ArticlePDF Available

Abstract and Figures

This paper presents performance enhancement of Si 3 N 4-gate Ion-Sensitive Field-Effect Transistor (ISFET) based pH sensor using Machine Learning (ML) techniques [1]. A robust SPICE macromodel is developed using experimental data which incorporates intrinsic temperature and temporal characteristics of the device, which is further used in sensor readout circuit (ROIC), which shows a non-ideal temperature and time dependence in the voltage output. To make the device robust to the critical drifts, we exploit six state-of-the-art ML models, which are trained using the data generated from ROIC for a wide range of pH, temperature, and temporal conditions. Thorough comparison between ML models shows Random Forests (RF) outperforms other models for drift compensation task. This work also shows a preliminary time series classification task. The ML models are implemented on a Xilinx ® PYNQ-Z1 Field Programmable Gate Array (FPGA) board to validate the performance in power and memory-restricted environment, crucial for IoT applications. A parameter, Implementation Factor (IF) is defined to evaluate best ML model for IoT deployment using FPGA/MCU hardware implementation. The significantly lower power consumption of FPGA board as compared to CPU with no noticeable performance drop is a pointer to the future of robust pH sensors used in industrial and remote IoT applications.
Content may be subject to copyright.
Machine Learning on FPGA for Robust Si3N4-gate
ISFET pH Sensor in Industrial IoT Applications
Soumendu Sinha, Member, IEEE, Nishad Sahu, Rishabh Bhardwaj, Aditya Mehta, Hitesh Ahuja, Satyam
Srivastava, Anubhav Elhence, Vinay Chamola, Senior Member, IEEE
Abstract—This paper presents performance enhancement of
Si3N4-gate Ion-Sensitive Field-Effect Transistor (ISFET) based
pH sensor using Machine Learning (ML) techniques [1]. A robust
SPICE macromodel is developed using experimental data which
incorporates intrinsic temperature and temporal characteristics
of the device, which is further used in sensor readout circuit
(ROIC), which shows a non-ideal temperature and time depen-
dence in the voltage output. To make the device robust to the
critical drifts, we exploit six state-of-the-art ML models, which
are trained using the data generated from ROIC for a wide
range of pH, temperature, and temporal conditions. Thorough
comparison between ML models shows Random Forests (RF)
outperforms other models for drift compensation task. This
work also shows a preliminary time series classification task.
The ML models are implemented on a Xilinx®PYNQ-Z1 Field
Programmable Gate Array (FPGA) board to validate the per-
formance in power and memory-restricted environment, crucial
for IoT applications. A parameter, Implementation Factor (IF)
is defined to evaluate best ML model for IoT deployment using
FPGA/MCU hardware implementation. The significantly lower
power consumption of FPGA board as compared to CPU with no
noticeable performance drop is a pointer to the future of robust
pH sensors used in industrial and remote IoT applications.
Index Terms—ISFET, Machine Learning, SPICE Macromodel,
pH Sensor, Random Forest, Neural Networks, Drift Compensa-
The field of biomedical and chemical sensing applica-
tions has been a fascinating area for research since several
decades [2]. With the advancement in micro/nano fabrication
technologies, it requires that miniaturized sensors should not
compromise on the measurement accuracy and precision [3].
Electrochemical sensors are widely used for pH sensing ap-
plications [4]. Ion-Sensitive Field-effect Transistors (ISFETs)
are one of the most promising electrochemical sensors due to
inherent advantages of easy fabrication, low cost and small size
Soumendu Sinha is with the Semiconductor Devices Area, CSIR-Central
Electronics Engineering Research Institute(CEERI), Pilani, India (e-mail:
Nishad Sahu, Hitesh Ahuja and Anubhav Elhence are with the Department
of Electrical and Electronics, BITS-Pilani, Pilani Campus, 333031, India. (e-
Aditya Mehta is with the Department of Computer Science and In-
formation Systems, BITS-Pilani, Pilani Campus, 333031, India. (e-mail: and
Satyam Srivastava is with the Intelligent Systems Group, CSIR-Central
Electronics Engineering Research Institute(CEERI), Pilani, India (e-mail:
V. Chamola is with the Department of Electrical and Electronics Engi-
neering & APPCAIR, BITS-Pilani, Pilani Campus, 333031, India. (e-mail:
TABLE I: Table listing out the applications of ISFET in various industries.
Application Application Description Ref.
Bio-medical /
and research
Rapid detection of SARS-CoV-2
virus causing the COVID-19 disease [7–9]
Malaria diagnosis [10]
Human Genome Project and
genome/DNA sequencing [6, 11–14]
Accuracte and high throughput pH sensning [11]
Robust label-free Micro RNA detection [15]
Breast cancer mutation detection [16]
pH and oxygen partial pressure measurement [17]
Point of care biomedical applications [18]
DNA methylation detection [19]
2. Agriculture
Development of wireless sensor
nodes for algae cultivation [20]
Precision Agriculture [21–23]
Soil and Crops Measuements [24]
In situ monitoring of soil nutrients/
soil nitrate sensing/ soil analysis/soil sensor [25–29]
3. Environmental
Monitoring of wastewater [24]
Geochemical Barriers Monitoring [24]
Pesticide and toxin concentration detection [30]
Coastal pH monitoring [31]
Pollution monitoring in liquid media [32]
Sea water alkalinity measurement [33]
4. Wearable
Sweat sensing [34, 35]
Skin temperature monitoring [35]
On body pH measurement [36]
Wearable Point of care applications [37]
Body Fluid pH monitoring [38]
5. Food
Heavy metal ions detection in vegetables [39]
Determination of glucose,
ascorbic and citric acids [40]
Controlling Phosphororgannic
pesticides in Water and Vegetables [41]
Nitrate ion determination in vegetables [42]
6. Horticulture
pH and potassium measurement
in nutrient solution [43]
Multi ion sensing [44]
Soil Analysis [45]
Nutrient sensing for rockwool culture [46]
Application in closed
loop system for green houses [47]
[5, 6] for chemical/biochemical sensing applications, popularly
utilized for pH sensing applications, as listed in Table I.
Despite the wide range of applications, ISFETs are prone
to temperature and temporal drift, restricting their accuracy
in field applications [48]. ISFETs show non-linear drift char-
acteristics due to temperature dependent semiconductor and
electrochemical parameters [49, 50]. The temporal drift in
ISFET devices is due to chemical modification of the sensing
film [51]. In order to deploy ISFET-based pH sensors for
critical field applications such as biomedical devices, environ-
mental monitoring and food quality monitoring, robust drift
compensation methodologies are critical to provide accurate
sensor data [5, 52]. Manual calibration of such deviations
is difficult when remotely deployed in large numbers, and
the non-linear drift behavior makes it extremely challenging
with varying temperature and time in field deployment [49].
Therefore, Artificial Intelligence (AI)-assisted techniques are
increasingly being utilized for automated sensor calibration
Machine learning techniques are powering major trans-
formations in the novel era of automation. The algorithms
have proven to be extremely effective in learning impact
of independent variables on the dependant variables [54].
As ISFET-ROIC shows output voltage (pH) dependence on
ambient conditions, we infer that ML techniques greatly suits
the compensation task [48]. From the data generated using
device-ROIC, ML algorithm learns approximate input-output
functions in the training phase. To test a model’s performance,
i.e., testing phase, we evaluate the model on an unseen test
data. Previous works have seen importance of ML based
approaches in temperature and temporal drift compensation
For the compensation task, we compare the accuracy of
various models, viz., multi-layer perceptron, linear regression,
polynomial regression, decision trees, random forest and sup-
port vector machine. For the purpose of data generation, a
robust SPICE macromodel of Si3N4-gate ISFET has been
developed using experimental data, which includes the temper-
ature and temporal dependence of electrochemical and device
parameters [49]. The model is essential to generate quality
data for later used ML algorithms for the task of concern.
This SPICE macromodel of ISFET is exported as a sub-circuit
block in a constant voltage constant current (CVCC) ROIC.
In comparison to our previously reported work [1], we extend
this work to the implementation and performance comparison
of the ML models on FPGA/MCUs using Xilinx PYNQ-Z1
board for drift compensation in sensors and study feasibility
for IoT deployment.
To the best of our knowledge, our five major contributions
We introduce ML-based models to significantly reduce
temperature and temporal dependence of Si3N4-gate IS-
FET pH sensor in a CVCC ROIC.
Based on our compensated results, we compare the per-
formance of various ML models to evaluate the best per-
forming ML technique for Si3N4-gate ISFET-pH sensor.
As a preliminary work, we show how by only using
drifted values of the reference voltage, we can estimate
the pH of the solution.
We also present for the first time implementation of ML
models on FPGA using Xilinx PYNQ-Z1 board for drift
compensation in sensors and compare the results in terms
of accuracy, speed and power consumption with CPU,
MCU and FPGA implementation.
We also propose a novel factor called Implementation
Factor (IF) to evaluate the performance of ML models
for FPGA/MCU implementation. The factor is calculated
by giving equal weights to accuracy, speed and energy
consumption of the ML models on the FPGA.
n-Si n-Si
Sensing film
pH solution
Fig. 1: Schematic of the cross-sectional view of ISFET device
The paper organization is as follows: Section II presents
an overview of ISFET devices and its SPICE macromodel
formulation along with the details of ROIC topology. Section
III presents an overview of the machine learning algorithms
and the experimental setups for drift compensation tasks.
Section IV discusses the FPGA and MCU implementation
of the ML models. Section V presents the results obtained
from the various experimental setups discussed. Section VI
presents the conclusion of the work.
