Scholarlink

<p>Memtranstor emulators offer a circuit-level approach to accurately emulate the charge-flux hysteresis phenomenon, but prediction is challenging with the inclusion of leakage, active-block nonlinearities, saturation, and noise. In this paper, a hard-constrained physics-residual neural digital twin for nonideal memtranstor emulator hysteresis prediction is presented. A research-grade synthetic dataset was created with a nonideal circuit-based teacher model, which includes flux and memory leakage, input perturbation, saturation-limited inverse memtranstance, and nonlinear transconductance variation. The emulator cases were created using Latin Hypercube Sampling, resulting in 300 cases with a split of 210 training, 45 validation, and 45 test cases, providing 1.8 million time-domain samples. The proposed model assumes an ideal compact memtranstor equation as a physics prior, and learns a neural residual correction to inverse memtranstance, and the charge-flux relation <i>q</i><sub>pred</sub> = <i>φ M</i><sub>T,pred</sub><sup>−1</sup> is enforced by construction. It was compared to an ideal compact physics baseline and a pure ANN baseline. The proposed model achieved a <i>R</i><sup>2</sup> = 0.9719 and <i>R</i><sup>2</sup> = 0.9205 for charge and inverse memtranstance prediction, respectively, and the RMSE for charge prediction was also improved by 32.60% and the RMSE for inverse memtranstance prediction was improved by 10.33% compared with the ANN baseline. These results show that hard-constrained physics-residual learning not only keeps the physically consistent hysteresis behavior but also enhances the generalization.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Hard-constrained physics-residual neural digital twin for nonideal memtranstor emulator hysteresis prediction

  • 期刊论文

  • 发表日期：

    2026年6月19日

查看全文
  • Sumit Vyas

摘要

Memtranstor emulators offer a circuit-level approach to accurately emulate the charge-flux hysteresis phenomenon, but prediction is challenging with the inclusion of leakage, active-block nonlinearities, saturation, and noise. In this paper, a hard-constrained physics-residual neural digital twin for nonideal memtranstor emulator hysteresis prediction is presented. A research-grade synthetic dataset was created with a nonideal circuit-based teacher model, which includes flux and memory leakage, input perturbation, saturation-limited inverse memtranstance, and nonlinear transconductance variation. The emulator cases were created using Latin Hypercube Sampling, resulting in 300 cases with a split of 210 training, 45 validation, and 45 test cases, providing 1.8 million time-domain samples. The proposed model assumes an ideal compact memtranstor equation as a physics prior, and learns a neural residual correction to inverse memtranstance, and the charge-flux relation qpred = φ MT,pred−1 is enforced by construction. It was compared to an ideal compact physics baseline and a pure ANN baseline. The proposed model achieved a R2 = 0.9719 and R2 = 0.9205 for charge and inverse memtranstance prediction, respectively, and the RMSE for charge prediction was also improved by 32.60% and the RMSE for inverse memtranstance prediction was improved by 10.33% compared with the ANN baseline. These results show that hard-constrained physics-residual learning not only keeps the physically consistent hysteresis behavior but also enhances the generalization.