirmed only the TabPFN–SVR difference. Across 10 repeated 80/20 splits, TabPFN obtained the highest mean test R2 (0.8614 ± 0.0691), ranked first in eight splits, and produced a higher R2 than every tuned baseline in at least eight splits. SHAP, permutation importance, and LOCO analyses emphasized pod-, seed-, and biomass-related predictors. Repeated-split ablation showed that derived features improved TabPFN consistently, whereas removing selected target-proximal yield variables reduced performance for every model. The framework is therefore a harvest-time trait-estimation and trait-prioritization tool rather than an early-season forecasting system. Notably, TabPFN achieved this performance without the 100-trial Optuna search used for each baseline; only n_estimators was screened over four prespecified values.
- Open Access
Explainable Artificial Intelligence Assisted Modeling of Malachite Green Adsorption onto SBA-15–Zn–Fe Composite
Dye-containing industrial effluents pose substantial risks to aquatic ecosystems. In this study, a material prepared using the stated SBA-15–Zn–Fe synthesis procedure was evaluated for malachite green (MG) removal and modeled using machine-learning and explainable artificial intelligence techniques. The dataset comprised 315 experimental observations covering initial MG concentrations of 100–500 mg L−1, adsorbent concentrations of 0.5–3.0 g L−1, pH values of 5–9, temperatures of 25–45 °C, and contact times of 0–360 min. At an initial MG concentration of 500 mg L−1, an adsorbent concentration of 1.0 g L−1, pH 9.0, 45 °C, and 90 min, the material achieved 99.65% MG removal. The Langmuir-estimated maximum monolayer capacity was 1428.57 mg g−1. A multilayer perceptron (MLP) with a 100–50–25–12 hidden-layer architecture, hyperbolic tangent activation, and L-BFGS optimization was developed to predict residual MG concentration. The prespecified MLP achieved an of 0.9624 on the strictly held-out 20% internal test subset. Five-fold cross-validation yielded a mean of 0.8420 with a fold-wise standard deviation of 0.1273, while pooled contact-time-based LOGO-CV yielded an of 0.7411, indicating reduced transferability when entire contact-time groups were excluded from fitting. Nominal 95% CV+ prediction intervals achieved 98.41% empirical coverage on the held-out test subset, although their relatively broad widths indicated non-negligible predictive uncertainty. PFI and SHAP showed that the fitted MLP relied most strongly on contact time and initial MG concentration. Longer contact times were generally associated with lower predicted residual concentrations, whereas higher initial concentrations were associated with higher predicted residual concentrations. Overall, the framework provided an interpretable assessment of MG adsorption within the investigated experimental domain; external validity and extrapolation beyond this domain were not established.Open AccessExplainable Artificial Intelligence Assisted Modeling of Malachite Green Adsorption onto SBA-15–Zn–Fe Composite
Dye-containing industrial effluents pose substantial risks to aquatic ecosystems. In this study, a material prepared using the stated SBA-15–Zn–Fe synthesis procedure was evaluated for malachite green (MG) removal and modeled using machine-learning and explainable artificial intelligence techniques. The dataset comprised 315 experimental observations covering initial MG concentrations of 100–500 mg L−1, adsorbent concentrations of 0.5–3.0 g L−1, pH values of 5–9, temperatures of 25–45 °C, and contact times of 0–360 min. At an initial MG concentration of 500 mg L−1, an adsorbent concentration of 1.0 g L−1, pH 9.0, 45 °C, and 90 min, the material achieved 99.65% MG removal. The Langmuir-estimated maximum monolayer capacity was 1428.57 mg g−1. A multilayer perceptron (MLP) with a 100–50–25–12 hidden-layer architecture, hyperbolic tangent activation, and L-BFGS optimization was developed to predict residual MG concentration. The prespecified MLP achieved an of 0.9624 on the strictly held-out 20% internal test subset. Five-fold cross-validation yielded a mean of 0.8420 with a fold-wise standard deviation of 0.1273, while pooled contact-time-based LOGO-CV yielded an of 0.7411, indicating reduced transferability when entire contact-time groups were excluded from fitting. Nominal 95% CV+ prediction intervals achieved 98.41% empirical coverage on the held-out test subset, although their relatively broad widths indicated non-negligible predictive uncertainty. PFI and SHAP showed that the fitted MLP relied most strongly on contact time and initial MG concentration. Longer contact times were generally associated with lower predicted residual concentrations, whereas higher initial concentrations were associated with higher predicted residual concentrations. Overall, the framework provided an interpretable assessment of MG adsorption within the investigated experimental domain; external validity and extrapolation beyond this domain were not established. - Open Access
