{"title":"Physics-informed neural modeling of TiO2 photoanode thickness-dependent electron transport and photovoltaic performance in dye-sensitized solar cells","authors":"K. Sanyasi Naidu, Nalluri Mohan Rao","doi":"10.1007/s10825-026-02658-6","DOIUrl":"10.1007/s10825-026-02658-6","url":null,"abstract":"<div><p>In dye-sensitized solar cells, the thickness of the photoanode has an influence on both the optical absorption and the dye-accessible surface area, as well as on the electron-transport distance and recombination. The work in this study builds upon a physics-informed neural model for thickness-dependent electron transport and photovoltaic performance for 6 compressed titanium dioxide (TiO<sub>2</sub>) photoanodes reported in the literature. The electron state in space is described by a parameterized random-feature neural field, under the steady-state diffusion-generation-recombination equation and proper extraction and zero-flux boundary conditions. As a result of the small size of the dataset, the electron diffusion coefficient, electron lifetime, absorption coefficient, and photon flux are assumed to be fixed effective modeling parameters to avoid non-identifiable inverse parameters. Low-capacity regularized correction maps the differential-equation response of the contact current into the short-circuit current density of the device and auxiliary relationships are regularized to represent open-circuit voltage and fill factor. The neural field showed good agreement with the analytical solution of constant coefficients with <i>L</i><sub>2</sub> relative error of 0.084–0.685%. The efficiency mean absolute percentage error was 12.40% for leave-one-thickness-out validation, while it was 27.76% for linear regression, 26.08% for quadratic regression, 78.42% for natural cubic splines, 19.34% for Gaussian process regression, 26.83% for Bayesian ridge regression, and 20.15% for a conventional ANN. There was a lot of variation in the optimum thickness from the jackknife analysis, which predicted the optimum thickness to be 29.72 µm with 8.70% efficiency in the continuous model. There was limited cross-system transfer seen in two independent runs of the thickness series, with efficiency mean absolute percentage errors of 17.10% and 19.78% for each series, respectively. This limited transferability is an important conclusion: There is no universal performance function for photoanodes, which only depends on their thickness. The framework should thus be considered as a proxy for the different systems, which has been regularized by physics, and as a foundation for future models involving multivariate descriptors of morphology, optical, processing, and electrochemical properties.</p></div>","PeriodicalId":620,"journal":{"name":"Journal of Computational Electronics","volume":"25 6","pages":""},"PeriodicalIF":2.8,"publicationDate":"2026-11-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149339007","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"Calculation of the per-unit-length capacitance matrix in multi-conductor transmission lines based on MoM","authors":"Yixing Gu, Zhisheng Hao","doi":"10.1007/s10825-026-02669-3","DOIUrl":"10.1007/s10825-026-02669-3","url":null,"abstract":"<div><p>Accurate and efficient extraction of per-unit-length capacitance matrices is essential for the modeling of multiconductor transmission lines (MTLs), particularly for nonuniform structures where the cross-sectional geometry varies along the transmission path. Conventional analytical formulas are computationally efficient but usually rely on simplified assumptions, such as sufficiently large conductor spacing, which may lead to inaccurate results. On the other hand, full-domain numerical methods generally require extensive volumetric discretization and become computationally expensive when repeated cross-sectional evaluations are required. To address these challenges, this paper develops an efficient boundary-based electrostatic extraction framework in computational electromagnetics for the per-unit-length capacitance matrix of MTLs. The proposed framework formulates the electrostatic problem using a scalar potential integral equation, in which only the conductor boundaries are discretized. The surface charge distribution is obtained using a pulse-basis, point-matching Method of Moments (MoM) scheme, and the capacitance matrix is directly extracted from the charge-potential relationship. Unlike conventional approaches, the proposed framework does not rely on the wide-spacing assumption and can accurately capture the electrostatic coupling among closely spaced conductors. The proposed method is validated using a representative bounded wave simulator (BWS). The extracted matrices show excellent agreement with Ansys simulations along the entire structure. Furthermore, compared with FEM-based extraction, the proposed method achieves significantly improved computational efficiency by avoiding volumetric meshing of the surrounding dielectric region. The proposed framework provides an accurate and efficient solution for repeated per-unit-length capacitance matrix extraction of practical nonuniform MTL structures.