Knowledge-based Neural Network Research Articles

A physics knowledge-based neural network method for three-dimensional fracture mechanics of attachment lugs

The lug type joint serves as a critical connecting structure in aerospace vehicles, allowing for convenient assembly and disassembly, and transfer the load during service. However, cracks often initiate at the stress concentration location near the lugs hole, posing a significant threat to the safety of the aircraft structure. Therefore, evaluating the stress intensity factors for complex lug structures is of great significance. In this study, a physics knowledge-based neural network method of calculating the stress intensity factors of attachment lugs is proposed. Three types of cracked lug structures are analyzed: through-thickness crack in straight lug, quarter elliptical corner crack in straight lug and quarter elliptical corner crack in tapered lug. Various complex influencing factors in actual situations are also considered, such as the impact of the interference fit between the pin and the lug, as well as different loading directions and crack positions of tapered lugs. The stress intensity factor dataset generated by 3D finite element method, and then the neural network with its powerful learning ability and prior physical knowledge are used to achieve accurate fitting of the data. The results demonstrate that incorporating physical knowledge features significantly improves the model performance of the neural network. This method can be extended to analyze crack in structures with arbitrary geometry and complex loading conditions.

Modeling and signal integrity analysis of silicon interposer channels based on MTL and KBNN

In this paper, the equivalent circuit model of interconnect channels in silicon interposer is developed for three-dimensional integrated circuit (3-D IC) based on the theory of multi-conductor transmission line (MTL). The deep neural network (DNN) is employed to learn the nonlinear mapping relationship between the geometric parameters and the distributed parameters, and it is demonstrated that the DNN has a high prediction accuracy. In the implementation, a coarse model is firstly developed by solving the scattering parameters of the model using the chain parameter matrix. The model is optimized using a knowledge-based neural network (KBNN), and a fine model is then established. The numerical examples demonstrate that the optimized model can achieve high accuracy. Finally, the inner contour of the worst-case eye diagrams of the channel is generated using the peak distortion algorithm (PDA), with the worst eye width and height obtained.

A Scalable Knowledge-Based Neural Network Model for GaN HEMTs With Accurate Trapping and Self-Heating Effects Characterization

In this article, a scalable knowledge-based neural network (KBNN) large-signal model of gallium nitride (GaN) high-electron-mobility transistors (HEMTs) with accurate trapping and self-heating effects characterization is developed. An improved empirical drain current model is proposed and added to the neural network as prior knowledge, thereby establishing the drain current model, including self-heating effect. A new empirical equation is proposed to model the buffer-related trapping effect more accurately. Taking Angelov capacitance models as prior knowledge, the KBNN capacitance models are completed. Moreover, the scaling characteristics of the proposed KBNN model are studied. The developed model has been fully verified by different sizes of GaN HEMTs. Good agreement between the model simulation results and the measurement data, including 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> ), <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$S$</tex-math> </inline-formula> -parameters, power characteristics, and load-pull data, confirms the effectiveness of the proposed model. The proposed scalable KBNN model is fast and accurate and would be useful for accurate large-signal modeling of large gate periphery GaN HEMTs for high-power radio frequency (RF) applications.

Modified Knowledge-Based Neural Networks Using Control Variates for the Fast Uncertainty Quantification of On-Chip MWCNT Interconnects

In this article, a modified knowledge-based artificial neural network (KBANN) metamodel is developed for the efficient uncertainty quantification of on-chip multiwalled carbon nanotube (MWCNT) interconnects. The proposed KBANN metamodel utilizes the notion of control variates to enable much faster training than what is possible with standard KBANNs. Importantly, techniques to calculate the optimal value of the control variates in an a priori manner without augmenting the training dataset have been developed in this article. Furthermore, techniques to exploit the control variates depending on whether one or multiple low-fidelity models of the MWCNT interconnects are available have also been developed in this article. The benefits of the proposed KBANN metamodel using control variates over standard KBANN metamodels have been validated using multiple MWCNT interconnect examples spanning multiple technology nodes.