ISFET device is similar to a MOSFET in terms of the device
structure, except the gate electrode in the MOSFET is replaced
by a reference electrode, a sensing layer and an ionic solution,
as shown in Fig. 1 [49]. A reference electrode is inserted in
the ionic solution to supply a reference potential. The ionic
solution is in contact with the gate oxide, which leads to the
formation of charges at the solution/oxide interface, which
further modulates the threshold voltage of the ISFET, causing
variation in the channel current, which is sensed by a suitable
read out circuit, by sensing the voltage at the output terminal.
The semiconductor and the electrochemical parameters of
ISFET have a dependence on temperature (see Supplementary
article for details) [50]. Thus, the threshold voltage of ISFET
gets affected by the concentration of ions in the electrolyte
solution and the temperature [51]. Sensing films such as Si3N4
also show a significant temporal drift [49] due to inherent
transient nature of chemical reactions taking place at the
surface of the sensing film, leading to chemical modification of
insulator with time, which affects the device threshold voltage
A. SPICE Macromodel Formulation of Si3N4-gate ISFET
The macromodel of ISFET is developed in SPICE because
it gives us the flexibility to define custom sub-circuit blocks,
which contains robust modeling of all the parameters pertain-
ing to the circuit elements involved. Sinha et. al. carried out
comprehensive modeling of the dependence of the parameters
present in the both electrochemical and electronic stages,
considering the parameters’ dependence on temperature and
Fig. 2: Schematic of the developed ISFET SPICE
8.5 k
+4 V
+5 V
Fig. 3: CVCC topology for readout for ISFET sensor [1].
time in [49] to develop an accurate Si3N4gate ISFET SPICE
macromodel that closely mimics real world ISFET device.
The study uses experimental data to develop a robust SPICE
macromodel and the simulated characteristics of the SPICE
model is very close to the experimental results. We follow a
similar approach and use the developed SPICE macromodel
in this study. The ISFET SPICE model consists of 5 nodes,
namely drain (D), source(S), bulk (B), the reference electrode
potential connected to the electronic stage of the device via
electrical double-layer capacitors and pH, which is the linear
mapping of pH of the solution as an input voltage terminal.
In CVCC circuit topology, the sub-circuit block shown in Fig.
2 is used to obtain the sensor response [48].
B. ISFET Readout Circuit Topology
CVCC topology (Fig. 3) is used as a read out circuit derived
from previous study [49]. The voltages at the source and the
drain nodes of the device are kept constant using op-amps
in negative feedback and voltage buffer modes, respectively.
Hence, due to constant Vds and Ids, the quantity (Vref -Vt) is
fixed. Thus, any change in the threshold voltage due to varying
temperature, time and pH are reflected by change in Vref of
the device. The Vref output relates to the pH of the solution
when the temperature and time are constant. For the same pH
solution, when the temperature and time is varied, the Vref
values obtained correspond to erroneous values of measured
pH, which needs to be corrected through intelligent techniques.
The ROIC output is given as an input to the ML algorithms
to compensate the undesired temporal and temperature drift.
C. Accuracy of Si3N4-gate ISFET SPICE Macromodel
The SPICE macromodel used in this study is based on our
previous study on development of an accurate Si3N4gate pH-
ISFET device [49]. The results of [49] indicate that the model
is able to obtain characteristics very close to the experimental
data for both temperature and temporal drift. For temperature
drift , the model was developed based on the experimental
data reported in [56]. In the transfer characteristics of the
device, the simulated isothermal point was obtained at (4.25 V,
457.45 uA) against the experimentally reported value of (4.10
V, 468.26 uA) i.e. an error of 3.53 % in the gate voltage and
error of 2.29 % in the drain to source current, which is within
the tolerance limits.
For the temporal drift, the experimental data was taken
from [57]. Using the experimental data, the parameters were
extracted for modeling the temporal drift in the ISFET SPICE
macromodel which provided an Adj. R2value of close to 1.
Thus, the simulated temporal drift was very close to the the
experimental data, which is confirmed with a high value of
Adj. R2.
Thus, the SPICE macromodel reported in [49] gives very
close accuracy to the real ISFET device characteristics. We
provide an in-depth explanation of the developed mathemat-
ical model used in the ISFET SPICE macromodel for the
calculation of ISFET threshold voltage, temporal drift and
temperature drift in the Supplementary Article (Sections I.B,
I.C and I.D respectively). The developed model has been used
in this study for generating the data for training the ML
models. Thus, a successful drift compensation on this data set
can be emulated with high confidence for drift compensation
in real life ISFET-pH sensors.
We formulate the compensation task as two sub-tasks. The
first task focuses on compensating the voltage drift induced
by only considering one parameter, i.e. temperature; whereas
the second task aims to tackle a more complicated challenge
of offsetting voltage drift caused by variability in both tem-
perature and time. Since both time and temperature lead to
voltage drift, it is a non-linear function estimation problem,
thereby increasing its complexity.
All models were built using the Sklearn Python library in
Python 3.5 [58]. In this study, we compare task performances
of six state-of-the-art ML models described in this section. The
pipeline of approach is depicted in Fig. 4. We encourage reader
to refer [58] for fine-grained details about ML model and
function optimization that lead to model parameter learning.
A. Machine Learning Algorithms
1) Linear [LR] and Polynomial Regression [PR] LR is the
most basic form of regression technique which involves
fitting a straight line on the training data with a strong
assumption i.e. linear relationship between input-output
variables. PR is an extension to LR incorporating nonlin-
ear dependencies. The algorithm tries to fit a polynomial
curve on the data provided by optimizing a cost function.
If x={x1, . . . , xn}is the input vector and ybe the
2+... +αnx2
Experimental data of
Si3N4 gate ISFET pH
device from literature
Accurate SPICE
macromodel development
based on experimental
data which accounts for
both temperature and
temporal drift
Six different ML Algorithms:-
trained on PC for compensating
temperature and temporal drift The SPICE macromodel is
incorporated in a CVCC readout
circuit and data is generated for
training and testing Machine
Learning(ML) algorithms
Data split
into 2 parts
80 % used for
20 % used for
Trained model
executed on PC
for testing
Trained model
executed on PYNQ Z1
board testing
Comparison made
between different ML
algorithms based on
speed, accuracy and
energy consumed
Results indicate that FPGA implementation is more
suitable for ISFET Industrial IoT applications as compared
to PC. Random Forests gives best results based on the
above three parameters for FPGA implementation
Fig. 4: Pipeline of the adopted methodology
Fig. 5: Schematic of RNN architecture.
where αiare weights which are learnt during training the
regressor model.
2) Support Vector Regression [SVR] The algorithm works
on a similar principle as SVM (Support Vector Machine).
In SVM, the algorithm finds a hyperplane, with function
f(x) = wx+b, given f(x)lies within margin of tolerance
() from the value y(x)of every data point, with the
greatest possible margin between the decision line and the
nearest point from it from samples representing a specific
class. The hyperplane with the maximum distance margin
will classify the data points correctly. However, in the
case of SVR, the support vectors come up with the closest
match between the data points and the actual function
that is represented by them. So, maximizing the distance
between the support vectors to the regressed curve, we
move towards the actual curve. Mathematically, this can
be represented as
minimize f=1
s.t. yiw1·xibε;w1·xi+byiε(3)
3) Decision Trees [DT] This algorithm builds the models
in the form of an inverted tree structure. Each branch
of the tree represents a possible decision, and with such
branches, the associated decision tree is incrementally
developed. The leaf nodes at the bottom represent the
final decision. In regression problem, the criteria to split
the branches is usually decided by selecting the setup
resulting in minimum mean square error or standard
4) Random Forest [RF] RF constructs multiple DTs during
training phase. It fits these DT on sub-samples of data
(a) (b)
(c) (d)
Fig. 6: Characteristic curves of Si3N4-based ISFET: a)Vref-pH
curves at different temperatures; b)Vref-Temperature curves at
different pH; c)Vref-time curves at different temperatures for
pH 7; d)Vref-time curves at different temperatures for pH 10
and gives the averaged output for the computation of
accuracy. There are two key concepts which causes the
randomness. Firstly, there is random sampling of training
data at the time of building trees. Secondly, there are ran-
dom subsets of features examined when splitting nodes.
For an input, select with replacement Krandom samples
from the training set followed by learning a separate
decision tree. Prediction on an unseen sample xnew can
be made by averaging prediction from all the individual
DT on xnew
Fig. 7: PYNQ-Z1 board booted with the PYNQ image and
drawing power from USB port via USB power tester.
f(xnew) = 1
5) Multi-layer Perceptron [MLP] An MLP is a feed-forward
artificial neural network model which maps the input data
to the output data. The framework of MLP consists of
neurons, which are the basic units of computations. This
fundamental unit receives input from other neurons or
an external source and computes the output. Each of the
inputs have an associated weight, which is being imposed
on the bias of its comparative importance to other inputs.
A neuron performs the following operation,
f(x) = ϕ n
where w represents the weight vector, x represents the
input vector, b adds the bias an ϕis the activation
6) Recurrent Neural Networks (RNN) Standard feed-forward
neural networks fail to capture sequential structure of the
input. To tackle problems where the input has a series
of values having notion of sequence, a special class of
neural network were designed. An RNN (directionally)
iterates over the element of the sequence and encodes
the necessary information from the sequence. The input
is processed as follows:
ht=fW(ht1, vt)(6)
where, fW, and fOare the function with trainable pa-
rameters Wand O, respectively. fOlearns to map hidden
representation of input. htis the hidden RNN generates
until sequence at step t, i.e., vt. For prediction, we use
hnt,ntbeing the sequence length. We feed hntto a
feed-forward NN with softmax activation to predict the
pH class,(as shown in Fig. 5.