</p></div>","PeriodicalId":620,"journal":{"name":"Journal of Computational Electronics","volume":"25 6","pages":""},"PeriodicalIF":2.8,"publicationDate":"2026-11-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149338066","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"A coupled computational framework for composition-dependent GaAsNP solar cells under indoor illumination","authors":"Volkan Kızıloğlu","doi":"10.1007/s10825-026-02670-w","DOIUrl":"10.1007/s10825-026-02670-w","url":null,"abstract":"<div><p>A coupled one-dimensional computational framework is developed for Si-lattice-matched GaAs<span>(_{1-x-y})</span>N<span>(_x)</span>P<span>(_y)</span> homojunction solar cells under indoor illumination. The framework integrates a band-anticrossing description of the composition-dependent bandgap, measured optical constants, finite-element minority-carrier transport, bulk and depletion-region recombination, surface recombination, antireflection losses, and non-ideal terminal resistances. Mesh convergence, carrier conservation, low-injection validity, and transfer of the formulation to published indoor GaInP measurements are examined before predictions are made for GaAsNP. For six lattice-matched compositions, the preferred nitrogen fraction depends on both spectral power distribution and illuminance. At 1000 lx it ranges from 2.5% for a 6500 K light-emitting diode to 3.5% for a 2700 K source; under the 6500 K source it shifts toward about 4.0% at 100 lx because shunt leakage becomes increasingly important relative to photocurrent. At 1000 lx and 6500 K, the baseline 3.0%-N device reaches 29.55% efficiency and 96.8 <span>(mu)</span>W cm<span>(^{-2})</span> maximum-power density. Material-quality scenarios span 21.31-−33.35%, demonstrating that bulk lifetime and surface recombination dominate the uncertainty, whereas a pure-GaAs mobility upper bound changes efficiency by less than one percentage point. A traced baseline-material sensitivity sequence gives a model ceiling of 33.60% only after ideal terminal resistances are imposed. The calculations identify material quality and shunt suppression, rather than fine composition tuning alone, as the primary requirements for experimentally viable indoor GaAsNP cells.</p></div>","PeriodicalId":620,"journal":{"name":"Journal of Computational Electronics","volume":"25 6","pages":""},"PeriodicalIF":2.8,"publicationDate":"2026-11-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149338063","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Yi Nan, Song Yue, Ran Wang, Minxia Ding, Bohan Liu, Man Li, Shunshuo Cai, Yue Fu, Guangtong Jiang, Mei Xue, Kunpeng Zhang
{"title":"Diffusion model-based phase transition modeling for GST optical weight devices","authors":"Yi Nan, Song Yue, Ran Wang, Minxia Ding, Bohan Liu, Man Li, Shunshuo Cai, Yue Fu, Guangtong Jiang, Mei Xue, Kunpeng Zhang","doi":"10.1007/s10825-026-02663-9","DOIUrl":"10.1007/s10825-026-02663-9","url":null,"abstract":"<div><p>Ge<span>(_2)</span>Sb<span>(_2)</span>Te<span>(_5)</span> (GST) phase-change materials enable non-volatile multi-level optical weighting in silicon photonic integrated circuits, yet conventional multi-physics coupled phase-change dynamics simulations are computationally prohibitive, hindering large-scale pulse parameter sweeping and array calibration. Existing surrogate models are limited to predicting scalar performance metrics and fail to capture spatially resolved phase-change morphologies or the inherent stochasticity of crystallization. This study proposes a conditional diffusion model for rapidly generating phase-change fields in waveguide-integrated GST optical weight devices. We establish an optical–thermal phase-change co-simulation framework and construct datasets covering two-step reset programming and multi-step incremental pulse schemes under fixed device geometry. The model learns the conditional spatial distribution of crystalline domains through denoising score matching. Evaluation shows that the two-step pulse model achieves a mean absolute error (MAE) of 0.144 and an Intersection over Union (IoU) of 0.72, with over two orders of magnitude acceleration compared to iterative physics-based simulation. The proposed method enables efficient pulse parameter exploration, inverse pulse design, batch array calibration, and system-level robustness analysis for large-scale photonic computing systems.</p></div>","PeriodicalId":620,"journal":{"name":"Journal of Computational Electronics","volume":"25 6","pages":""},"PeriodicalIF":2.8,"publicationDate":"2026-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149338207","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Djamila Ziani, Allel Mokaddem, Mohammed Belkheir, Mehdi Rouissat