Knowledge‐based neural network with Bayesian optimization for efficient nonlinear RF device modeling

AbstractThis paper proposes a novel approach for modeling the nonlinear behavior of microwave devices, with a particular focus on addressing the challenging problem of modeling compound semiconductor transistors under large‐signal excitation. The approach utilizes a knowledge‐based neural network (KBNN) to construct a frequency‐domain behavioral model of the transistor, which maps incident and scattered waves from a coarse model to a fine model in cases where the coarse model fails to match the behavior of a new device. This neural network‐based method aims to achieve better alignment between the extracted model parameters obtained from the coarse model and the behavior of the fine model. To improve mapping efficiency and accuracy, Bayesian optimization techniques are employed to automatically adjust the hyper parameters in the KBNN within a custom‐defined hyper parameter space. The proposed method uses simulation and measurement of a 0.25 um GaN HEMT device operating at 8 GHz, with VGS at −2 V and VDS at 18 V. Furthermore, the proposed model exhibits good interpolation capability at different input power levels, indicating its broad applicability in the design of high‐speed device models. Specifically, the accuracy of the simulation data and test data reached −53.83 and −51.68 dB, respectively, with the test data used in model training being less than 7.5% of the simulation data.

Knowledge-Based Neural Network for Multiphysical Field Modeling

Knowledge-Based Neural Network for Multiphysical Field Modeling

Knowledge-based neural network for surface roughness prediction of ball-end milling

Theoretical modeling for surface roughness of ball-end milling generally results in low prediction accuracy due to the lack of machining information, excessive computational costs, and uncertainty of manual adjustment parameters. Besides, data-driven methods are limited in practical applications due to the lack of theoretical support or ‘black-box problems’ of neural network. To deal with these problems as mentioned above, this paper presents a coupling model for surface roughness prediction by combining the theoretical model of surface roughness and the data-driven approach. The core process is the secondary modeling with regard to the predictive result of theoretical model, which is realized by utilizing the knowledge-based neural networks with radial basis functions (KBaNN_RBF). Firstly, a theoretical model is established for the surface roughness prediction of ball-end milling by using the tool parameters and machining parameters. The output of the theoretical model is embedded into KBaNN_RBF as a priori knowledge. Secondly, low-dimensional features are extracted from the monitoring signals and adopted as the input of KBaNN_RBF for secondary modeling. And, the network structure is optimized by utilizing particle swarm optimization (PSO). Finally, ball-end milling experiment is carried out to verify the effectiveness of the presented KBaNN_RBF. Experimental results show that KBaNN_RBF is highly effective in improving the prediction accuracy and compressing the confidence interval (CI). Moreover, the comparison results show that the effectiveness of KBaNN_RBF is far superior to the traditional methods, such as back propagation neural network (BPNN) and support vector machine (SVM). This paper provides theoretical guidance for accurate monitoring of surface roughness in real industrial settings.

Knowledge-based neural network SPICE modeling for MOSFETs and its application on 2D material field-effect transistors

Knowledge-based neural network SPICE modeling for MOSFETs and its application on 2D material field-effect transistors

Artificial Neural Networks for Microwave Computer-Aided Design: The State of the Art

This article presents an overview of artificial neural network (ANN) techniques for a microwave computer-aided design (CAD). ANN-based techniques are becoming useful for performing forward/inverse modeling for active/passive components to enhance a circuit design. With measured or simulated data of microwave devices, ANNs can be trained to learn relevant microwave relationships, which are, otherwise, computationally expensive or for which efficient analytical formulas are not available. Fundamental concepts of the ANN structure and training, such as feedforward neural networks (FFNNs), recurrent neural networks (RNNs)/dynamic neural networks (DNNs)/time-delay neural networks (TDNNs), deep neural networks, and neural network training and extrapolation, are described. Knowledge-based neural networks (KBNNs) are described for improving the accuracy and reliability of modeling and design optimization. Various advanced ANN techniques, such as neuro-transfer function (neuro-TF) modeling, neural network inverse modeling, and deep neural network modeling, are discussed. The existing and emerging applications of ANN in microwave CAD are identified, such as electromagnetic (EM)/multiphysics modeling, modeling of nonlinear circuits and transistors, filter design, very large-scale integration (VLSI) interconnects, oscillator, transmitter and receiver modeling, and CAD applications in such as gallium nitride (GaN) high electron-mobility transistor (HEMT), wireless power transfer (WPT), microelectromechanical system (MEMS), and substrate-integrated waveguide (SIW).