Synthetic Data-Guided Symmetric Neural Network Approximation in Banach Spaces
This paper develops a Banach space-valued approximation framework based on symmetrized neural network (SNN) operators generated by a deformation-dependent sigmoidal activation function. Symmetry is introduced directly at the activation level through a reciprocal-deformation mechanism, yielding a positive, even, normalized, and localized density kernel satisfying the partition of unity. The resulting construction provides normalized compact-interval and whole-line quasi-interpolation operators for Banach space-valued functions. Quantitative pointwise and uniform convergence estimates are established through the first modulus of continuity and are extended to higher-order and Caputo–Bochner fractional approximation. Numerical diagnostics support the theoretical kernel properties, and fractional approximation experiments compare the SNN and classical NN operators under common computational conditions. A controlled blind-prediction experiment on a synthetic monthly temperature-like series uses a strict fit–validation–test protocol and a parameter-matched operator comparison, with seasonal ARIMA and MLP models as external baselines. Across five independent realizations, the SNN attains the best mean predictive performance, with , NMAE , and NRMSE . A vector-valued experiment in further illustrates the non-scalar applicability of the Banach space framework. In addition, the normalized SNN kernel weights provide an intrinsic node-level interpretation mechanism without requiring an external post hoc explainability method.Open AccessSynthetic Data-Guided Symmetric Neural Network Approximation in Banach Spaces
This paper develops a Banach space-valued approximation framework based on symmetrized neural network (SNN) operators generated by a deformation-dependent sigmoidal activation function. Symmetry is introduced directly at the activation level through a reciprocal-deformation mechanism, yielding a positive, even, normalized, and localized density kernel satisfying the partition of unity. The resulting construction provides normalized compact-interval and whole-line quasi-interpolation operators for Banach space-valued functions. Quantitative pointwise and uniform convergence estimates are established through the first modulus of continuity and are extended to higher-order and Caputo–Bochner fractional approximation. Numerical diagnostics support the theoretical kernel properties, and fractional approximation experiments compare the SNN and classical NN operators under common computational conditions. A controlled blind-prediction experiment on a synthetic monthly temperature-like series uses a strict fit–validation–test protocol and a parameter-matched operator comparison, with seasonal ARIMA and MLP models as external baselines. Across five independent realizations, the SNN attains the best mean predictive performance, with , NMAE , and NRMSE . A vector-valued experiment in further illustrates the non-scalar applicability of the Banach space framework. In addition, the normalized SNN kernel weights provide an intrinsic node-level interpretation mechanism without requiring an external post hoc explainability method. - Open Access
- 23 July 2026
- Published by MDPI AG in AppliedMath
- Vol. 6 (8) , 119
Activation-Induced Symmetric Kernels for Neural Network Approximation with Quantitative Error Analysis
