Session Chair: Lining Zhang, Peking University Shenzhen
Formulation and Applications of Industry-standard MVSG GaN HEMT Compact Model
Invited Speaker: Lan Wei, University of Waterloo
Abstract: Given its high mobility, high breakdown voltage and decent thermal conductivity, GaN HEMT devices are a front-runner for the next generation power RF applications and power electronics applications. With commercial roll-out of GaN technology, development of accurate, scalable and efficient compact model is essential. The MIT Virtual-source Gallium Nitride (MVSG) model is one of the two industry-standard compact models for GaN HEMTs.
This presentation will provide a brief overview of the physics-based MVSG GaN HEMT compact model, including the model formulation and various features. Its recent progress for RF GaN and power GaN devices will also be presented with application examples, showing the potentials of this physics-based compact model.
HiRL-DCMPE：A Combined HiMOSS and RL Framework for Device Compact Model Parameter Extraction
Presenter: Ye Lu, Fudan University
Abstract: The relentless scaling of semiconductor device technology has significantly increased the complexity and computational demands of device compact model (DCM) parameter extraction and model creation. The extraction workflow typically needs to simultaneously optimize hundreds of interdependent parameters while meeting stringent requirements for both computational efficiency and model physical accuracy. In this work, a new method that combines HiMOSS with RL algorithms for DCM parameter extraction is proposed. Specifically, we (i) propose HiRL-DCMPE, a novel algorithm that combines a modified Bayesian optimization and reinforcement learning. This framework narrows down the high-dimensional search space in the early stage and subsequently performs fine-grained optimization locally in the later stage. This not only improves the performance but also enhances the efficiency of the automated DCM parameter extraction; (ii) conduct experiments of parameter extraction task for both BSIMCMG and BSIMSOI models, and more than 100 parameters are extracted at once. It is shown that the RMSE of the created model output v.s. target data is < 5%, satisfying the needs of practical usage. In addition, the results outperform the benchmark algorithms by at least 20% in terms of RMSE value; and (iii) Overall, this framework improves the search efficiency and facilitates the optimal solution for complex DCM parameter extraction tasks. It takes at least 11.8% less time compared to the benchmark algorithm to reach the same accuracy level. These results demonstrate the great potential of HiRL-DCMPE for future DCM parameter extraction purpose.
APEX: Automating Parameter Extraction of Compact Models with Differential Neural Network Approximation
Presenter: Jianing Zhang, East China Normal University
Abstract: The traditional compact model parameter extraction highly depends on engineers' expertise, leading to a time-consuming and iterative process. To address the above issue, this paper proposes an automatic parameter extraction method for compact models, APEX. The proposed APEX framework adopts an artificial neural network (ANN) method as an approximation of compact models using model parameters as inputs and IV/CV data as outputs. The model parameters are efficiently extracted using an automatic differential mechanism based on the ANN-approximated compact model. Experimental results demonstrate that our proposed framework achieves good fitting accuracy and scalability across device structures when evaluating GAA and FinFET devices. A fitting error of less than 4% is achieved on the open-source benchmark.
MOIL: An Efficient Multi-objective Optimization Framework for SRAM Cell with Incremental Learning
Presenter: Baokang Peng, Peking University
Abstract: This paper proposes MOIL, an efficient multi-objective optimization framework for SRAM design that combines neural network (NN) surrogate modeling with NSGA-II evolutionary algorithms. The framework features: 1) An NN surrogate model employing a two-phase training strategy — initially trained on Latin Hypercube Sampling data and progressively refined through incremental learning cycles — achieving 99.83% prediction accuracy on critical SRAM metrics (read/write delay, SNM, leakage power) while eliminating iterative SPICE simulations; 2) An adaptive NSGA-II optimizer incorporating dynamic crowding distance calculation to effectively explore 5-dimensional device parameters (Fin width/height, Lg, PHIG) and generate Pareto-optimal solutions. Experimental results demonstrate MOIL's superior efficiency with 20.8× faster decision-making and 13.7× speedup over Bayesian optimization method, while maintaining solution diversity (hypervolume ratio of 3.52). The framework establishes a robust methodology for rapid SRAM evaluation and optimization in advanced technology nodes.
Formulation and Applications of Industry-standard MVSG GaN HEMT Compact Model
Invited Speaker: Lan Wei, University of Waterloo
This presentation will provide a brief overview of the physics-based MVSG GaN HEMT compact model, including the model formulation and various features. Its recent progress for RF GaN and power GaN devices will also be presented with application examples, showing the potentials of this physics-based compact model.