Aperture coupled dielectric resonator antenna optimisation using machine learning techniques

In this article, various machine learning (ML) algorithms such as Artificial neural network (ANN), Random Forest, XG boost, K nearest neighbor (KNN), and Knowledge-based neural network (KBNN) are used for efficient optimization of dielectric resonator antenna (DRA). ML models are used to predict the |S11| for a particular set of frequency, resonator height, aperture radius, and resonator radius. Finally, a comparative performance analysis of different ML algorithms has been done with the outcomes of the HFSS EM simulator. Error percentage is in between 1.0% and 7.0% with different ML algorithms. Antenna design is also fabricated and tested. The performance of the fabricated prototype is very close to different ML algorithms and HFSS obtained outcomes.

Knowledge-Based Neural Networks for Fast Design Space Exploration of Hybrid Copper-Graphene On-Chip Interconnect Networks

In this article, an artificial neural network (ANN) is developed in order to predict the per-unit-length (p. u. l.) parameters of hybrid copper-graphene on-chip interconnects from a prior knowledge of their structural geometry and layout. The salient feature of the proposed ANN is that it combines knowledge of the p. u. l. parameters extracted from empirical models along with that extracted from a rigorous full-wave electromagnetic solver. As a result, the proposed ANN is referred to as a knowledge-based neural network (KBNN). The KBNN has been found to converge to the same accuracy as a conventional ANN but at the expense of far smaller training time costs. As a result, the KBNN is much more suitable for performing design space explorations.

Local bi-fidelity field approximation with Knowledge Based Neural Networks for Computational Fluid Dynamics

This work presents a machine learning based method for bi-fidelity modelling. The method, a Knowledge Based Neural Network (KBaNN), performs a local, additive correction to the outputs of a coarse computational model and can be used to emulate either experimental data or the output of a more accurate, but expensive, computational model. An advantage of the method is that it can scale easily with the number of input and output features. This allows bi-fidelity modelling approaches to be applied to a wide variety of problems, for instance in the bi-fidelity modelling of fields. We demonstrate this aspect in this work through an application to Computational Fluid Dynamics, in which local corrections to a velocity field are performed by the KBaNN to account for mesh effects. KBaNNs were trained to make corrections to the free-stream velocity field and the boundary layer. They were trained on a limited data-set consisting of simple two-dimensional flows. The KBaNNs were then tested on a flow over a more complex geometry, a NACA 2412 airfoil. It was demonstrated that the KBaNNs were still able to provide a local correction to the velocity field which improved its accuracy. The ability of the KBaNNs to generalise to flows around new geometries that share similar physics is encouraging. Through knowledge based neural networks it may be possible to develop a system for bi-fidelity, computer based design which uses data from past simulations to inform its predictions.

Machine learning materials physics: Multi-resolution neural networks learn the free energy and nonlinear elastic response of evolving microstructures