We used mean absolute error (MAE), mean square error
(MSE) and R2metrics to test the above models. [59].
B. Experiment 1 - Temperature Drift Compensation
As discussed above, temperature is one of the primary
factors leading to ROIC output drift provided by an ISFET
sensor. We plot the temperature dependent characteristics of
ISFET ROIC as shown in Fig. 6 (a) and (b). We tested several
ML models in this first experiment, which attempts to predict
the actual pH by compensating for Vref drift. Here, we have
used the data points from the set of reference voltages and
corresponding pH of the solution, ranging from 0 to 14 at
a resolution of 0.01. The temperature is varied from 15oC
to 65oC. Hence, the total number of samples obtained were
66,351. The whole data was split into training and testing
data sets in the ratio of 80:20%. For efficient preparation,
preprocessing techniques, such as normalization and mean
centering, were taken into account. We have taken ROIC
output voltage and ambient temperature as input features for
the algorithms.
(a) pH-7 (b) pH-10
Fig. 8: Prediction curve with and without ROIC temperature
drift compensation for pH 7 and pH 10 [1].
C. Experiment 2 - Temperature-Temporal Drift Compensation
For this experiment, we have taken data points for tempera-
ture ranging from 15oC to 65oC and time varying from 0 hours
to 50 hours. We plot temperature-temporal dependent ROIC
output as shown in Fig. 6 (c) and (d). Data corresponding
to pH 7 and 10 were considered for this experiment, which
resulted in a total of 81,254 data samples. Again, the data
is split in the ratio 80:20% for training and testing purposes
and similar pre-processing as in Experiment 1 was applied. We
have considered pH, ambient temperature and time as the input
features for training all the models, with an aim to capture the
changing voltage for a particular pH with change in time and
D. Experiment 3 - Time series classification
Owing to the limited temporal data for pH values 7 and 10,
we also performed time-series classification. Given a set of 10
consecutive Vref reading given by the pH sensor, the binary
classification task is to predict the pH of the solution. The
problem formulation is slightly harder from the previous two
experiments as we do not have the time stamp information at
which the readings are noted, and also the data is scarce. For
TABLE II: Comparison of results for experiments on the Xilinx PYNQ-Z1 board(PS) and the Intel Core CPU for temperature drift compensation.
Model Metrics for Xilinx PYNQ-Z1 Board(PS) Metrics for Intel(R) Core(TM) i5-8250U
CPU @ 1.6 GHz - Quad core
(W) MAE MSE R2Time
LR 0.234 1.237e-01 0.99121 0.004 2.4552 0.234 1.237e-01 0.99121 0 6.0
PR 0.071 1.118e-02 0.99916 0.013 2.7181 0.071 1.118e-02 0.99916 0.0009 9.7
DT 0.058 6.713e-03 0.99952 0.023 2.7181 0.058 6.713e-03 0.99952 0.0029 9.6
RF 0.007 8.947e-05 0.99999 0.272 2.7280 0.007 8.627e-05 0.99999 0.0568 10.8
MLP 0.025 9.809e-04 0.99993 0.035 2.52464 0.025 9.809e-04 0.99993 0.0366 9.6
SVR 0.074 1.060e-02 0.99925 73.878 2.66352 0.074 1.059e-02 0.99925 1.1847 21.2
(a) (b)
(c) (d)
Fig. 9: Prediction curve with both temporal and temperature
drift compensation and only temperature compensation: a) pH-
Time curve for pH 7, temperature 20oC; b) pH-Time curve
for pH 10, temperature 20oC; c) pH-Time curve for pH 7,
temperature 35oC; d) pH-Time curve for pH 10, temperature
35oC [1].
the time-series (sequence) classification, we use RNNs. The
input consists of two feature vectors, one corresponding to
the ten consecutive Vref readings and the other is the natural
logarithm of each voltage reading. As we suspect the voltage
should follow an exponential decay curve, the logarithm of the
voltage might show linear dependence with time. This comes
from the idea that learning linear relationships is significantly
easier (since it facilitates analytical solutions) than capturing
a non-linear relationship between dependent and independent
variables from the dataset. We sample 2,500 non-overlapping
sequences of Vref through time and for both the pH 7 and 10.
Training set constitutes randomly shuffled 70% of the dataset
samples, 10% is used for development, 20% for performance
TABLE III: Summarizing experiments for input and output
Experiment Input features Output
1:Temp Comp Temperature+Vref pH
2:Temp-time Comp Temperature+Vref +Time pH-shift
3:Series classification Sequence of Vref values pH
We have summarised the input features and expected output
of the above experiments in Table III.
In addition to conventional implementation on CPU, we also
implemented the algorithms discussed in the Section III on
FPGA and MCU Hardware. The ML models discussed in this
study are relatively simple and therefore in addition to FPGAs,
they can also be implemented using micro-controllers with
specialised instruction set, although it may take long time. In
this study, we choose to implement them on Programmable
System (PS) and Programmable Logic (PL) of the Xilinx
PYNQ-Z1 board. The implementation on the PS is considered
as an MCU implementation while the implementation on the
PS+PL is considered as an FPGA implementation. FPGAs
provide the feature of rapid prototyping. In comparison to
FPGAs, micro-controllers with specialise instruction set con-
sume lesser area but consume more time as well as power
for implementing the same ML algorithms. PYNQ (Python
Productivity for ZYNQ) is the latest open-source project from
Xilinx which makes it easier to program FPGAs using the
python language [60]. Traditionally, Verilog HDL code was
essential to be written for implementing algorithms on hard-
ware like FPGAs but with the aid of the PYNQ framework,
such algorithms can now be directly programmed into FPGA
using languages such as Python and C/C++ and with the help
of hardware libraries/overlays, using a hardware-software co-
design approach. The hardware libraries or overlays are written
in conventional languages such as Verilog and can be used
directly to program the FPGA on board the PYNQ-Z1 board.
Here, we choose the PYNQ-Z1 board as it has very low power
consumption, which makes it suitable for IoT and industrial
applications. To monitor the power consumption of the PYNQ-
Z1 board in real time, we used a USB power tester. A real time
image of the setup is given in Fig. 7. The PYNQ-Z1 board has
TABLE IV: Programmable logic resource in the PYNQ Z1
board[60, 61].
No. Programmable logic resource No. of Units
1. 6-input LUTs 53,200
2. Flip-flops 106,400
3. BRAM (block RAM) 630 KB
Clock management tiles each
with a phase-locked loop(PLL)
and mixed-mode-clock manager
5. DSP slices 220
6. On chip analog-to
digital converter 1
Fig. 10: Screenshot of the Xilinx Power Estimator Window for
estimating the worst case power consumption on the PYNO-
Z1 board.
a 650 MHz dual core Cortex-A9 processor and PL equivalent
to Artix- FPGA. It has 13,300 logic slices, each with four 6-
input LUTs and 8 flip flops. Details of its PL resources are
given in Table IV.
In addition to this, we also perform worst case power
estimation using the Xilinx Power estimator (XPE) tool in
the following subsection.
A. Power Estimation using Xilinx Power Estimator(XPE)
It is important to determine the power and cooling speci-
fications of SoC(System on Chip) and FPGA designs early.
Accurate worst case power analysis helps in avoiding ma-
jor roadblocks in the product development life-cycle. The
Xilinx SPE is a spreadsheet tool designed for this purpose.
we performed the power estimation using the Xilinx Power
Estimator for Zynq 7000 family (PYNQ Z1 board belongs
to ZYNQ XC7Z020-1CLG400C) [60, 61]. On choosing the
Fig. 11: Total on chip power consumption on the PYNQ Z1
board for different amount of resource utilisation of the PL
(estimated using Xilinx Power estimator).
suitable device settings and 100 % resource utilisation in
the Programmable Logic (PL) , the total on-chip power was
estimated to be 2.861 W as shown in Figure 10. This proves
the lower power consumption of the board and makes it
feasible for checking designs meant for IoT deployment. This
is also very close to the results obtained with the USB power
tester as discussed in Tables II and VII . Also we plot
the power consumption estimate using the XPE for different
degrees of PL utilisation from 20 % to 100 % as shown in
Figure 11.
B. Adopted methodology for FPGA and MCU implementation
Next, we followed the following step by step methodology
for the FPGA and MCU implementation of the various ML
algorithms discussed earlier for temperature and temperature
+ temporal drift compensation in ISFET devices:
1) We installed the SciKit Learn (SK Learn) library on the
PYNQ Z1 board using the reference [62].
2) Next, we trained the ML models using the data set
generated from SPICE.
3) Next, we dump the trained model on the PS of PYNQ-Z1
and run the prediction. While the prediction is running,
we measure the peak power consumption of the board by
visually observing the USB power tester display.
4) We implement the ML algorithms for temperature and
temporal drift data on the PL of the PYNQ -Z1 Board
using the sklearn hardware library for PYNQ in [63].
5) We calculate the time taken for prediction using the
time.time() function in the Jupyter terminal of the PYNQ
Z1 board.
6) The performance metrics, such as MAE, MSE, RMS and
R2are calculated for the ML models using the python
code in the Jupyter terminal.
7) We note the obtained values and multiply the correspond-
ing elapsed time with peak power consumption values to
calculate the energy consumption per prediction.
8) Steps 2-6 are repeated for all the six ML algorithms
for both temperature and temperature + temporal drift
9) As the best results were achieved by RF, we choose
to implement it from scratch using Verilog and imple-
Fig. 12: Schematic of the trained RF algorithm implementation on the PL of PYNQ-Z1 board using generated using Xilinx
Vivado software for only temperature ( top) and temperature + temporal drift(bottom).
ment it using Xilinx Vivado software. We selected the
Xc7z020CLG400-1 FPGA platform (PL of PYNQ-Z1)
for implementation. The verilog code was extracted for
the trained RF model from [64].