{"title":"Advanced LCP and LCP/SiO2 polymer substrates for AI-enhanced THz photonic band gap antenna design","authors":"Djamila Ziani, Allel Mokaddem, Mohammed Belkheir, Mehdi Rouissat","doi":"10.1007/s10825-026-02662-w","DOIUrl":"10.1007/s10825-026-02662-w","url":null,"abstract":"<div><p>This study presents a performance enhancement framework of terahertz (THz) circular microstrip antennas operating at 0.3 THz based on liquid crystal polymer (LCP) and LCP/SiO<sub>2</sub> composite substrates integrated with photonic band gap (PBG) structures. By combining low-loss polymer dielectrics, periodically arranged cylindrical air gap PBG lattices, graphene-based conductive layers, and machine learning-assisted performance prediction, the proposed design addresses the persistent trade-off between bandwidth, gain, and radiation efficiency that limits conventional THz antenna configurations. The effective relative permittivity of the LCP/SiO<sub>2</sub> composite was calculated as 3.08 without PBG loading, the proposed LCP and LCP/SiO<sub>2</sub> antennas exhibit impedance bandwidths of 19.58 and 19.51 GHz, realized gains of 4.70 and 4.69 dBi, and radiation efficiencies of approximately 93%, respectively. A detailed parametric analysis of the PBG air gap radius, lattice constant, and substrate thickness demonstrates a strong dependence of the antenna performance on the PBG geometry. The optimized LCP configuration achieves an impedance bandwidth of 72.8 GHz, a realized gain of 7.21 dBi, and a radiation efficiency of 96%, whereas the optimized LCP/SiO<sub>2</sub> configuration provides the highest bandwidth of 104.74 GHz and realized gain of 8.12 dBi with an efficiency of 85%. These results demonstrate that the LCP/SiO<sub>2</sub> composite combined with an optimized PBG structure can substantially offers at trade-off enhancing bandwidth and gain, maintaining acceptable radiation efficiency. To support rapid design evaluation, four machine learning models, random forest, support vector regression, gradient boosting regression, and linear regression models, were trained using CST-generated parametric data to predict the antenna performance. The machine learning models provide highly accurate predictions for bandwidth and efficiency, while gain and reflection coefficient prediction remain more challenging because of their nonlinear dependence on the geometrical parameters. On-body simulations using a three-layer human tissue model confirm compliance with the considered FCC and ICNIRP specific absorption rate limits. The proposed combination of polymer-composite substrates, PBG engineering, graphene-based conductors, and machine learning provides a promising framework for the development and optimization of compact THz antennas for next-generation wireless and sensing applications.</p></div>","PeriodicalId":620,"journal":{"name":"Journal of Computational Electronics","volume":"25 6","pages":""},"PeriodicalIF":2.8,"publicationDate":"2026-10-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149339492","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"A stable, high-order time-stepping scheme for the drift-diffusion model in modern solar cell simulation","authors":"Jun Du, Jun Yan","doi":"10.1007/s10825-026-02656-8","DOIUrl":"10.1007/s10825-026-02656-8","url":null,"abstract":"<div><p>This paper presents a one-dimensional transient drift–diffusion simulator for advanced solar cells, integrating a structure-preserving finite-volume spatial discretization with Scharfetter–Gummel–type fluxes and a high-order, L-stable implicit Runge–Kutta (Radau IIA) temporal integrator. The scheme ensures local charge conservation, handles sharp material interfaces, achieves second-order spatial and fifth-order temporal convergence, and exhibits excellent computational efficiency and numerical stability. Its accuracy is verified against the classical depletion approximation in <i>p</i>–<i>n</i> junction and validated against four independent software implementations for an organic photovoltaic device under a common protocol. The framework’s extensibility is demonstrated by incorporating exciton kinetics in organic solar cells, capturing multi-timescale dynamics, and by modeling mobile ions in perovskite solar cells, reproducing characteristic <span>(textit{J})</span>–<span>(textit{V})</span> hysteresis without empirical parameters. This work provides a robust, high-order numerical foundation for simulating coupled charge, exciton, and ion transport in next-generation photovoltaic devices.</p></div>","PeriodicalId":620,"journal":{"name":"Journal of Computational Electronics","volume":"25 6","pages":""},"PeriodicalIF":2.8,"publicationDate":"2026-10-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149339066","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
