Machine-Learning-Based Compact Modeling for Sub-3-nm-Node Emerging Transistors

TFETs (Tunnel Field-Effects Transistors) are assumed to be ultra-low-power devices because they have steep subthreshold swing and a more favourable scaling characteristic. Their electrical characteristics, however, cannot be adequately simulated using extensive Technology Computer-Aided Design (TCAD) simulations, which are computationally intensive and time-consuming to analyze. In an attempt to overcome this bottleneck, a proposal for a machine learning (ML)-based surrogate modeling framework is provided where the prediction of the critical TFET properties of transfer curves, Ion, Ioff, threshold voltage (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$V_{\text{th}}$</tex>), and subthreshold swing (SS) is endeavored with the help of the Random Forest Regression (RFR) model. The data collected based on TCAD simulations and literature reports was processed, trained, and tested on various ML models such as the Artificial Neural Networks (ANN), Support Vector Regression (SVR) and RFR. The calculated R2 of the proposed RFR model was better than ANN and SVR, with a much smaller MSE (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{3. 2 1} \times \mathbf{1 0}^{-\mathbf{7}}$</tex>) and an RMSE (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1. 7 9} \times \mathbf{1 0}^{-\mathbf{4}}$</tex>), showing that the model is robust and exhibits better generalization within the conditions of the devices used. Also, the ranking of intrinsic feature importance offered by RFR has a better interpretation rate and complies well with the previous findings on TCAD. These findings place RFR as a sound and efficient substitute for TFET modelling, with lower computational expense and intact physical stability, and thus add another facet and a practical approach to future device design optimization, assisted by a novel and powerful ML

We propose a physics-based analytical model that accurately captures the effects of buffer traps on dc characteristics of gallium nitride (GaN)-based high-electron-mobility transistors (HEMTs). The model is then semi-analytically extended to additionally include the transient behavior. Analytical formulations for the shift in the threshold voltage <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${(}{V}_{\text {OFF}}{)}$ </tex-math></inline-formula> and two-dimensional electron gas (2-DEG) density due to the presence of buffer traps in the steady state are presented. In pulsed operation, technology computer-aided design (TCAD) simulations indicate that a time-dependent negative potential (NP) is developed under the gate, resulting in a modified <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${V}_{\text {OFF}}$ </tex-math></inline-formula> and current collapse (CC). An expression for the modified <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${V}_{\text {off}}$ </tex-math></inline-formula> helps capture the pulsed current–voltage characteristics. The model captures the dependence of bias, time, temperature, trap concentration, capture cross section area, and activation energy of traps on the steady-state and transient characteristics. The model is implemented in Verilog-A in an existing compact model framework using a diode and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">RC</i> sub-circuit and validated using measured data and TCAD simulations. The modeling results are in excellent agreement with the experimental data and TCAD simulations. Since the model is physics-based, it requires fewer number of parameters compared to that in the existing models.

The artificial neural network (ANN)-based compact model has significant advantages over physics-based standard compact models such as BSIM-CMG because it can achieve higher accuracy over a wide range of geometric parameters. This makes it particularly suitable for design space exploration and optimization. However, the ANN-based compact model using only one set of model parameters (global-ANN) requires larger model sizes to achieve wider coverage and higher accuracy in order to capture the unpredictable nonlinearities of emerging devices. This results in reduced simulation speed and a trade-off between simulation accuracy, model coverage, and simulation speed makes it difficult to utilize ANN-based compact models in a variety of ways. To solve this problem, we propose the first ANN-based compact modeling flow using a binning method (binning-ANN) and we address the training requirements and data sparsity issues that may occur due to the binning method in ANNs. In addition, we develop a bin size optimization guideline for the binning-ANN. As a result, the binning-ANN not only has higher accuracy, but also much better expandability than existing methods.