10) As RF consists of multiple DTs, we also implemented
DT using the Xilinx Vivado Software on PL platform.
Overlays given in [65] can also be used to run DT on the
PYNQ-Z1 board directly.
11) Finally all the the results are compiled and presented in
tabular format. The PL implementation of PYNQ Z1 is
compared with the Intel Core CPU implementation in
terms of accuracy, power consumption and speed.
12) We calculate the new parameter proposed in Section V-E
called IF score for the six ML algorithms. It indicates the
best performing ML model for IoT deployment, giving
equal importance to accuracy, power consumption and
time taken for prediction for both the PS and PL of the
Recorded video and code snippets from Jupyter terminal of
the PYNQ-Z1 board for the experiments conducted on the PS
are given in the Supplementary material [66].
C. Issues with fixed point and floating point representation
In FPGA implementation, it is important to decide the num-
ber representation to be used. There are two popular methods
1) Fixed Point representation, 2) Floating point representation.
In fixed point representations, it is easier to perform arithmetic
operations but the range of the numbers and precision is
limited whereas in floating point representation (such as IEEE
754), it is difficult to perform arithmetic operations but the
range of numbers which can be represented and the precision
is very high. In our implementation of ML models on the PL,
we used 32 bits fixed point number representation with 20 bits
for fraction part. A better solution can be obtained by using
standard floating point representation, such as IEEE 754 single
precision, half precision etc., which is part of our future work.
D. Details of FPGA implementation
For LR and PR models, we used a common hardware library
from [63]. The library used 32 bit fixed point number system
with 20 bits in fraction. For LR the training:testing ratio in the
data-set was 80:20 for temperature and temporal drift compen-
sation. However, for PR the same was 95:5 due to the limited
buffer size available on the PYNQ-Z1 board for storing the
test data.We also implemented the best performing RF model
along with DT algorithms using the Xilins Vivado Software
on the PYNQ-Z1 PL platform, i.e., XC7020CLG400-1. The
Verilog code for the trained RF and DT models were extracted
using [64]. It uses 64 bit number representation with only 8 bit
for fractions. For both temperature and temperature+temporal
drift compensation, the RF had 12 estimators with a maximum
depth of 4. Figure 12 shows the schematic of the Trained
RF models for temperature and temperature+temporal drift
obtained using the Xilinx Vivado Software. For DT, in both
temperature and temperature+temporal drift compensation, the
depth was 4.
A. Temperature Drift Compensation
Table II encompasses Si3N4-based ISFET ROIC results on
various performance measures for achieving temperature drift
compensation on both the Xilinx PYNQ-Z1 board and the
Intel(R) Core(TM) i5-8250 CPU @ 1.6 GHz - Quad core.
The developed ML models achieved very high R2score (cor-
relation) on both the platforms. The strong correlation value
between the real pH value and its expected value indicates that
the temperature drift can be precisely accounted for by the ML
models. The error rates achieved with various models are very
small, with Random Forests (RF) outperforming other models.
We suspect that with a large number of data samples used for
training, MLP might outperform the classical ML algorithms.
In terms of the time taken for prediction and average power
TABLE V: Performance metrics for ML algorithms practically implemented using the PYNQ-Z1 board PL (Xc7z020CLG400-1)
and power measured using USB power tester.
Accuracy Metrics Time taken
(per prediction)
Energy consumed
(per prediction) (J)
LR 0.244 0.125 0.991 3.80E-04 6.83E-04 1351.861
PR 0.015 0.003 0.999 4.24E-04 7.70E-04 1248.598
+ Temporal
LR 1.505 3.963 0.644 3.68E-04 6.24E-04 1425.462
PR 0.695 1.001 0.907 3.67E-04 6.50E-04 1407.147
TABLE VI: Performance metrics for trained RF model imple-
mentations on Vivado on XC7Z020CLG400-1 FPGA board(
PL of PYNQ-Z1)
Accuracy Metrics Total on chip
power (W)MAE MSE R2
RF 0.598 0.546 0.961 1.13
DT 0.692 0.7248 0.948 1.13
+ Temporal
RF 0.297 0.335 0.965 0.926
DT 0.316 0.397 0.958 1.13
consumption during prediction, Linear Regression (LR) out-
performs other models on both the platforms but gives poorer
performance metrics as compared to other models. Thus, there
is a trade-off between better performance metrics with power
and time consumed for prediction.
In addition, we examine the performance of the RF model
with respect to a simple curve fitting model (we chose Linear
Regression (LR) for this purpose). Illustration Fig. 8 shows the
expected pH vs temperature curves for pH values 7 and 10.
The obtained curves are almost same for both CPU and FPGA
implementation. This is explained by results shown in Table
II, where it is seen that the performance metrics of FPGA
implementation on PYNQ board and CPU implementation are
similar for RF and LR models. The curves in blue correspond
to a random forest regressor fitted on all samples of training
data, except for a number of pH values. For example, all
samples corresponding to pH 0-5 and 9-14, i.e. except all of
the samples corresponding to pH 6-8, will be the training data
to predict pH 7. On similar lines, another model was trained
to predict pH 10.
B. Temperature-Temporal Drift Compensation
Table VII shows the performance of ML models for com-
pensating drift in Vref due to both time and temperature using
the Xilinx PYNQ-Z1 board and the Intel(R) Core(TM) i5-
8250 CPU @ 1.6 GHz - Quad core. All ML models except
LR, perform decently well for the task at hand. This drop
in performance of linear regression model can be explained
due to the non-linearity of the experiment. Even though it
consumes the least time in both platforms and least power
on FPGA board, its poor accuracy makes it unfit for imple-
mentation. Similar to experiment 1, a high R2score shows
prominent correlation between the real Vref and the predicted
Vref for particular temperature, pH and time. Random Forest
achieves better results than any other model, showing its worth
for compensating drift in voltage for Si3N4-based ISFET.
Moreover, here we examine the performance of the RF
model for temperature-temporal drift compensation with re-
spect to best model obtained from experiment 1 (only temper-
ature compensation). Illustration Fig. 9 shows the expected pH
vs time curves for pH values 7 and 10 at two temperatures (low
and high) . For training the RF regressor, for both temporal and
temperature drift compensation, it was fitted on all samples
of training data except for a number of temperatures. For
example, if the model is used to compensate Vref for pH
7 at 20oC, it was trained for all samples of pH 7 and
all temperatures except 18,19,20,21,22 oC. On similar lines,
another model was trained to predict Vref for pH 10. The
output is mapped to corresponding pH value. Similar to the
previous case, the results are nearly same for both the FPGA
and CPU implementations due to the very close performance
metrics, as shown in Table VII.
C. Time series classification
We compare the performance on 4 different RNN-based
models—1) LSTM, 2) biLSTM, 3) GRU, 4) biGRU [67]. The
number of RNN nodes in LSTM and GRU is 32 while in
biLSTM and biGRU is 16. The feed-forward NN has 200
neurons which is chosen based on grid search in the set
10,20,50,100,150,200,300,500,1000. We minimize the binary
cross-entropy loss with a batch size of 64 and run for 10
epochs. Table VIII shows biLSTM performs best amongst all
the RNN variants with minimum entropic loss and highest
accuracy of 79.8%. As the obtained accuracy of RNN is not
very high in CPU, future work is required to optimize it further
before going for FPGA implementation. As a future work,
when the data scarcity is not a limitation, the classification
task can be effortlessly adapted to a regression task where a
sequence of voltages can predict precise pH of the solution
over a wide range.
D. Comparison of FPGA, MCU and CPU implementation
In this section we compare the FPGA, MCU and CPU
implementation of the various ML algorithms in terms of
three parameters, i.e., accuracy, speed and power consumption.
Here, FPGA implementation is on the PS+PL of the PYNQ-Z1
board, MCU implementation is only on the PYNQ-Z1 board
and CPU implementation is on the Intel Quad core processor.
1) Accuracy: Accuracy is measured in terms of perfor-
mance metrics, i.e., Mean Absolute Error (MAE), Mean
Squared Error (MSE) and Correlation (R2). For an accurate
model, MAE and MSE should be very low and R2should
be very high. We see that in Tables II and VII, V and VI
that the accuracy of all the three, the CPU, MCU and FPGA
implementation are nearly the same. FPGA has slightly lower
accuracy but that can be easily overcome by using more
number of bits and floating point number representation.
TABLE VII: Comparison of results for experiments on the Xilinx PYNQ-Z1 board(PS) and the Intel core CPU for temperature + temporal drift compensation.
Model Metrics for Xilinx PYNQ-Z1 Board(PS) Metrics for Intel(R) Core(TM) i5-8250U
CPU @ 1.6 GHz - Quad core
(W) MAE MSE R2Time
LR 1.4217 3.5449 0.642554 0.0051 2.485 1.4217 3.5449 0.642554 0.0009 9.6
PR 0.0577 0.0220 0.997777 0.0751 2.743 0.0581 0.0220 0.997776 0.0069 11.6
DT 0.0057 0.0005 0.999943 0.0229 2.847 0.0057 0.0005 0.999950 0.0019 5.5
RF 0.0017 0.0003 0.999965 0.4405 2.768 0.0018 0.0003 0.999966 0.0419 6.5
MLP 0.0552 0.0151 0.998472 0.0994 2.599 0.0552 0.0151 0.998473 0.0049 12.3
SVR 0.0704 0.8175 0.991757 60.636 2.644 0.0705 0.0818 0.991753 0.9520 13.4
TABLE VIII: Comparison of results obtained for different
RNN on CPU for Temperature Drift Compensation.
