The novelity of this work is that the sensitivity of a hetero dielectric Bio Tunnel Field-Effect Transistor (HD-BioTFET) has been successfully predicted using autoregressive integrated moving average (ARIMA) machine learning model with a limited dataset obtained from TCAD simulations. HD-BioTFET is a charge plasma based label-free biosensor where, a high-K dielctric \(TiO_2\) introduced over source region promotes band-to-band tunneling and hence, a improvement in senstivity of \(2\times 10^{7} \) in HD-BioTFET than \(1.6\times 10^{7}\) of BioTFET for K=10. Also, the senstivity improved for charged biomolecules is \(2.6\times 10^{8}\) \(A/\mu m \) in HD-BioTFET than \(1.35\times 10^{8}\) \(A/\mu m \) in BioTFET and \(1.34\times 10^{3}\) \(A/\mu m \) in HD-BioTFET than \(5\times 10^{2}\) in BioTFET for ± 1e13 charge values, respectively. A small dataset of 40 rows and 4 coloumns obtained during optimization of HD-BioTFET is then used for training of Machine Learning (ML) models such as convolutional neural nework (CNN), artificial neural network (ANN), and ARIMA that serves the purpose of low computational power. However, due to the limited dataset, CNN and ANN fail, whereas ARIMA excels by handling sequential data and nonlinearities. ARIMA successfully predicted the drain current of the device, achieved 98% accuracy and F1 score = 1 for unknown K =3, 4.1, 4.6, 5, and 7 values and 98% accuracy and F1 score = 0.5 for unknown charged ( range± 4e11, ± 8e12, and ± 3e13) biomolecules. Hence, the sensitivity of HD-BioTFET for ARIMA predicted output and simulated output are closely matched, which justify the integration of ML to the biosensing application that promises a cost-effective, label-free, low-powered, and a higly accurate sensitive prediction solution.