D. Joseph Jeyakumar, V. Rajkumar, D. Shanthi Chelliah, R. Amutha, C. Karthiga
{"title":"Electrothermal–RF co-design of segmented-field-plate MgZnO/ZnO HEMTs","authors":"D. Joseph Jeyakumar, V. Rajkumar, D. Shanthi Chelliah, R. Amutha, C. Karthiga","doi":"10.1007/s10825-026-02660-y","DOIUrl":"10.1007/s10825-026-02660-y","url":null,"abstract":"<div><p>Enhancement-mode MgZnO/ZnO high-electron-mobility transistors are promising candidates for oxide-based RF and power-oriented electronics, but their design requires simultaneous control of electric-field crowding, RF parasitics, electrothermal loading, trap-induced dispersion, and fabrication variability. Present MgZnO/ZnO studies usually examine these aspects distinctly, whereas field-plate and robust-optimization notions established mainly for GaN-based HEMTs cannot be directly transferred without material-specific recalibration. This work presents a literature-calibrated, physics-constrained multi-fidelity computational context for the co-design of an asymmetric segmented-field-plate MgZnO/ZnO HEMT. The work combines reduced-order device relations, a two-dimensional finite-difference electrostatic benchmark, an electrothermal correction, trap-dispersive small-signal equivalent-circuit modeling, multi-objective robust selection, and local Sobol sensitivity analysis. All numerical values reported in this study are outputs of the stated theoretical-computational benchmark; <span>(V_{{B,{text{eq}}}})</span> is a model-derived field-limited comparative index. Within the stated benchmark, the robust-knee design reduces the local gate-edge electrostatic metric by approximately 36% compared with the unplated baseline and gives <span>(V_{{B,{text{eq}}}} approx 1.05,{text{kV}})</span>, <span>(f_{T} approx 118,{text{GHz}})</span>, <span>(f_{{{text{max}}}} approx 326,{text{GHz}})</span>, <span>(NF_{{{text{min}}}} approx 1.48,{text{dB}})</span>, and <span>(T_{{{text{max}}}} approx 373,{text{K}})</span>. The dispersive small-signal model decreases the complex broadband S-parameter error relative to a constant-element model in the synthetic high-fidelity scale. The results indicate that segmented field shaping, dielectric coupling, copper-aware parasitic reduction, trap dispersion, and uncertainty-aware design must be treated together. The study delivers a reproducible computational route for future experimental validation by matched baseline, continuous-field-plate, and segmented-field-plate MgZnO/ZnO devices.</p></div>","PeriodicalId":620,"journal":{"name":"Journal of Computational Electronics","volume":"25 6","pages":""},"PeriodicalIF":2.8,"publicationDate":"2026-10-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149235030","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"A direct and physically transparent iterative framework for solving the four-diode photovoltaic model: efficient parameter extraction and accurate prediction performance","authors":"Yueran Gao","doi":"10.1007/s10825-026-02651-z","DOIUrl":"10.1007/s10825-026-02651-z","url":null,"abstract":"<div><p>The four-diode model (FDM) enables high-accuracy photovoltaic characterization but poses numerical challenges in solving its implicit current–voltage relationship. This paper modifies the state-of-the-art Iterative Lambert <i>W</i> Function (ILWF) methodology to propose a direct and physically transparent iterative framework. First, it replaces the dimensionless iteration variable in ILWF with a physically meaningful initial current guess <i>I</i><sub>init</sub>, enabling principled initialization. Second, it adapts concise and explicit iterative expressions that map <i>I</i><sub>init</sub> directly to the estimated current <i>I</i><sub>cal</sub>, establishing a clear functional mapping <i>I</i><sub>cal</sub>=<i>F</i>(<i>I</i><sub>init</sub>) without implicit dependencies. Third, it designs two goal-oriented termination criteria: Criterion 1 for parameter extraction using root-mean-square error (RMSE) monitoring, which requires measured current data; and Criterion 2 for parameter extraction and current prediction using a fixed three-iteration maximum-absolute-difference selection strategy. Validated on five photovoltaic devices under diverse conditions and measurement noise, the proposed framework outperforms ILWF and six recent FDM solvers. For parameter extraction, it reduces RMSE by 3.07–41.3% and average computation time by up to 75.4%. Improvements are statistically significant via one-tailed Wilcoxon signed-rank tests across 128 independent configurations; 18 non-significant cases (14.1%) arise from deterministic initialization–solver interactions rather than stochastic variability. Under noise, win rates against ILWF reach 74.5–92.8%. For current prediction on xSi12922 and mSi0188 modules under varying operating conditions, the framework exhibits superior extrapolation capability, and this advantage persists after enhancing physical consistency via the proposed diode ideality factor correction. This work provides a rigorously validated, efficient, and interpretable paradigm for solving the FDM, advancing high-fidelity photovoltaic modeling.</p></div>","PeriodicalId":620,"journal":{"name":"Journal of Computational Electronics","volume":"25 6","pages":""},"PeriodicalIF":2.8,"publicationDate":"2026-10-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149235073","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Ritika Singh, Aditya Singh, Guru Pyari Jangid, Ashok Jangid