RNN-model Loss Accuracy
LSTM 0.4135 79.20%
GRU 0.4061 79.00%
BiLSTM 0.4063 79.80%
BiGRU 0.486 79.20%
2) Speed: Speed refers to the time duration ML model takes
to predict the pH value. The lesser is the time taken, faster is
the speed. From Tables II and VII and V we observe that
the speed is much better for the CPU implementation in both
temperature and temperature+temporal compensation tasks as
compared to MCU implementation.It should be noted that the
FPGA implementation is faster than both in most cases. Hence,
FPGA helps in achieving hardware acceleration.
3) Power Consumption: Power is measured while the ML
models are running on the FPGA, MCU and the CPU. The
lesser is the power consumption, better is the model. Here, for
both the PS and PL, we measure the entire power consumption
by the PYNQ-Z1 board with the help of a USB power tester
as shown in Fig. 7, while for the CPU, we measure the total
power consumed by the CPU package, CPU cores , CPU
graphics and CPU DRAM with the help of the Open Hardware
Monitor application. It is important to note that the PYNQ-
Z1 board has many on-board peripherals such as audio/video,
USB, Ethernet, UART, CAN controllers, etc., which will not
be required for final FPGA deployment in an IoT or industrial
scenario for ISFETs, hence the power consumed by the ML
models solely on the FPGA will be much lesser than what
is given, i.e., the given power consumption can be treated
as worst case. However, for the CPU, we are measuring the
power consumed by the CPU itself and not other peripherals
of the computer, hence it is the best case. It can be seen
from II and VII that the the worst case power consumed
by FPGA implementation is much lesser than the best case
power consumption by CPU. Similarly, from Table V it is
seen that the FPGA implementation provides a lot of hardware
acceleration as compared to MCU or software implementation.
Additionally power consumption can also be measured using
Xilinx Power estimator or Xilinx Vivado Design Suite.
IoT and industrial scenarios such as biomedical applications,
environmental monitoring, agriculture, food industry, etc.,
where ISFETs are used may not require high speed but require
low power consumption and high accuracy. Hence FPGA
Temperature compensation Temperature+Temporal compnesation
Score ML models
Fig. 13: IF score comparison of the various ML models for
both temperature and temperature+temporal compensation.
implementation is the best approach for such applications.
In this study, we used the PYNQ framework to implement
the algorithms on FPGA but designs implemented using these
are not optimised for performance parameters such as speed,
power consumption, area etc. Hence, customized RTL and/or
ASIC implementations would consume much less power and
would be preferred for eventual large-scale fabrication. It will
give better results than the results obtained in this study. Table
X summarises the pros and cons of both the implementations.
After concluding that the FPGA implementation is better than
CPU or MCU for deployment of ML model in IoT and
industrial application of ISFETs, the next step is to evaluate
the best ML model to be deployed on FPGA. This is discussed
in the next subsection.
E. Evaluation of ML models for FPGA and MCU hardware
for IoT deployment
As discussed previously, certain ML models such as RF
give very high accuracy whereas some ML models such as
LR consume very less power and time but give relatively
poorer accuracy and some models such as Decision Trees
(DT) give very moderate accuracy and consume moderate
power and time. Hence, choosing the optimum ML model
for implementation on FPGA/MCU needs further analysis. In
order to solve this problem, we define a new parameter called
Implementation Factor (IF) which gives weights to accuracy,
speed or time taken per-prediction and energy consumed per-
prediction to calculate a final score, which allows the ML
models to be compared. The weights are given according to
the scheme below :-
1) Accuracy: 33 % ( 11% - MAE, 11% - MSE , 11% R2)
2) Speed or Time taken per prediction : 33 %
TABLE IX: Final comparison of various ML algorithms based on the accuracy of prediction, time taken for prediction, and energy consumed per prediction.
The IF score gives the overall performance score on the basis of which random forests is best for implementation on FPGA platform for both temperature
and temperature+temporal compensation.
Temperature Compensation Temperature + Temporal Compensation
Accuracy Time
Energy per
Accuracy Time
Energy per
LR 0.234 1.237e-01 0.99121 0.004 9.921 117.22 1.4217 3.5449 0.642554 0.0051 12.817 89.91
PR 0.071 1.118e-02 0.99916 0.013 36.654 44.39 0.0577 0.0220 0.997777 0.0751 205.906 13.01
DT 0.058 6.713e-03 0.99952 0.023 63.698 37.66 0.0057 0.0005 0.999943 0.0229 65.349 235.33
RF** 0.007 8.947e-05 0.99999 0.272 743.599 1246.07 0.0017 0.0003 0.999965 0.4405 1219.228 378.85
MLP 0.025 9.809e-04 0.99993 0.035 88.908 129.74 0.0552 0.0151 0.998472 0.0994 258.366 13.96
SVR 0.074 1.060e-02 0.99925 73.878 196777.458 11.96 0.0704 0.8175 0.991757 60.636 160303.618 1.81
TABLE X: Overall comparison of FPGA and CPU implemen-
tation of ML models for ISFET IoT and Industrial implemen-
requirement for
(PS of
Accuracy Essential 3= =
Speed Non-Essential
but helpful 3 7 3
Power Essential 3 3 7
3) Energy consumed per prediction : 33 %
The IF score for each ML model can be calculated using
the formula below:-
IF =0.11
MAE +0.11
MSE + 0.11 ×R2+0.33
where, MAE, MSE and R2are the performance metrics
for accuracy of the ML model, T is the time taken by the
ML model per prediction in SI unit, i.e., seconds and E is
the energy consumed by the ML model per prediction in SI
unit, i.e., Joules. The higher is the IF score, better is the
ML model for implementation on FPGA/MCU, giving equal
weights to all the three parameters, i.e., accuracy, speed and
energy consumption. To the best of our knowledge, there is no
parameter defined in literature till now which considers all the
three factors: energy consumed, speed and accuracy to give a
single score to evaluate the best ML algorithm for FPGA/MCU
implementation, which necessitates the creation of new IF
parameter. Existing performance metrics for ML algorithms
consider only one of these metrics, for example, F score and
F1 score only gives information about the accuracy. Hence,
there is a need to define a score which accounts collectively
for all three factors. Moreover, the weights assigned to the
factors may change according to type of application. For
example, in bio-medical and research applications, accuracy
will have more weight while for long term IoT deployment
such as environmental monitoring, power consumption will
have more weight. In this study, as a general case, we have
given equal weights to all the three factors. On the basis of
IF score calculation, we see from Table IX that RF has the
highest IF for both temperature and temperature + temporal
compensation. This is also graphically shown in Fig. 13.
F. Advantage of drift compensation using ML implemented on
FPGA/MCU platforms
As compared to conventional drift compensation methods
in ISFETs, implementing ML on FPGA/MCUs has several
advantages. Drift compensation in ISFETs may be done by
other compensation techniques such as ISFET sensor design
or simple modeling. However, the temporal drift compensation
is a complex problem, where the sensor needs to compensate
for the drift temporally for various pH values. In order to
learn the sensor behavior, ML models are suitable to capture
the drift patterns in the acidic and basic pH solutions, and
predict more accurate measurements in comparison to simple
modeling techniques. Moreover, the ML models developed
in this work are designed to obtain early estimation of pH
even before the VRef stabilizes. Thus, its prediction can be
much faster. We have discussed pros and cons of using ML
models implemented on FPGA platforms for sensor drift com-
pensation in comparison with drift compensation by sensor
design or using simple modeling, as listed in Table XI. Drift
compensation using ML implemented on FPGA platforms is
suitable for Industrial IoT applications, as it is re-configurable
and cheaper compared to other techniques.
In this work, we present the modeling of non-idealities in
Si3N4-gate ISFET due to temperature and temporal variations
using SPICE and the drift compensation using state-of-the-
art ML models which have been implemented on FPGA.
The SPICE macromodel is used as a subcircuit block in a
CVCC readout topology to generate temperature and temporal
drift data for a wide temperature (15-65 oC) and pH (0-14)
range. The data is used to train state-of-the-art ML models
to compensate the temperature and temporal drift and it was
found that Random Forest (RF) technique outperforms other
ML models. Seeing promising results on CPU, the ML models
were implemented in the PYNQ Z1 board that houses both
an ARM processor and an FPGA. It is a power and memory
constrained environment suitable for IoT applications. We used
hardware-software co-design based tool, PYNQ framework
to implement the design but customized RTL and/or ASIC
implementations would consume much less power and would
be preferred for eventual large-scale fabrication. We also
propose a novel Implementation Factor(IF) which gives equal
weights to accuracy, speed and energy consumption while
running on FPGA and assigns the ML models an overall score.
A higher IF score indicates that the ML model is better suited
TABLE XI: Comparison of drift compensation by ML implemented on platforms like FPGA/MCUs vs. other drift compensation
Criteria Compensation by sensor design Compensation using simple modelling Compensation using post data processing
techniques like ML on platforms like FPGA
Very complex as precise operations need to be performed for adjusting
appropriate doping to compensate the temporal drift at
fabrication level by ion-implantation. Also, it is highly vulnerable to
process variation which may adversely impact the compensation [51].
Difficult to implement the modeling based compensation in
hardware as modeling software like SPICE and MATLAB
use complex mathematical tools and equations which
are difficult to be implemented in IoT Embedded systems [68, 69].
ML models can be easily implemented in
external hardware processing chips like
FPGAs and trained to learn the drift
patterns using the well-established ML models.
Size No external chip based compensation
required, hence overall system is small in size.
