6976 IEEE SENSORS JOURNAL, VOL. 23, NO.

7, 1 APRIL 2023

Bayesian Sensor Calibration of a

CMOS-Integrated Hall Sensor Against
Thermomechanical Cross-Sensitivities
Moritz Berger , Member, IEEE, Christian Schott, Member, IEEE, and Oliver Paul , Senior Member, IEEE

Abstract—For the first time, Bayesian sensor calibration is

used to identify efficient calibration procedures for a sensor
cross-sensitive to two parasitic influences. The object under
study is a thermomechanically cross-sensitive sensor sys-
tem for determining the magnetic induction B. The packaged
system comprises a Hall sensor, a stress sensor, and a
temperature sensor. The three sensor signals are combined
in a polynomial sensor response model with 11 parameters to
determine B compensated for offset and cross-sensitivities.
For the calibration, sensors are exposed to mechanical
stress values between 0 and −68 MPa, temperatures between
−40 and 100 ◦ C, and B values between −25 and 25 mT.
A sample of 35 sensors serves to extract the prior model
parameter distribution of their fabrication run. The Bayesian
experimental design is applied to identify sets of 2–8 optimal
calibration conditions under I-optimality and G-optimality.
The Bayesian inference then allows to obtain the posterior
model parameter distribution of any uncalibrated sensor from the same run. Any such sensor is thereby turned into
a B measuring device with individually quantified accuracy. The method was successfully applied to 15 validation
sensors. In the case of I-optimality, the median root-mean-square (rms) σ values of the ±1σ confidence intervals for
the extracted B values were found to be 113–71 µT after near-I-optimal calibrations based on 2–8 measurements.
Over the entire range of temperature and mechanical stress and for applied |B| ≤ 25 mT, corresponding experimentally
determined medians of the rms deviations between predicted and applied B values were found to be 89–71 µT. Analogous
observations apply to G-optimality. In short, Bayesian calibration made it possible to obtain functional B sensors of
known accuracy with significantly fewer calibration measurements than model parameters. This was enabled by prior
knowledge collected by the thorough characterization of 35 prior-generating specimens.
Index Terms— Bayesian inference, calibration, compensation, experimental design, Hall sensor, multiple cross-
sensitivities, multisensor system.

I. I NTRODUCTION the stability of Hall sensors [1], [2], [3], [4], pressure sen-
NWANTED, yet often unavoidable parasitic influences sors [5], [6], [7], and mechanical stress sensors [8], [9], [10].
U affect the operation of many sensors. Uncompensated,
such cross-sensitivities cause systematic errors in the values
In fact, many sensors exhibit more than one cross-sensitivity.
For example, semiconductor Hall sensors are affected not only
inferred from a sensor system’s output signals and thus dimin- by temperature but also by mechanical stress [11]. Volatile
ish its accuracy. Temperature is well known to modulate the organic compound (VOC) sensors used as air quality sensors
response of virtually every sensor. It has been shown to impair lack selectivity to individual air compounds [12]. Electronic
tongues are cross-sensitive to various components in solu-
Manuscript received 23 January 2023; accepted 7 February 2023. tions [13]. Electronic noses have exhibited similar selectivity
Date of publication 14 February 2023; date of current version 31 March
2023. This work was supported by Melexis Technologies SA. The issues, for instance in classifying water, methanol, and ethanol
associate editor coordinating the review of this article and approving it vapors [14]. Likewise, cross-sensitivities to other gases, such
for publication was Prof. Shakeb A Khan. (Corresponding author: Moritz as C2 H5 OH, SO2 , and NO, have been suspected for an
Berger.)
Moritz Berger and Oliver Paul are with the Department of Microsys- ozone–humidity–temperature sensor system with a sensitive
tems Engineering (IMTEK), University of Freiburg, 79110 Freiburg, WO3 film [15], as found in [16] for NO2 sensors relying on
Germany (e-mail: [email protected]; [email protected]). this sensitive material. Among devices for physical measur-
Christian Schott is with Melexis Technologies SA, 2022 Bevaix,
Switzerland (e-mail: [email protected]). ands, mechanical sensors often possess cross-sensitivities to
Digital Object Identifier 10.1109/JSEN.2023.3243783 other mechanical constraints. A six-degree-of-freedom force-

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
BERGER et al.: BAYESIAN SENSOR CALIBRATION OF A CMOS-INTEGRATED HALL SENSOR 6977