This paper studies symmetrized neural network (SNN) operators generated by an adjustable half-hyperbolic tangent activation function. The construction is based on the paired density kernels and , whose average defines the symmetric kernel This kernel is positive, even, normalized, and preserves the partition of unity. Using , we define finite-interval and whole-line SNN operators in the Banach space-valued setting. Pointwise and uniform convergence estimates are obtained through the first modulus of continuity. Higher-order and fractional approximation estimates are also derived, the latter using Caputo–Bochner fractional derivatives. The numerical part compares the nonsymmetrized operator and the symmetrized operator . For , the uniform error decreases from to , the root mean square error (RMSE) decreases from to , and the coefficient of determination () improves from to . This improvement is accompanied by an increase in central processing unit (CPU) time from s to s. The parameter tests further show that the performance depends on the joint choice of n, t, and . Overall, the results indicate that symmetrization improves approximation accuracy, while parameter tuning remains necessary.- 23 July 2026
- Published by MDPI AG in AppliedMath
- Vol. 6 (8) , 119
Open AccessActivation-Induced Symmetric Kernels for Neural Network Approximation with Quantitative Error Analysis
This paper studies symmetrized neural network (SNN) operators generated by an adjustable half-hyperbolic tangent activation function. The construction is based on the paired density kernels and , whose average defines the symmetric kernel This kernel is positive, even, normalized, and preserves the partition of unity. Using , we define finite-interval and whole-line SNN operators in the Banach space-valued setting. Pointwise and uniform convergence estimates are obtained through the first modulus of continuity. Higher-order and fractional approximation estimates are also derived, the latter using Caputo–Bochner fractional derivatives. The numerical part compares the nonsymmetrized operator and the symmetrized operator . For , the uniform error decreases from to , the root mean square error (RMSE) decreases from to , and the coefficient of determination () improves from to . This improvement is accompanied by an increase in central processing unit (CPU) time from s to s. The parameter tests further show that the performance depends on the joint choice of n, t, and . Overall, the results indicate that symmetrization improves approximation accuracy, while parameter tuning remains necessary. - Open Access
- 17 July 2026
- Published by MDPI AG in Mathematics
- Vol. 14 (14) , 2599
Parameter Recovery in Noisy Tri-Periodic Horadam Sequences: A Fourier-Embedded Recurrence-Regularized Neural Estimator (RRNE)
This study investigates tri-periodic Horadam-type recurrence systems from both algebraic and computational perspectives. Algebraically, by adapting standard discrete Floquet-type and monodromy matrix viewpoints to the tri-periodic setting, the sequence is analyzed through induced recurrence relations, a monodromy matrix formulation, Binet-type representations, and an effective spectral radius . Computationally, the inverse parameter estimation problem under noisy observations is shown to be highly ill-conditioned. In particular, classical phase-wise Ordinary Least Squares (OLS) estimators may suffer from attenuation bias under an Errors-in-Variables mechanism, causing coefficient shrinkage and over-smoothed trajectory reconstruction in noisy recursive settings. To address this difficulty, we propose a Fourier-Embedded Recurrence-Regularized Neural Estimator (RRNE). By combining Fourier feature embeddings for period-three phase encoding, a Log-Cosh loss for robust residual control, a teacher-forced recurrence constraint, and hybrid Adam–L-BFGS optimization, the RRNE acts as a structure-aware inverse estimator. Numerical experiments across varying noise levels and asymmetry regimes show that the proposed framework provides stable parameter recovery and long-horizon reconstruction in low- and moderate-asymmetry regimes, while exposing severe instability under extreme phase-switching dynamics. The results further indicate that inverse-estimation collapse cannot be explained by the effective spectral radius alone; rather, it reflects the joint effect of coefficient asymmetry, spectral amplification, observation noise, and optimization stiffness. Overall, this study connects tri-periodic recurrence theory, monodromy-based spectral analysis, classical regression limitations, and recurrence-regularized neural inverse modeling.- 17 July 2026
- Published by MDPI AG in Mathematics
- Vol. 14 (14) , 2599
Open AccessParameter Recovery in Noisy Tri-Periodic Horadam Sequences: A Fourier-Embedded Recurrence-Regularized Neural Estimator (RRNE)