System size increases due to involvement
of large and bulky computing systems. System size slightly increases due to an additional chip.
Cost Very costly as there has to be customization at the fabrication level. Less costly Less costly
Time Fast, as there is compensation at the point of sensing. Slow, as the data will have to be transferred to
computing systems and calculated by complex softwares. Fast, as it can have high level of parallelism.
Power Not much power consumption. High power consumption by computing systems. Low power consumption.
Overall IoT
deployment scenario
Feasible for Real-Time IoT Applications but
will be very costly and cannot be reconfigured.
Implementation is challenging for real-time IoT
applications and it may not be reconfigurable.
Feasible for real-time IoT applications and
it will be cheap and reconfigurable
for FPGA implementation. Based on the evaluation, we find
that RF has a higher IF score for both temperature and tem-
perature+temporal drift compensation in both the PS and PL
implementations on the PYNQ-Z1 board. This paper presents a
promising application of ML in drift compensation of ISFET-
based pH sensor, which has a huge scope for further research
to improve the robustness of smart chemical sensors, which
can be deployed for various Industrial IoT applications such as
biomedical diagnosis, environmental monitoring, agriculture,
wearable sensing, food industry and horticulture.
The authors are thankful to Director, CSIR-CEERI, Pilani
for the generous support. They would like to thank all the
scientific and technical staff of CSIR-CEERI, Pilani for the
discussions and technical support. Financial support from
CSIR, New Delhi is gratefully acknowledged.
[1] A. Mehta, H. Ahuja, N. Sahu, R. Bhardwaj, S. Srivastava,
and S. Sinha, “Machine learning techniques for perfor-
mance enhancement of si 3 n 4-gate isfet ph sensor,” in
2020 IEEE 17th India Council International Conference
(INDICON). IEEE, 2020, pp. 1–7.
[2] M. J. Madou and S. R. Morrison, Chemical sensing with
solid state devices. Elsevier, 2012.
[3] B. R. Eggins, Chemical sensors and biosensors. John
Wiley & Sons, 2008, vol. 28.
[4] L. Manjakkal, D. Szwagierczak, and R. Dahiya, “Metal
Oxides based Electrochemical pH Sensors: Current
Progress and Future Perspectives,Progress in Materials
Science, p. 100635, 2019.
[5] M. Kaisti, “Detection Principles of Biological and
Chemical FET Sensors,” Biosensors and Bioelectronics,
vol. 98, pp. 437–448, 2017.
[6] N. Moser, T. S. Lande, C. Toumazou, and P. Georgiou,
“Isfets in cmos and emergent trends in instrumentation:
A review,” IEEE Sensors Journal, vol. 16, no. 17, pp.
6496–6514, 2016.
[7] G. Seo, G. Lee, M. J. Kim, S.-H. Baek, M. Choi, K. B.
Ku, C.-S. Lee, S. Jun, D. Park, H. G. Kim et al.,
“Rapid detection of covid-19 causative virus (sars-cov-2)
in human nasopharyngeal swab specimens using field-
effect transistor-based biosensor,” ACS nano, vol. 14,
no. 4, pp. 5135–5142, 2020.
[8] J. Zimmer, “Lab-on-a-Chip (LoC) COVID-
19 Test Advances to Clinical Trials,” https:
test-advances-to- clinical-trials, 2020, [Online;accessed
[9] D. Sadighbayan and E. Ghafar-Zadeh, “Portable sensing
devices for detection of covid-19: A review,IEEE
Sensors Journal, vol. 21, no. 9, pp. 10219–10 230, 2021.
[10] T. Shaffaf and E. Ghafar-Zadeh, “Covid-19 diagnostic
strategies. part i: Nucleic acid-based technologies,Bio-
engineering, vol. 8, no. 4, p. 49, 2021.
[11] X. Huang, H. Yu, X. Liu, Y. Jiang, M. Yan, and D. Wu,
A dual-mode large-arrayed cmos isfet sensor for accu-
rate and high-throughput ph sensing in biomedical diag-
nosis,” IEEE Transactions on Biomedical Engineering,
vol. 62, no. 9, pp. 2224–2233, 2015.
[12] A. M. Dinar, A. M. Zain, and F. Salehuddin, “Cmos
isfet device for dna sequencing: Device compensation,
application requirements and recommendations,” Int. J.
Appl. Eng. Res, vol. 12, no. 21, pp. 11 015–11 028, 2017.
[13] P. Zarkesh-Ha, J. Edwards, and P. Szauter, “Avalanche
isfet: A highly sensitive ph sensor for genome sequenc-
ing,” in 2015 IEEE Biomedical Circuits and Systems
Conference (BioCAS). IEEE, 2015, pp. 1–4.
[14] N. Miscourides and P. Georgiou, “Impact of technology
scaling on isfet performance for genetic sequencing,”
IEEE Sensors Journal, vol. 15, no. 4, pp. 2219–2226,
[15] A. Ganguli, Y. Watanabe, M. T. Hwang, J.-C. Huang,
and R. Bashir, “Robust label-free microrna detection
using one million isfet array,Biomedical microdevices,
vol. 20, no. 2, pp. 1–10, 2018.
[16] G. Alexandrou, N. Moser, J. Rodriguez-Manzano,
P. Georgiou, J. Shaw, C. Coombes, C. Toumazou, and
M. Kalofonou, “Detection of breast cancer esr1 p.e380q
mutation on an isfet lab-on-chip platform,” in 2020
IEEE International Symposium on Circuits and Systems
(ISCAS), 2020, pp. 1–5.
[17] M. Lehmann, W. Baumann, M. Brischwein, H.-J. Gahle,
I. Freund, R. Ehret, S. Drechsler, H. Palzer, M. Kleintges,
U. Sieben et al., “Simultaneous measurement of cellular
respiration and acidification with a single cmos isfet,”
Biosensors and Bioelectronics, vol. 16, no. 3, pp. 195–
203, 2001.
[18] J.-H. Jeon and W.-J. Cho, “Ultrasensitive coplanar dual-
gate isfets for point-of-care biomedical applications,”
ACS omega, vol. 5, no. 22, pp. 12 809–12 815, 2020.
[19] M. Kalofonou and C. Toumazou, “Semiconductor tech-
nology for early detection of dna methylation for cancer:
From concept to practice,” Sensors and Actuators B:
Chemical, vol. 178, pp. 572–580, 2013.
[20] J. Malinowski and E. J. Geiger, “Development of a
wireless sensor network for algae cultivation using isfet
ph probes,” Algal Research, vol. 4, pp. 19–22, 2014.
[21] G. A. Taylor, C. Parra, H. Carrillo, and A. Mouazen,
A decision framework reference for isfet sensor-based
electronic systems design for agriculture industry appli-
cations,” in 2020 IEEE 17th India Council International
Conference (INDICON). IEEE, 2020, pp. 1–6.
[22] D. Walvoort, J. Bouma, P. Peters, and J. de Gruijter, “Us-
ing isfets for proximal sensing in precision agriculture,”
in Proceedings of the Fifth International Conference on
Precision Agriculture, Madison USA, 2000, 2000, pp.
[23] G. A. Taylor, H. B. Torres, F. Ruiz, M. N. Mar´
D. M. Chaves, L. T. Arboleda, C. Parra, H. Carrillo,
and A. M. Mouazen, “ph measurement iot system for
precision agriculture applications,” IEEE Latin America
Transactions, vol. 17, no. 05, pp. 823–832, 2019.
[24] C. Jimenez-Jorquera, J. Orozco, and A. Baldi, “Isfet
based microsensors for environmental monitoring,Sen-
sors, vol. 10, no. 1, pp. 61–83, 2010.
[25] M. Joly, L. Mazenq, M. Marlet, P. Temple-Boyer,
C. Durieu, and J. Launay, “Multimodal probe based on
isfet electrochemical microsensors for in-situ monitoring
of soil nutrients in agriculture,” in Multidisciplinary
Digital Publishing Institute Proceedings, vol. 1, no. 4,
2017, p. 420.
[26] S. J. Birrell and J. W. Hummel, “Real-time multi isfet/fia
soil analysis system with automatic sample extraction,
Computers and Electronics in Agriculture, vol. 32, no. 1,
pp. 45–67, 2001.
[27] ——, “Membrane selection and isfet configuration eval-
uation for soil nitrate sensing,” Transactions of the ASAE,
vol. 43, no. 2, p. 197, 2000.
[28] M. Joly, L. Mazenq, M. Marlet, P. Temple-Boyer,
C. Durieu, and J. Launay, “All-solid-state multimodal
probe based on isfet electrochemical microsensors for
in-situ soil nutrients monitoring in agriculture,” in 2017
19th International Conference on Solid-State Sensors,
Actuators and Microsystems (TRANSDUCERS). IEEE,
2017, pp. 222–225.
[29] V. I. Adamchuk, J. W. Hummel, M. Morgan, and
S. Upadhyaya, “On-the-go soil sensors for precision
agriculture,” Computers and electronics in agriculture,
vol. 44, no. 1, pp. 71–91, 2004.
[30] N. F. Starodub, “Efficiencies of biosensors in envi-
ronmental monitoring,” Book: Portable biosensing of
food toxicants and environmental pollutants”, Series in
sensors, Ed. By Nikolelis DP and et al., CRC Press,
Taylor&Francis Group, pp. 487–514, 2013.
[31] K. McLaughlin, A. Dickson, S. B. Weisberg, K. Coale,
V. Elrod, C. Hunter, K. S. Johnson, S. Kram, R. Kudela,
T. Martz et al., “An evaluation of isfet sensors for coastal
ph monitoring applications,” Regional Studies in Marine
Science, vol. 12, pp. 11–18, 2017.