A machine learning (ML) method was used to optimize the trap distribution of the charge trap nitride (CTN) to simultaneously improve its performance/reliability (P/R) characteristics, which are tradeoffs in 3-D NAND flash memories. Using an artificial neural network (ANN), we modeled the relationship between trap distributions and P/R characteristics. The ANN was trained using a large experimentally calibrated technology computer-aided design (TCAD) simulation dataset. The gradient descent method was adapted to optimize the trap distribution, achieving the best P/R characteristics based on the well trained ANN. Eventually, we found the best trap profile distributed in both space and energy. In particular, the energetic trap distribution had a larger impact on the P/R characteristics than that of the spatial trap distribution. Furthermore, in terms of the P/R characteristics, it was generally preferable to increase all inputs of the energetic trap distribution. However, the acceptor-like trap energy level (ETA) and its standard deviation (sigma EA) caused a tradeoff between P/R characteristics; therefore, ML was used to determine their optimal points. The proposed ML method allows the optimization of trap distribution to obtain the best P/R characteristics rapidly and quantitatively. Our findings could be used as a guideline for determining the physical properties of CTN in 3-D NAND flash cells.

Compact Modeling of Physical Mechanisms Present in the Weak Breakdown Regime of Advanced Si and SiGe Bipolar Transistors

This study intended to use optimized artificial neural network (ANN) for the design of pure cohesive slopes (by means of considering sufficient safety factors (SF) of stability). A total of 630 finite element limit equilibrium analysis were completed to provide datasets of training and testing for preparation of the ANN-based predictive models. The database consisted of 504 training datasets (80% of the database) and 126 testing datasets (20% of the database). To construct an optimized ANN structure, variables of ANN algorithm such as a number of neurons in hidden layer and the number of hidden layers were changed (e.g. with a series of trial and error process). The input parameters that used in the analysis include setback distance ratio (b/B), the undrained shear strength of the cohesive soil (Cu), applied stresses on the slope (Fy) and slope angle (β) while the final output was the value of SF. The predicted results for datasets from ANN models were assessed based on several well-known statistical indices namely, R2, RMSE and VAF. Also, in order to evaluate the performance of predicted network two ranking systems of (1) colour intensity rating (CIR) and (2) total ranking method (TRM), i.e. based on the result of statistical indices, were utilized. The optimal architecture of 4 × 6 × 1 was found for the ANN structure. After the performance of the optimized ANN model, the results are provided as a simple tansig formula which can be used for further application of engineering purposes. In the optimized ANN network and based on R2, RMSE and VAF, values of (0.999, 0.0350 and 99.9978) and (0.999, 0.0379 and 99.998) were found, respectively, for both of the normalized training and testing datasets. This proves the excellent performance of the proposed ANN model in estimating the SF of the slope. As the final outputs and to show the reliability of the proposed ANN formula, a series of optimized ANN design solution charts are compared to a similar example of measured charts.

In this study, SiGe channel nanosheet field-effect transistor (NSFET) with a novel channel release stopping layer (CRSL), namely CRSL-SiGe NSFET, is proposed to addresses the issue of sub-fin recess during channel release in conventional SiGe (C-SiGe) NSFET. Through a three-dimensional technology computer-aided design (TCAD) simulations, we systematically compare the DC and AC electrical characteristics of C-SiGe and CRSL-SiGe NSFETs. The results show that the proposed CRSL-SiGe NSFET achieves a 16% reduction in gate capacitance (C<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">GG</sub>) while maintaining a comparable DC performance to C-SiGe NSFET. Additionally, an obvious improvement in gain of the inverter is achieved under various power. Moreover, the oscillation frequency of ring oscillators is increased and their delay is reduced under various power. These results indicate that CRSL-SiGe NSFET is a promising candidate for suppressing process variations and improving circuit performance in the advanced technology node.