{"title":"Inverse modeling of tunable ferrite RMSAs addressing non-uniqueness using a multivalued neural network","authors":"Ritika Singh, Aditya Singh, Guru Pyari Jangid, Ashok Jangid","doi":"10.1007/s10825-026-02661-x","DOIUrl":"10.1007/s10825-026-02661-x","url":null,"abstract":"<div><p>Inverse electromagnetic (EM) design problems often exhibit non-unique mappings, where multiple geometrical configurations can produce identical electrical responses. Conventional artificial neural networks trained using standard loss functions typically learn an averaged solution, limiting their ability to accurately model such multivalued relationships. This paper proposes a Multivalued Neural Network (MVNN) framework for inverse modeling of ferrite-based tunable rectangular microstrip antennas (RMSAs). The proposed architecture employs a feedforward neural network with a multibranch output structure, enabling the simultaneous prediction of multiple candidate geometrical solutions for a given set of electrical specifications. A custom multivalued loss function is introduced to guide the training process by selecting the prediction branch with the minimum error for each sample, allowing the network to learn multiple feasible inverse mappings without penalising alternative solutions. The model is trained using datasets generated from EM simulations based on the spectral-domain immittance approach (SDIA). The influence of the number of output branches <i>M</i> is systematically investigated. Experimental results show that the optimal configuration (<i>M</i> = 6) achieves approximately 42% reduction in mean squared error compared to a conventional single-output ANN while maintaining high prediction accuracy with <i>R</i><sup>2</sup> > 0.999. The proposed MVNN framework provides an effective and computationally efficient approach for solving multivalued inverse EM design problems.</p></div>","PeriodicalId":620,"journal":{"name":"Journal of Computational Electronics","volume":"25 6","pages":""},"PeriodicalIF":2.8,"publicationDate":"2026-10-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149234774","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"Handling class imbalance in high-throughput materials’ simulation data: a cross-validated ablation benchmark on an ab initio transport database","authors":"Johaimen M. Omar, Kemal Akyol","doi":"10.1007/s10825-026-02652-y","DOIUrl":"10.1007/s10825-026-02652-y","url":null,"abstract":"<div><p>Materials databases built from high-throughput simulations often contain rare property classes, so classifiers trained on them face severe class imbalance. We benchmarked imbalance-handling strategies on 9,036 compounds from one such database, the BoltzTraP (Boltzmann Transport Properties) ab initio transport database, of which 445 (4.92%) have a hole effective mass <span>({m}_{p}^{*})</span> < 1.0 <span>({m}_{e})</span> (imbalance ratio 19.3:1). The inputs were three n-type descriptors: the electron effective mass, the n-type power factor, and the n-type Seebeck coefficient. Because the inputs and the label come from the same calculation, the classifier is not a screening tool; the task serves as a controlled testbed in which every label is known. All resampling was confined to training folds, as standard practice requires. Relative to no resampling, the borderline synthetic minority oversampling technique (Borderline-SMOTE) with eXtreme Gradient Boosting (XGBoost) raised minority-class recall from 0.411 to 0.798 and gave the highest F2-score (0.608). It also exceeded random undersampling on F2-score, F1-score, and Matthews correlation coefficient (MCC) (0.608 vs. 0.550; 0.448 vs. 0.355; 0.461 vs. 0.391). The no-resampling baseline, however, had the highest F1-score (0.520) and MCC (0.523), so the preferred strategy depends on whether recall or precision matters more. With fivefold, the fold-level Wilcoxon tests could not reach significance. The n-type descriptors carried information about the hole-mass class, an expected result because both carrier masses derive from the same band structure.</p></div>","PeriodicalId":620,"journal":{"name":"Journal of Computational Electronics","volume":"25 6","pages":""},"PeriodicalIF":2.8,"publicationDate":"2026-10-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149234496","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}