[32] N. Jaffrezic-Renault, A. Senillou, C. Martelet, K. Wan,
and J. Chovelon, “Isfet microsensors for the detection
of pollutants in liquid media,” Sensors and Actuators B:
Chemical, vol. 59, no. 2-3, pp. 154–164, 1999.
[33] E. M. Briggs, S. Sandoval, A. Erten, Y. Takeshita,
A. C. Kummel, and T. R. Martz, “Solid state sensor for
simultaneous measurement of total alkalinity and ph of
seawater,ACS sensors, vol. 2, no. 9, pp. 1302–1309,
[34] E. Voulgari, F. Krummenacher, and M. Kayal, “Antigone:
A programmable energy-efficient current digitizer for an
isfet wearable sweat sensing system,” Sensors, vol. 21,
no. 6, p. 2074, 2021.
[35] S. Nakata, T. Arie, S. Akita, and K. Takei, “Wearable,
flexible, and multifunctional healthcare device with an
isfet chemical sensor for simultaneous sweat ph and skin
temperature monitoring,” ACS sensors, vol. 2, no. 3, pp.
443–448, 2017.
[36] M. Douthwaite, E. Koutsos, D. C. Yates, P. D. Mitcheson,
and P. Georgiou, “A thermally powered isfet array for on-
body ph measurement,” IEEE transactions on biomedical
circuits and systems, vol. 11, no. 6, pp. 1324–1334, 2017.
[37] F. Bellando, L. J. Mele, P. Palestri, J. Zhang, A. M.
Ionescu, and L. Selmi, “Sensitivity, noise and resolution
in a beol-modified foundry-made isfet with miniaturized
reference electrode for wearable point-of-care applica-
tions,” Sensors, vol. 21, no. 5, p. 1779, 2021.
[38] A. Vilouras and R. Dahiya, “Ultra-thin chips with
current-mode isfet array for continuous monitoring of
body fluids ph,” in 2021 IEEE International Symposium
on Circuits and Systems (ISCAS). IEEE, 2021, pp. 1–5.
[39] N. F. Starodub, W. Torbicz, V. M. Starodub, M. I. Kanjuk,
and K. S. Ternovoj, “Enzymatic sensors based on the
isfets for determination of heavy metal ions in some
vegetables,” in Proceedings of International Solid State
Sensors and Actuators Conference (Transducers’ 97),
vol. 2. IEEE, 1997, pp. 1383–1384.
[40] V. Volotovsky and N. Kim, “Determination of glucose,
ascorbic and citric acids by two-isfet multienzyme sen-
sor,Sensors and Actuators B: Chemical, vol. 49, no. 3,
pp. 253–257, 1998.
[41] N. Starodub, Y. M. Shirshov, W. Torbicz, N. Kanjuk,
V. Starodub, and A. Kukla, “Biosensors for in field
measurements: Optimisation of parameters to control
phosphororganic pesticides in water and vegetables,” in
Biosensors for direct monitoring of environmental pollu-
tants in field. Springer, 1998, pp. 209–219.
[42] H. Ito, S. Wakida et al., “Nitrate ion determination
of vegetables using a portable isfet-nitrate ion sensor.”
Bulletin of the National Institute of Vegetable and Tea
Science, no. 15, pp. 11–17, 2016.
[43] H. Van den Vlekkert, J. Kouwenhoven, and
A. Van Wingerden, “Application of isfets in closed-loop
systems for horticulture,” in I International Workshop
on Sensors in Horticulture 304, 1991, pp. 309–320.
[44] H. Van den Vlekkert, U. Verkerk, P. Van der Wal,
A. Van Wingerden, D. Reinhoudt, J. Haak, G. Honig, and
H. Holterman, “Multi-ion sensing device for horticultural
application based upon chemical modification and special
packaging of isfets,” Sensors and Actuators B: Chemical,
vol. 6, no. 1-3, pp. 34–37, 1992.
[45] J. Artigas, A. Beltran, C. Jimenez, A. Baldi, R. Mas,
C. Domınguez, and J. Alonso, “Application of ion sensi-
tive field effect transistor based sensors to soil analysis,
Computers and electronics in agriculture, vol. 31, no. 3,
pp. 281–293, 2001.
[46] I. Amemiya, H. Yagi, and T. Sakai, “An isfet-based
nutrient sensor for rockwool culture,IFAC Proceedings
Volumes, vol. 24, no. 11, pp. 361–366, 1991.
[47] T. H. Gieling and H. Van Den Vlekkert, “Application of
isfets in closed-loop systems for greenhouses,” Advances
in Space Research, vol. 18, no. 4-5, pp. 135–138, 1996.
[48] N. Sahu, R. Bhardwaj, H. Shah, R. Mukhiya, R. Sharma,
and S. Sinha, “Towards development of an isfet-based
smart ph sensor: Enabling machine learning for drift
compensation in iot applications,” IEEE Sensors Journal,
[49] S. Sinha, N. Sahu, R. Bhardwaj, H. Ahuja, R. Sharma,
R. Mukhiya, and C. Shekhar, “Modeling and Simulation
of Temporal and Temperature Drift for the Development
of an Accurate ISFET SPICE Macromodel,” Journal of
Computational Electronics, vol. 19, no. 1, pp. 367–386,
[50] P. Barabash, “Analysis of the Threshold Voltage and
its Temperature Dependence in Electrolyte-Insulator-
Semiconductor Field-Effect Transistors (EISFET’s),”
IEEE Transactions on Electron Devices, vol. 34, no. 6,
pp. 1271–1282, 1987.
[51] A. Elyasi, M. Fouladian, and S. Jamasb, “Counteract-
ing Threshold-Voltage Drift in Ion-Selective Field Ef-
fect Transistors (ISFETs) Using Threshold-Setting Ion
Implantation,” IEEE Journal of the Electron Devices
Society, vol. 6, pp. 747–754, 2018.
[52] V. Pachauri and S. Ingebrandt, “Biologically sensitive
field-effect transistors: from isfets to nanofets,Essays
in biochemistry, vol. 60, no. 1, pp. 81–90, 2016.
[53] N. Zimmerman, A. A. Presto, S. P. Kumar, J. Gu, A. Hau-
ryliuk, E. S. Robinson, A. L. Robinson, and R. Subrama-
nian, “A Machine Learning Calibration Model using Ran-
dom Forests to Improve Sensor Performance for Lower-
Cost Air Quality Monitoring.” Atmospheric Measurement
Techniques, vol. 11, no. 1, 2018.
[54] A. Moraru, M. Pesko, M. Porcius, C. Fortuna, and
D. Mladenic, “Using machine learning on sensor data,”
Journal of computing and information technology,
vol. 18, no. 4, pp. 341–347, 2010.
[55] S. Sinha, R. Bhardwaj, N. Sahu, H. Ahuja, R. Sharma,
and R. Mukhiya, “Temperature and Temporal Drift Com-
pensation for Al2O3-gate ISFET-based pH Sensor using
Machine Learning Techniques,Microelectronics Jour-
nal, vol. 97, p. 104710, 2020.
[56] S. Naimi, B. Hajji, Y. Habbani, I. Humenyuk, J. Lau-
nay, and P. Temple-Boyer, “Modeling of the ph-isfet
thermal drift,” in 2009 International Conference on
Microelectronics-ICM. IEEE, 2009, pp. 288–291.
[57] S. Jamasb, S. Collins, and R. L. Smith, “A physical model
for drift in ph isfets,” Sensors and Actuators B: Chemical,
vol. 49, no. 1-2, pp. 146–155, 1998.
[58] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel,
B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer,
R. Weiss, V. Dubourg et al., “Scikit-learn: Machine
Learning in Python,” Journal of Machine Learning Re-
search, vol. 12, no. Oct, pp. 2825–2830, 2011.
[59] E. Alpaydin, Introduction to Machine Learning. MIT
Press, 2020.
[60] Xilinx, “Pynq: Python productivity,” https:http://www., [Online;accessed 18-May-2021].
[61] Digilent, “Pynq-z1 refernece manual,” https:
z1/reference-manual, [Online;accessed 25-July-2021].
[62] I. Andrei, “Pynq-z1 error use pip and install sklearn,”
install-sklearn/129, [Online;accessed 26-July-2021].
[63] sfox14, “Github: sfox/pynq sklearn,”
sfox14/pynq-sklearn, [Online;accessed 15-Aug-2021].
[64] B. Snyder, “Github: johnbensnyder/fpgarandomforest, random forest,
[Online;accessed 15-Aug-2021].
[65] Y. Qu, “Github: yunqu/pynq-decisiontree,” https://github.
com/yunqu/PYNQ-DecisionTree, [Online;accessed 15-
[66] S. Sinha, N. Sahu, R. Bhardwaj, A. Mehta, and H. Ahuja,
“Machine learning on fpga for robust si3n4-gate isfet
ph sensor in industrial iot applications,” 2021. [Online].
[67] W. Yu, I. Y. Kim, and C. Mechefske, “Analysis of
different rnn autoencoder variants for time series classi-
fication and machine prognostics,” Mechanical Systems
and Signal Processing, vol. 149, p. 107322, 2021.
[68] P. K. Chan and D. Chen, “A cmos isfet interface circuit
with dynamic current temperature compensation tech-
nique,” IEEE Transactions on Circuits and Systems I:
Regular Papers, vol. 54, no. 1, pp. 119–129, 2007.