Two-dimensional (2D) transition metal dichalcogenides (TMDs) have emerged as a first-rate candidate for post-silicon channels in field-effect transistors (FETs), particularly in the context of monolithic three-dimensional (M3D) integration. Despite the recent synthetic advances including industry-compatible wafer-scale production, they have predominantly been prepared on the ultra-flat, non-processed surfaces, limiting device architectures to outdated planar FETs and geometrically restricting their versatile applications. Integrating 2D semiconductors into 3D device architectures currently presents two significant challenges: i) stackablity, and ii) scalablility. While 2D material-based FETs have been primarily limited to traditional planar devices structures, their applications to vertically integrated 3D devices have been hindered by low-fidelity transfer method which is certainly challenging in conformally and controllably transferring atomically thin materials onto steep sidewalls. Here, we report a directly growth-and-fabrication approach enabling the dimensional transition from 2D planar FETs to a 3D vertical-channel FETs (VCFETs) through the atomically conformal direct growth of monolayer (ML) MoS2 onto the pre-fabricated VCFET structures (Fig. 1A). Leveraging the inherently anisotropic in-plane growth kinetics and precisely controlled supersaturation based on metal-organic chemical vapor deposition, we achieved 100% step coverage even with sub-1-nm channel thickness without compromised crystallinity along an exceptionally high aspect ratio exceeding 15,000 (Fig. 1B). We proved that monolayer MoS2 grown along the SiO2 vertical sidewall is a single crystal by comparing the angles of the selected-area electron diffraction (SAED) pattern in transmission electron microscope (TEM) at the three vertices of MoS2 (Fig. 1C). Fig. 1D shows a cross-sectional scanning transmission electron microscope (STEM) image of the VCFET with and spatial distribution of each element was also identified by energy dispersive X-ray spectroscopy (EDS) mapping images, and it convincingly suggests that atomically conformal monolayer MoS2 was indeed integrated without discontinuity and void even along a steep trench edge (Fig. 1D). Fig. 1E illustrates the unit device schematic of the monolayer MoS2 VCFET. The experimental and simulated transfer characteristic (I DS-V GS) of VCFET with vertical monolayer MoS2 well-coincide each other, showing the subthreshold swing (SS) of 77mV·dec−1 and the ON/OFF ratio of 108 under a drain-to-source bias (V DS) of 1V (Fig. 1F). Output characteristic (I DS-V DS) indicates the change in conductivity of the MoS2 channel by regulating the V G at steps of 1V from −2V to 5V (Fig. 1G). To evaluate device-to-device variation and statistic distribution of their performance, electrical characterization of the 7×7 VCFETs array was performed and their I DS-V GS curves under V DS = +1V exhibited the excellent reproducibility (Fig. 1H). Fig. 1I shows low OFF-state current density of 10−13A∙μm−1 at all devices in our monolayer MoS2 VCFET array, which is 10 times lower than the International Roadmap for Devices and Systems (IRDS) 2028 low-standby power device requirement (~10−12A∙μm−1). Sentaurus technology computer-aided design (TCAD) simulations for monolayer MoS2 VCFET were performed to investigate the electric field distribution of ON-state (V GS = +5V) and OFF-state (V GS = −2V), respectively (Fig. 1J-K). To demonstrate how electron concentration changes in monolayer MoS2 at the sidewall control the switching operation of the transistor, we fabricated a multi-gate VCFET (Fig. 1L). The additional terminals, a modulation gate (MG) and a back gate (BG), facilitate to form an effective gate length near the control gate (CG), enabling a switching behavior of the transistor. The MG screens the electric field toward the top of the trench from the CG, maintaining the electron concentration of the monolayer MoS2 on the top region of the trench and restricting switching operation to monolayer MoS2 on the vertical sidewall. I DS-V CG curve was obtained by applying the CG bias (V CG) and the BG bias (V BG) of +30V with the floated MG bias (V MG) (Fig. 1M). The fabricated multi-gate VCFET further validated the superior sidewall-gate controllability of the ultrathin monolayer MoS2 channel on the vertical sidewall, with aid of TCAD simulation. In the ON-state (V CG = +2.5V), the MG screens electric field generated by CG, allowing the electric field of CG to accumulate electrons effectively along the vertical monolayer MoS2 sidewall, as shown in Fig. 1N. in the OFF-state (V CG = −2.5V), the electric field contour plot reveals that the electric field from the CG does not reach the top region of trench due the MG, while forming a strong electric field near the CG, leading to deplete the MoS2 channel (Fig. 1O). Our work establishes a tailored pathway for the bespoke monolithic integration of ML semiconductors, positioning them as viable channels for high-density, high-performance computational device. Our integration strategy is undisputed pathway for M3D integration and usher in a new era of atomic-level fabric. Figure 1