Many important multi-component crystalline solids undergo mechanochemical spinodal decomposition: a phase transformation in which the compositional redistribution is coupled with structural changes of the crystal, resulting in dynamically evolving microstructures. The ability to rapidly compute the macroscopic behavior based on these detailed microstructures is of paramount importance for accelerating material discovery and design. Here, our focus is on the macroscopic, nonlinear elastic response of materials harboring microstructure. Because of the diversity of microstructural patterns that can form, there is interest in taking a purely computational approach to predicting their macroscopic response. However, the evaluation of macroscopic, nonlinear elastic properties purely based on direct numerical simulations (DNS) is computationally very expensive, and hence impractical for material design when a large number of microstructures need to be tested. A further complexity of a hierarchical nature arises if the elastic free energy and its variation with strain is a small-scale fluctuation on the dominant trajectory of the total free energy driven by microstructural dynamics. To address these challenges, we present a data-driven approach, which combines advanced neural network (NN) models with DNS to predict the homogenized, macroscopic, mechanical free energy and stress fields arising in a family of multi-component crystalline solids that develop microstructure. The microstructures are numerically generated by solving a coupled, mechanochemical spinodal decomposition problem governed by nonlinear strain gradient elasticity and the Cahn–Hilliard phase field equation. The hierarchical structure of the free energy’s evolution induces a multi-resolution character to the machine learning paradigm: We construct knowledge-based neural networks (KBNNs) with either pre-trained fully connected deep neural networks (DNNs), or pre-trained convolutional neural networks (CNNs) that describe the dominant characteristic of the data to fully represent the hierarchically evolving free energy. We demonstrate multi-resolution learning of the materials physics; specifically of the nonlinear elastic response for both fixed and evolving microstructures.

A neural network structure specified for representing and storing logical relations

Logical representation and reasoning is an important aspect of intelligence. Current ANN models are good at perceptual intelligence while they are not good at cognitive intelligence such as logical representation, so researchers have tried to design novel models so as to represent and store logical relations into the neural network structures, called the type of Knowledge-Based Neural Network. However, there is an ambiguous problem that the same neural network structure represents multiple logical relations. It causes the corresponding logical relations not to be read out from these neural network structures which are constructed according to them. To let logical relations stored in the format of neural network and read out from it, this paper studies the direct mapping method between logical relations and neural network structures and proposes a novel model called Probabilistic Logical Generative Neural Network, which is specified for logical relation representation by redesigning the neurons and links. It can make neurons solely for representing things while making links solely for representing logical relations between things, and thus no extra logical neurons and layers are needed. Moreover, the related construction and adjustment methods of the neural network structure are also designed making the neural network structure dynamically constructed and adjusted according to logical relations.

A Novel Estimation Method for the State of Health of Lithium-Ion Battery Using Prior Knowledge-Based Neural Network and Markov Chain

The state of health (SOH) of lithium-ion batteries (LIBs) is a critical parameter of the battery management system. Because of the complex internal electrochemical properties of LIBs and uncertain external working environment, it is difficult to achieve an accurate SOH determination. In this paper, we have proposed a novel SOH estimation method by using a prior knowledge-based neural network (PKNN) and the Markov chain for a single LIB. First, we extract multiple features to capture the battery aging process. Due to its effective fitting ability for complex nonlinear problems, the neural network with a prior knowledge-based optimization strategy is adopted for the battery SOH prediction. The Markov chain, with the advantageous prediction performance for the long-term system, is established to modify the PKNN estimation results based on the prediction error. Experimental results show that the maximum estimation error of the SOH is reduced to less than 1.7% by adopting the proposed method. By comparing with the group method of data handling and the back-propagation neural network in conjunction with the Levenberg-Marquardt algorithm, the proposed estimation method obtains the highest SOH accuracy.

Optimisation of reflection coefficient of microstrip antennas based on KBNN exploiting GPR model

When microstrip antennas (MSAs) are optimised, the full-wave electromagnetic simulation takes long time to get the result which is very time consuming. Hence, the knowledge-based neural network (KBNN) is used here to replace the electromagnetic simulation to shorten the calculating time and raise efficiency. However, prior knowledge of KBNN is always obtained by empirical formulas and neural networks, both of them are heavy and complicated. In this study, Gaussian process regression (GPR) is proposed to get the prior knowledge. The modelling of MSA is realised by the prior knowledge input method. The short optimal time and excellent optimal results prove that the proposed KBNN is a fast and effective method for the optimisation of MSAs.