This study investigates tri-periodic Horadam-type recurrence systems from both algebraic and computational perspectives. Algebraically, by adapting standard discrete Floquet-type and monodromy matrix viewpoints to the tri-periodic setting, the sequence is analyzed through induced recurrence relations, a monodromy matrix formulation, Binet-type representations, and an effective spectral radius . Computationally, the inverse parameter estimation problem under noisy observations is shown to be highly ill-conditioned. In particular, classical phase-wise Ordinary Least Squares (OLS) estimators may suffer from attenuation bias under an Errors-in-Variables mechanism, causing coefficient shrinkage and over-smoothed trajectory reconstruction in noisy recursive settings. To address this difficulty, we propose a Fourier-Embedded Recurrence-Regularized Neural Estimator (RRNE). By combining Fourier feature embeddings for period-three phase encoding, a Log-Cosh loss for robust residual control, a teacher-forced recurrence constraint, and hybrid Adam–L-BFGS optimization, the RRNE acts as a structure-aware inverse estimator. Numerical experiments across varying noise levels and asymmetry regimes show that the proposed framework provides stable parameter recovery and long-horizon reconstruction in low- and moderate-asymmetry regimes, while exposing severe instability under extreme phase-switching dynamics. The results further indicate that inverse-estimation collapse cannot be explained by the effective spectral radius alone; rather, it reflects the joint effect of coefficient asymmetry, spectral amplification, observation noise, and optimization stiffness. Overall, this study connects tri-periodic recurrence theory, monodromy-based spectral analysis, classical regression limitations, and recurrence-regularized neural inverse modeling. - Open Access
Interpretable Machine Learning Approach for Photocatalytic Degradation in Mn-Doped Semiconductors Using Multilayer Perceptron and SHAP Analysis
This study comprehensively investigates the degradation performance of a Mn-doped Zn2SnO4 photocatalyst based on time-dependent UV-Vis absorption spectra. Before machine learning modelling, the effects of experimental parameters such as UV–Vis measurement wavelength, reaction time, and Mn doping ratio were statistically validated using One-Way Analysis of Variance (ANOVA) and Multiple Linear Regression (MLR) methods. To overcome the limitations of linear models in representing complex physical systems, an optimized Multi-Layer Perceptron (MLP) architecture was developed to capture the system’s nonlinear dynamics with high accuracy. To validate the model’s out-of-sample prediction capability and prevent data leakage potentially arising from spectral data correlation, the “Leave-One-Doping-Level-Out” (LODLO) cross-validation strategy was applied, during which performance metrics of R2 and were recorded. To make the neural network’s decision-making mechanism transparent, a dual-validation explainability framework comprising Shapley Additive Explanations (SHAP) and Permutation Feature Importance analyses was employed. By quantifying the relative contributions of the experimental parameters to the model predictions, this approach revealed that the UV–Vis measurement wavelength was the dominant predictive variable, followed by the Mn doping ratio and reaction time. This study presents a transparent methodology that offers both strong predictive capability and physically grounded data to shed light on the complex interactions in doped semiconductor photocatalysts.Open AccessInterpretable Machine Learning Approach for Photocatalytic Degradation in Mn-Doped Semiconductors Using Multilayer Perceptron and SHAP Analysis
This study comprehensively investigates the degradation performance of a Mn-doped Zn2SnO4 photocatalyst based on time-dependent UV-Vis absorption spectra. Before machine learning modelling, the effects of experimental parameters such as UV–Vis measurement wavelength, reaction time, and Mn doping ratio were statistically validated using One-Way Analysis of Variance (ANOVA) and Multiple Linear Regression (MLR) methods. To overcome the limitations of linear models in representing complex physical systems, an optimized Multi-Layer Perceptron (MLP) architecture was developed to capture the system’s nonlinear dynamics with high accuracy. To validate the model’s out-of-sample prediction capability and prevent data leakage potentially arising from spectral data correlation, the “Leave-One-Doping-Level-Out” (LODLO) cross-validation strategy was applied, during which performance metrics of R2 and were recorded. To make the neural network’s decision-making mechanism transparent, a dual-validation explainability framework comprising Shapley Additive Explanations (SHAP) and Permutation Feature Importance analyses was employed. By quantifying the relative contributions of the experimental parameters to the model predictions, this approach revealed that the UV–Vis measurement wavelength was the dominant predictive variable, followed by the Mn doping ratio and reaction time. This study presents a transparent methodology that offers both strong predictive capability and physically grounded data to shed light on the complex interactions in doped semiconductor photocatalysts. - Open Access
