In this article, we have presented the use of prior physics knowledge-based artificial neural networks (KBANNs) to improve the simulation speed and accuracy of compact models for circuit simulations. Multiple architectures such as source-difference artificial neural network (SD-ANN), prior knowledge input artificial neural network (PKI-ANN), and prior knowledge input with difference artificial neural network (PKID-ANN) have been developed to evaluate the best KBANN framework for runtime improvement, accuracy improvement as well as development (training) time improvement. In particular, the proposed PKID-ANN is the best KBANN since it matches the accuracy of rigorous physics-based technology computer-aided design (TCAD) solvers while exhibiting the most compact architecture, the fastest learning rate, and the fastest runtime. To the best of our knowledge, we have reported the first PKID-based architecture for device modeling in this article, as well as the fastest (simulation time) artificial neural network (ANN) compact model for FinFETs, which even outperforms conventional ANNs (C-ANNs) and industry standard compact models (BSIM-CMG) while retaining the physics and predictive accuracy. Moreover, the reported model also has the lowest training/development cost among all alternatives.
27 December 2023 in IEEE Transactions on Electron Devices
In this work, using multiobjective optimization (MOO) technique, design optimization of a gate-all-around field-effect transistor (GAAFET) has been performed for improved device logic and RF parameters. By using fast and accurate machine learning (ML) surrogate model, we have emulated the logic and RF performance figures of merit as analytic functions of the design objectives. Datasets required to train and test the ML model are generated using the well-calibrated TCAD setup. The multiobjective optimizers automate and perform extremely fast multispace design optimization. Contrary to MOO, TCAD optimization is tedious and time-intensive. Keeping in view of the International Roadmap of Devices and Systems (IRDS) target, optimal design trade-offs between ${I}_{ \mathrm{\scriptscriptstyle ON}}/{I}_{ \mathrm{\scriptscriptstyle OFF}}$ ratio and speed for logic; gain and cut-off frequency ( ${f}_{T}$ ) for RF operation are obtained for a sub-2 nm node GAAFET. Circuit simulation is performed to further validate the design optimization methodology. Moreover, it has been demonstrated that an efficient and faster trade-off between complex nonlinear design parameters can be automated by leveraging the ML-coupled MOO framework for any advanced-node FET.
23 June 2020 in IEEE Transactions on Components, Packaging and Manufacturing Technology
In this paper, a closed-form matrix rational approximation (MRA) model is presented for the reliability assessment of copper-graphene hybrid on-chip interconnect networks. The key feature of this MRA model is its capacity to predict how different values of temperature and dielectric roughness affect the signal integrity performance of the hybrid interconnect networks. As a result, the proposed MRA model is well-suited for very fast parametric sweeps and worst-case analysis of the hybrid interconnect networks, which has not been possible using existing closed-form models or even SPICE simulations. Numerical examples show that the proposed model is significantly more efficient than conventional models while exhibiting error less than 5%.
15 June 2023 in IEEE Transactions on Electromagnetic Compatibility
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.
12 February 2024 in IEEE Transactions on Electromagnetic Compatibility
In this article, artificial neural network (ANN) metamodels have been developed for the fast statistical signal integrity analysis of multiwalled carbon nanotube and multilayer graphene nanoribbon interconnect networks. These ANN metamodels, referred to as primary ANNs, are trained using transfer learning strategies where the initial guess of the weights and bias terms are learned from a pretrained secondary ANN. The secondary ANN is trained using data extracted from a cheap and approximate equivalent single conductor model of the interconnects. Starting with this informed initial guess, the weights and bias terms of the primary ANN can be optimized using a very small dataset generated from the rigorous but massive multiconductor circuit model of the interconnects. This makes the primary ANN significantly more efficient to train than existing metamodels. Two distinct transfer learning strategies based on the full and partial transfer of knowledge between the primary and secondary ANNs have been developed in this work.
1 August 2024 in IEEE Transactions on Components, Packaging and Manufacturing Technology
In this paper, an artificial neural network (ANN) enabled uncertainty quantification technique is developed for graphene on-chip interconnects. In the proposed technique, primary ANNs are trained to emulate the signal integrity characteristics of multi-walled carbon nanotube (MWCNT) and multilayer graphene nanoribbon (MLGNR) interconnects. The training of the primary ANNs is accelerated by using information elicited from other ANNs, known as secondary ANNs, that have been pre-trained to perform related tasks. In this paper, the elicited information takes two forms – (a) the optimized values of the weights and bias terms and (b) the output features of the secondary ANNs. Algorithms to intelligently infuse these two distinct types of information using a multistep training process has been proposed to ensure the best possible speedup. Validation examples spanning both MWCNT and MLGNR interconnect networks and different technology nodes have been presented.