A Unified Automated Parametric Modeling Algorithm Using Knowledge-Based Neural Network and ${l}_{1}$ Optimization

Knowledge-based neural network modeling techniques using space-mapping concept have been demonstrated in the existing literature as efficient methods to overcome the accuracy limitations of empirical/equivalent circuit models when matching new electromagnetic data. For different modeling problems, the mapping structures can be different. In this paper, we propose a unified automated model generation algorithm that uses $\boldsymbol {l}_{\mathbf {1}}$ optimization to automatically determine the type and the topology of the mapping structure in a knowledge-based neural network model. By encompassing various types of mappings of the knowledge-based neural network model in the existing literature, we present a new unified model structure and derive new sensitivity formulas for the training of the unified model. The proposed $\boldsymbol {l}_{\mathbf {1}}$ formulation of modeling can force some weights of the mapping neural networks to zeros while leaving other weights as nonzeros. We utilize this feature to allow $\boldsymbol {l}_{\mathbf {1}}$ optimization to automatically determine which mapping is necessary and which mapping is unnecessary. Using the proposed $\boldsymbol {l}_{\mathbf {1}}$ optimization method, the mapping structure can be determined to address different needs of different modeling problems. The structure of the final knowledge-based model can be flexible combinations of some or all of linear mapping, nonlinear mapping, input mapping, frequency mapping, and output mapping. In this way, the proposed algorithm is more systematic and can further speed up the knowledge-based modeling process than existing knowledge-based modeling algorithms. The proposed method is illustrated by three microwave filter modeling examples.

KBNN Based on Coarse Mesh to Optimize the EBG Structures

The microwave devices are usually optimized by combining the precise model with global optimization algorithm. However, this method is time-consuming. In order to optimize the microwave devices rapidly, the knowledge-based neural network (KBNN) is used in this paper. Usually, the a priori knowledge of KBNN is obtained by the empirical formulas. Unfortunately, it is difficult to derive the corresponding formulas for the most electromagnetic problems, especially for complex electromagnetic problems; the formula derivation is almost impossible. We use precise mesh model of EM analysis as teaching signal and coarse mesh model as a priori knowledge to train the neural network (NN) by particle swarm optimization (PSO). The NN constructed by this method is simpler than traditional NN in structure which can replace precise model in optimization and reduce the computing time. The results of electromagnetic band-gap (EBG) structures optimally designed by this kind of KBNN achieve increase in the bandwidth and attenuation of the stopband and small passband ripple level which shows the advantages of the proposed KBNN method.

Modeling and optimization of microwave filter by ADS-based KBNN

When microwave devices are designed by knowledge-based neural network (KBNN), the empirical formula is always used as priori knowledge. However, it is difficult to derive the corresponding formulas for the most electromagnetic problems, especially for complex electromagnetic problems, the formula derivation is almost impossible. In this article, they combine neural network with simulation software and use results of Agilent ADS as priori knowledge and HFSS as teaching signal to train the neural network by particle swarm optimization (PSO), which solves the difficulty in obtaining priori knowledge and effectively reduces the complexity of the neural network structure. Based on the KBNN, the microwave filters are designed. The results of optimization satisfy the required specifications which show the effectiveness and superiority of the method.

Knowledge-based artificial neural network model to predict the properties of alpha+ beta titanium alloys

In view of emerging applications of alpha+beta titanium alloys in aerospace and defense, we have aimed to develop a Back propagation neural network (BPNN) model capable of predicting the properties of these alloys as functions of alloy composition and/or thermomechanical processing parameters. The optimized BPNN model architecture was based on the sigmoid transfer function and has one hidden layer with ten nodes. The BPNN model showed excellent predictability of five properties: Tensile strength (r: 0.96), yield strength (r: 0.93), beta transus (r: 0.96), specific heat capacity (r: 1.00) and density (r: 0.99). The developed BPNN model was in agreement with the experimental data in demonstrating the individual effects of alloying elements in modulating the above properties. This model can serve as the platform for the design and development of new alpha+beta titanium alloys in order to attain desired strength, density and specific heat capacity.