Providing the real-time working system able to localize the dangerous contaminant source is one of the main challenges of the cities emergency response groups. Unfortunately, all proposed up to now frameworks capable of estimating the contamination source localization based on recorded by the sensors network the substance concentrations are not able to work in real-time. The reason is the significant computational time required by the applied dispersion models. In such reconstruction systems, the parameters of the given dispersion model are sampled to fit the model output to the registrations; thus, the dispersion model is run tens of thousands of times. In this paper, we test the possibility of training an artificial neural network (ANN) to effectively simulate the atmospheric toxin transport in the highly urbanized area. The use of a fast neural network in place of computationally costly dispersion models in systems localizing the source of contamination can enable its fast response time. As a training domain, we have chosen the center of London, as it was used in the DAPPLE field experiment. The training dataset is generated by the Quick Urban & Industrial Complex (QUIC) Dispersion Modeling System. To achieve the ANN capable of estimating the contaminant concentration, we tested various ANN structures, i.e., numbers of ANN layers, neurons, and activation functions. The performed tests confirm that trained ANN has the potential to replace the dispersion model in the contaminant source localization systems.

Finding the optimal parameters in laser processing applications can be time-consuming given the large parameter space and various sources of error. This problem is exacerbated by day-to-day variation in laser beam characteristics and a large variety of materials that need to be processed. The ideal laser processing system should be “smart,” meaning that it can sense changes in the environment, make proper adjustment, and predict parameters for new materials. As a step toward this goal, we propose a method to efficiently extract the areas of a large number of laser-induced damages in situ using an automated data acquisition system that can control laser parameters, motorized stage movement, image capturing/processing, and feature extraction. The damage areas are extracted and compared with direct measurements. Damage areas are fed into an artificial neural network (ANN) for prediction. Various ANN structures and training functions are tested to create the optimal ANN for prediction. ANN predictions were found to be capable enough to accurately model and optimize the laser processing parameters that were investigated. With the capability of collecting a large amount of usable data in a short period of time, this acquisition system can be used to train sophisticated ANNs for complicated tasks such as quality control and failure prediction.

Complex nonlinear dependence of ultra-scaled transistor performance on its channel geometry and source/drain (S/D) doping profile bring obstacles in the advanced technology path-finding and optimization. A machine learning-based multi-objective optimization (MOO) workflow is proposed to optimize the sub-3-nm node gate-all-around (GAA) three-layer-stacked nanosheet transistors (NSFETs) accounting for the key performance knob of channel geometry and S/D doping profile. The artificial neural network (ANN) is trained to learn the compact current–voltage ( <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${I}$ </tex-math></inline-formula> – <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${V}$ </tex-math></inline-formula> ) relationship of NSFETs from 3-D technology computer-aided design (TCAD) simulation results. Based on the artificial neural network (ANN) model, MOO between threshold swing, on–off ratio, and ON-state current of NSFETs is performed with adaptive weighted sum theory. The proposed workflow efficiently suggests an optimized design window of channel geometry and doping profile of NSFETs. The proposed devices satisfy the 2025 International Roadmap for Devices and Systems (IRDSs) target in terms of electrical characteristics for digital circuits.