- 12 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 80511-80530
Early and Accurate Diagnosis of Ectopic Pregnancy: A SHAP-Based Explainable Machine Learning Approach
Ectopic pregnancy (EP) is a significant cause of maternal morbidity and mortality, particularly in the first trimester. It occurs when a fertilized ovum implants outside the uterine cavity, with the fallopian tubes being the most common location. Early and accurate diagnosis is critical to prevent complications such as tubal rupture and hemorrhage. Traditional diagnostic methods rely on transvaginal ultrasound (TVUS) and serial β-hCG measurements, but these approaches can sometimes yield inconclusive results. Recent progress in artificial intelligence has demonstrated potential for improving the precision of ectopic pregnancy detection. This study examines the use of artificial intelligence-based classification models to predict pregnancy outcomes using routinely available clinical data. The dataset comprises 310 patient records collected from real-world clinical practice. Several classification approaches were evaluated, with particular emphasis on tree-based ensemble models. Experimental results indicate that the LightGBM DART model outperformed the other evaluated classifiers, achieving the highest overall performance with an accuracy of 0.9161, a macro-AUC of 0.9409, and a balanced accuracy of 0.9080, followed by CatBoost and Histogram-based Gradient Boosting models. To enhance model transparency and clinical interpretability, Shapley Additive Explanations (SHAP) were applied to identify the most influential features contributing to model predictions. These findings suggest that explainable artificial intelligence (XAI)-based predictive models may support early and accurate diagnosis during the early follow-up period (Day 0–12), assist clinicians in timely decision-making, and potentially improve patient management, rather than serving as a standalone tool for initial Day-0 triage.
- 12 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 80511-80530
Open AccessEarly and Accurate Diagnosis of Ectopic Pregnancy: A SHAP-Based Explainable Machine Learning Approach
Ectopic pregnancy (EP) is a significant cause of maternal morbidity and mortality, particularly in the first trimester. It occurs when a fertilized ovum implants outside the uterine cavity, with the fallopian tubes being the most common location. Early and accurate diagnosis is critical to prevent complications such as tubal rupture and hemorrhage. Traditional diagnostic methods rely on transvaginal ultrasound (TVUS) and serial β-hCG measurements, but these approaches can sometimes yield inconclusive results. Recent progress in artificial intelligence has demonstrated potential for improving the precision of ectopic pregnancy detection. This study examines the use of artificial intelligence-based classification models to predict pregnancy outcomes using routinely available clinical data. The dataset comprises 310 patient records collected from real-world clinical practice. Several classification approaches were evaluated, with particular emphasis on tree-based ensemble models. Experimental results indicate that the LightGBM DART model outperformed the other evaluated classifiers, achieving the highest overall performance with an accuracy of 0.9161, a macro-AUC of 0.9409, and a balanced accuracy of 0.9080, followed by CatBoost and Histogram-based Gradient Boosting models. To enhance model transparency and clinical interpretability, Shapley Additive Explanations (SHAP) were applied to identify the most influential features contributing to model predictions. These findings suggest that explainable artificial intelligence (XAI)-based predictive models may support early and accurate diagnosis during the early follow-up period (Day 0–12), assist clinicians in timely decision-making, and potentially improve patient management, rather than serving as a standalone tool for initial Day-0 triage.