[69] W.-Y. Chung, C.-H. Yang, M.-C. Wang, D. G. Pi-
janowska, and W. Torbicz, “Temperature compensation
electronics for isfet readout applications,” in IEEE Inter-
national Workshop on Biomedical Circuits and Systems,
2004. IEEE, 2004, pp. S1–6.
... Table 1 compared the [16] and our work, which reduces the number of SAs and area overhead in the similar reading bandwidth. It is essential to accurate the execution time results that can avoid significant roadblocks, and the energy consumption is further evaluated with Xilinx Power Estimator (XPE) [1] in various current studies [13,25,27]. Note that the LUT and FF are the most mentioned resources in the related hardware architecture, we do not apply other hardware resources for a relatively fair comparison, such as the DSP. ...
Full-text available
Monitoring of pH is crucial for several chemical and biochemical processes. ISFET (Ion-Sensitive Field-Effect Transistors)-based pH sensors are promising candidates for pH monitoring applications. However, ISFET devices are prone to temporal and temperature drifts, which severely affects the precision of pH measurements. In this work, we collect experimental data of temporal and temperature drifts in an ISFET sensor to formulate an accurate SPICE macro model, incorporating both temporal and temperature non-idealities. The developed macro model is utilized for generating training data for state-of-the-art machine learning models for drift compensation, with a primary focus on the temporal characteristics. We utilize recurrent neural networks (RNNs) to model the temporal characteristics of ISFET, and thus, compensate the non-ideality. The sensor data is collected in various pH buffer solutions and a data set of sequences containing time-dependent voltage readings are generated by the device and the RNNs are trained to learn the crucial features from the data and map them to the precise pH of the solution. We compare two variants of RNNs, i.e. LSTM (long short-term memory) and GRU (gated recurrent unit), and their bidirectional low computational cost variants - biLSTM and biGRU. Each model is tested in a memory-constrained environment with the availability of a 32-bit and 64-bit floating-point number. Empirically, we find biLSTMs to perform best, where the achieved root mean square error (RMSE) between the model predicted pH and the true pH of the test solution is less than 0.212 pH, with an average RMSE of 0.126 pH. For temperature drift compensation, we collect data for four different temperatures and adapt well-established MLPs (Multi-layer Perceptrons) to compensate the intrinsic temperature drift in the sensor. We observe an average RMSE of the model predicted pH to the true pH to be less than 0.286 pH. The developed RNN models were implemented on Xilinx ZCU104 FPGA development kit using PYNQ framework, which demonstrates low power consumption. The proposed framework establishes the efficacy of Machine Learning (ML) techniques for drift compensation in ISFET-based pH sensors for deployment in IoT applications.
Full-text available
The novel Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) has caused respiratory infection, resulting in more than two million deaths globally and hospitalizing thousands of people by March 2021. A considerable percentage of the SARS-CoV-2 positive patients are asymptomatic or pre-symptomatic carriers, facilitating the viral spread in the community by their social activities. Hence, it is critical to have access to commercialized diagnostic tests to detect the infection in the earliest stages, monitor the disease, and follow up the patients. Various technologies have been proposed to develop more promising assays and move toward the mass production of fast, reliable, cost-effective, and portable PoC diagnostic tests for COVID-19 detection. Not only COVID-19 but also many other pathogens will be able to spread and attach to human bodies in the future. These technologies enable the fast identification of high-risk individuals during future hazards to support the public in such outbreaks. This paper provides a comprehensive review of current technologies, the progress in the development of molecular diagnostic tests, and the potential strategies to facilitate innovative developments in unprecedented pandemics.
Full-text available
This article describes the design and the characterization of the ANTIGONE (ANalog To dIGital cONvErter) ASIC (Application Specific Integrated Circuit) built in AMS 0.35 m technology for low dc-current sensing. This energy-efficient ASIC was specifically designed to interface with multiple Ion-Sensitive Field-Effect Transistors (ISFETs) and detect biomarkers like pH, Na+, K+ and Ca2+ in human sweat. The ISFET-ASIC system can allow real-time noninvasive and continuous health monitoring. The ANTIGONE ASIC architecture is based on the current-to-frequency converter through the charge balancing principle. The same front-end can digitize multiple currents produced by four sweat ISFET sensors in time multiplexing. The front-end demonstrates good linearity over a dynamic range that spans from 1 pA up to 500 nA. The consumed energy per conversion is less than 1 J. The chip is programmable and works in eight different modes of operation. The system uses a standard Serial Peripheral Interface (SPI) to configure, control and read the digitally converted sensor data. The chip is controlled by a portable device over Bluetooth Low Energy (BLE) through a Microcontroller Unit (MCU). The sweat sensing system is part of a bigger wearable platform that exploits the convergence of multiparameter biosensors and environmental sensors for personalized and preventive healthcare.
Full-text available
Ion-sensitive field-effect transistors (ISFETs) form a high sensitivity and scalable class of sensors, compatible with advanced complementary metal-oxide semiconductor (CMOS) processes. Despite many previous demonstrations about their merits as low-power integrated sensors, very little is known about their noise characterization when being operated in a liquid gate configuration. The noise characteristics in various regimes of their operation are important to select the most suitable conditions for signal-to-noise ratio (SNR) and power consumption. This work reports systematic DC, transient, and noise characterizations and models of a back-end of line (BEOL)-modified foundry-made ISFET used as pH sensor. The aim is to determine the sensor sensitivity and resolution to pH changes and to calibrate numerical and lumped element models, capable of supporting the interpretation of the experimental findings. The experimental sensitivity is approximately 40 mV/pH with a normalized resolution of 5 mpH per µm2, in agreement with the literature state of the art. Differences in the drain current noise spectra between the ISFET and MOSFET configurations of the same device at low currents (weak inversion) suggest that the chemical noise produced by the random binding/unbinding of the H+ ions on the sensor surface is likely the dominant noise contribution in this regime. In contrast, at high currents (strong inversion), the two configurations provide similar drain noise levels suggesting that the noise originates in the underlying FET rather than in the sensing region.
The coronavirus pandemic is the most challenging incident that people have faced in recent years. Despite the time-consuming and expensive conventional methods, point-of-care diagnostics have a crucial role in deterrence, timely detection, and intensive care of the disease’s progress. Hence, this detrimental health emergency persuaded researchers to accelerate the development of highly-scalable diagnostic devices to control the propagation of the virus even in the least developed countries. The strategies exploited for detecting COVID-19 stem from the already designed systems for studying other maladies, particularly viral infections. The present report reviews not only the novel advances in portable diagnostic devices for recognizing COVID-19, but also the previously existing biosensors for detecting other viruses. It discusses their adaptability for identifying surface proteins, whole viruses, viral genomes, host antibodies, and other biomarkers in biological samples. The prominence of different types of biosensors such as electrochemical, optical, and electrical for detecting low viral loads have been underlined. Thus, it is anticipated that this review will assist scientists who have embarked on a competition to come up with more efficient and marketable in-situ test kits for identifying the infection even in its incubation time without sample pretreatment. Finally, a conclusion is provided to highlight the importance of such an approach for monitoring people to combat the spread of such contagious diseases.
Conference Paper
This paper presents the performance enhancement of Si 3 N 4 -gate Ion-Sensitive Field-Effect Transistor (IS-FET) based pH sensor using machine learning (ML) techniques. The temporal and temperature characteristics of the ISFET device are modeled using SPICE tool. The developed macromodel incorporates the electrochemical and device parameters, which enhances the robustness of the model in order to produce accurate characteristics of ISFET device over a wide temperature range as well as long term usage. The ISFET readout circuit tends to show a drift in the measured pH values with temperature and time variations.To make the device functioning more accurate, we incorporate state-of-the-art models based on ML techniques. We show how such auxiliary ML models reduce the effects of most common undesired ambience variations to get consistent output from CVCC (Constant Voltage Constant Current) Read Out Integrated Circuit (ROIC). We first simulate the quality data from the CVCC ROIC developed using ISFET macromodel. Next, we learn the auxiliary ML model parameters focused on capturing ROIC ambience dependence. This includes learning how individual ambient factors and their correlation impact the ROIC functioning. To make the compensations effective for a wide range of practical scenarios, we carry out compensation on a wide range of data values with temperature ranging from 15-65 o C and time ranging from 0-50 hours. Over this range, we compare our various ML models and observe that Random Forest is the most efficient in terms of lowest error rate to carry out the temperature and temporal drift compensation of pH values obtained from ISFET-based pH sensor circuitry.
Recurrent neural network (RNN) based autoencoders, trained in an unsupervised manner, have been widely used to generate fixed-dimensional vector representations or embed-dings for varying length multivariate time series. These embeddings have been demonstrated to be useful for time series reconstruction, classification, and creation of health index (HI) curves of machines being used in industrial applications, based on which the remaining useful life (RUL) of machines can be estimated. In this study, we extend the traditional form of RNN autoencoders as a feature extractor for multivariate time series to a more general form in terms of arranging the order of input or output sequences and the hidden unit architecture. We apply the embeddings obtained by different variants of RNN autoencoders for a time series classification task and a machine RUL estimation problem using two publicly available datasets. A random research strategy is used to find the optimal hyperparameters of all variants for each task in order to give a fair comparison of the general performance among different variants over a large hyperparameter space, as well as the best performance that each variant can achieve compared with the best reported values in the literature. Our results show that traditional reversing the order of output time series while maintaining the order of input time series when training an RNN autoencoder does not show improved performance for the two studied cases. Thus, intentionally arranging the input or output order seems unnecessary for training the RNN autoencoder as a feature extractor of time series. We further observe that only the RNN architectures with gating mechanism can achieve the functionality of encoding for the time series, and none of the three common gated architectures we studied shows significantly and consistently improved performance compared to the others on the two studied cases. However, the bidirectional RNN autoencoders yield slightly better performance than their unidirectional counterparts.