In this paper, we study a new model that represents the symmetric connection between capacitance–voltage and Schottky diode. This model has a symmetrical shape towards the horizontal direction. In recent times, works conducted on artificial neural network structure, which is one of the greatest actual artificial intelligence apparatuses used in various fields, stated that artificial neural networks are apparatuses that proposal very high forecast performance by the side of conventional structures. In the current investigation, an artificial neural network structure has been generated to guess the capacitance voltage productions of the Schottky diode with organic polymer edge, contingent on the frequency with a symmetrical shape. Of the dataset, 130 were grouped for training, 28 for validation, and 28 for testing. In order to evaluate the effect of the number of neurons on the prediction accuracy, three different models with different neuron numbers have been developed. This study, in which an artificial neural network model, although well-trained, could not predict the output values correctly, is a first in the literature. With this aspect, the study can be considered as a pioneering study that brings a novelty to the literature.

Several researchers have investigated the relationships among different physical attributes of the fruits. For proper design and operation of grading systems, important relationships among the mass and other properties of fruits such as length, width, thickness, arithmetic mean diameter, geometric mean diameter, sphericity, surface area, volume, projected area, shape index, aspect ratio and elongations must be known. Recent researches have focused on artificial neural network (ANN) approaches to predict hard-to-find attributes of the fruits from easily-determined and readily available values. In this study, Modular Neural Network (MNN) and Radial Basis Neural Network (RBNN) structures of Artificial Neural Network (ANN) were employed to predict walnut mass from the physical attributes of the walnuts. Root mean square errors (RMSE) of MNN structure ranged from 0.60 to 0.89, while RMSE of RBNN structure were found to be very low (0.0002) in all of walnut varieties. These results showed that RBNN structures of Artificial Neural Network could potentially be used to estimate mass of walnuts and various physical attributes of walnuts were sufficient to predict the mass characteristics of a walnut.

An artificial neural network (ANN) is a network designed with adaptation to a computer system. The developed computer system will perform functions oriented to the way the brain works (neuron concept). This study is an extension to the study of the suitability of ANN to be applied in numbers of applications, especially in the field of medical engineering. ANN has been widely being used in medicine, ANN is widely applied in education, research, and even decision making. In this study, ANN will be trained for pre-testing to predict the cardiac abnormalities symptom based on selected reference parameters. This reference parameter is better known as the input parameter to the ANN to detect cardiac abnormalities, among which are the of the height of peak/wave (amplitude) and time occurrence of peak/wave (duration of time) extracted from the electrocardiogram (ECG) signal. A complete ECG complex contains a P peak, a QRS wave, and a T peak. For each P peak, QRS wave, and T peak, amplitude height and duration will be measured to serve as input parameters. This makes six parameters defined as inputs to the ANN. This study has used a Multilayer Perceptron (MLP) network as ANN structure by being trained using three different training algorithms namely Backpropagation (BP), Lavenberg Marquardt (LM) and Bayesian Regularization (BR). At the end of the study, it showed the MLP network which by BR training algorithm gave the highest accuracy prediction (94.04%), followed by LM (92.95%) and BP (88.77%). In this study all MLP networks were activated using the Tansig activation function.

Wide bandgap (WBG) semiconductors such as 4H-SiC MOSFETs are key enablers for next-generation power electronics due to their superior efficiency and high-temperature capability. However, their electrical performance is strongly affected by process variations, and conventional Technology Computer-Aided Design (TCAD) simulations are computationally demanding and difficult to scale. This work presents a novel explainable machine learning framework that integrates artificial neural networks (ANNs) with explainable artificial intelligence (XAI) to enable accurate and interpretable device modeling. The ANN is trained on comprehensive TCAD-generated datasets covering a broad range of structural and doping parameters, while SHapley Additive exPlanations (SHAP) are employed to quantify the influence of each design parameter on electrical characteristics. The proposed model achieves a Pearson correlation coefficient exceeding 0.99 for on-state current prediction, and SHAP analysis reveals physically consistent trends such as the inverse dependence of drain current on oxide thickness and channel length. This study is among the first to apply XAI for interpreting design-performance correlations in SiC MOSFETs, establishing a transparent and data-driven framework for device understanding and optimization. The methodology can be readily extended to other semiconductor technologies where both modeling accuracy and interpretability are essential.