Steep Sub Threshold Phase Transition FETs

This paper elucidates the application of physics informed recurrent neural networks in formulating a predictive model for the promising steep sub-threshold phase transition FET. Given the ongoing exploration into the phase transition mechanisms within transition metal oxides - whether dominated by peierls or Mott-Hubbard interactions-neural models emerge as a compelling alternative to conventional physics-driven equations. The model developed herein demonstrates a robust capability to accurately predict eight critical parameters of Phase-FET, contingent in variations in both input voltages. Empirical results indicate a reduction in test mean squared error to 0.04 for the basic RNN configuration and 0.05 for the LSTM variant when compared with the existing model of phase transition FET. Furthermore, the median absolute percentage error in the prediction of the Phase-FETs drain current is observed to be 0.03 using the RNN, while an even lower value of 0.01 is achieved with the LSTM. Cumulative distribution function of error gives a probability of 99 percent for accurate prediction especially in case of GmoverId, drain to source capacitance, source to gate capacitance, drain to bulk capacitance and on resistance of phase transition FET.