moment transducer for example has exhibited for each load Several approaches have aimed to cure these disturbances
component cross-sensitivities to the other five components at the levels of device operation and system architecture.
in addition to temperature [17], [18], [19]. Similarly, inertial Integration of temperature sensors has allowed to effectively
sensors, such as gyroscopes, require compensations against compensate thermal output signal drifts [2], [54]. The current
stress, temperature, and the quadrature error [20], [21], [22]. spinning technique allows to obtain an averaged output signal
All these examples highlight the importance of calibration largely cleared of contributions caused by the shear piezore-
for guaranteeing the cross-sensitivity-free and, thus, accurate sistance effect [55], [56], [57], [58], [59], [60], [61]. The
operation of sensor systems. cointegration of temperature and stress sensors together with a
When calibrating sensors, well-chosen calibration condi- Hall sensor has enabled the analog compensation of the cross-
tions are applied to each individual device. The outcome of sensitivities [3], [62], [63]. Alternatively, the digital signal
the measurements performed under these conditions allows processing of the sensor signals has also allowed to obtain Hall
to model the individual relationship between measurand and sensor signals largely cleared of the parasitic contributions [3],
output signals. Understandably, it is of interest to keep the [45], [54], [64], [65]. As stated in [45], accuracies better than
number of calibration conditions low [23], [24]. This is 1% for the temperature range required by automotive appli-
particularly true in the context of large production volumes, cations can be achieved only by compensating the long-term
where calibration is known to be time-consuming and greedy drift of the Hall sensitivity associated with mechanical stress.
of resources [25], [26], [27], [28], [29], causing up to 50% It is noteworthy, however, that questions concerning the
of overall sensor costs [29]. However, a minimized calibration efficiency of the required calibration procedures and their
procedure jeopardizes one’s ability to guarantee a specified relation to the predictive accuracy of the compensated sen-
high sensor accuracy. sors were systematically addressed in none of these previous
The present study takes up the challenge of calibrating a studies.
sensor system possessing two cross-sensitivities by applying On a different track, data science concepts, such as machine
the method of Bayesian sensor calibration [4]. This method learning (ML), have recently gained popularity in the field of
was formulated in [4] in general terms and demonstrated on sensor calibration. Therefore, they give reason to hope that
a simple model case with a single, thermal cross-sensitivity. the open questions may be addressed by numerically based
Here, the demonstration of its usefulness is therefore expanded methods. Among the most widespread ML methods in sensor
to a more demanding case. The object under study is a Hall– calibration are artificial neural networks (ANNs) in the form of
stress–temperature sensor system sensitive to thermal drift multilayer perceptrons (MLPs) [66], [67], [68], convolutional
and mechanical loads, fabricated in complementary metal– neural networks (CNNs) [69], [70], and fuzzy neural networks
oxide–semiconductor (CMOS) technology. Sensor elements (FNNs) [71]. Other approaches have relied on random forests
for the magnetic field, temperature, and mechanical stress are (RFs) [67], [72], [73], Gaussian process regression (GPR) [73],
cointegrated with elaborate analog-to-digital circuitry [30]. [74], [75], and Bayesian neural networks [76]. These methods
Without any doubt, CMOS Hall sensors have already have been applied for temperature compensation [66], [67],
reached a high level of development. The 1-D Hall sen- temporal drift compensation of field-effect transistor sen-
sors allow to measure the out-of-plane [11] or an in-plane sors [70], and compensating commercial water quality sensors
component of the magnetic induction B [11], [31], [32], in order to extend the calibration lifetime [69]. ANNs are much
[33], [34], [35]. The 2-D Hall sensors serve to determine appreciated for their effectiveness in classification tasks [77],
angular information by measuring the in-plane components [78], [79]. An advantage of ML approaches is their ability to
of B [34], [36], while 3-D Hall sensors give access to all handle unknown, potentially complex input–output relations.
three components. Such 3-D Hall sensors have, for example, This often comes at the cost of an intense training effort and
combined vertical and horizontal Hall plates [37], [38] and the need to determine the predictive accuracy by additional
were realized as isotropic monolithic devices [39], [40]. validation data [66], [67], [71] or by cumbersome numerical
Nevertheless, as understandable from semiconductor trans- sampling [76].
port theory, such B sensors are affected by temperature Since in the present case, the relationship between the three
variations and mechanical stress. Temperature T acts via input quantities (B, s, T ) of the sensor system under study
the temperature-dependent Hall mobility of the charge car- and its three output signals is well described by low-order
riers [11], [41]. The cross-sensitivity to mechanical stress has polynomials [80], we opt for the Bayesian approach that has
its origin in piezoresistance and the piezo-Hall effect [11], already proven effective in cases with a single, thermal cross-
[42], [43], [44]. In the former, pseudo-Hall signals are sensitivity [4], [6]. In [4], a Hall-temperature sensor system
caused by shear components of the mechanical stress tensor. was calibrated for the range between −30 and 150 ◦ C. This
In the latter, the magnetic sensitivity of planar Hall plates was accomplished by applying between one and three thermal
is affected, e.g., by the sum s = σx x + σ yy of the in-plane calibration conditions, implying fewer measurements than the
normal components σx x and σ yy of the mechanical stress seven parameters contained in the sensor model. The root-
tensor [45], [46]. Mechanical stresses responsible for both mean-square (rms) accuracies after these modest calibration
effects are caused by thermomechanical properties of the het- procedures were 78, 41, and 34 µT.
erogeneous sensor assemblies [47], [48], [49], [50], [51] and The prerequisite for the effectivity of the method was
by the swelling of the packaging materials when exposed to the availability of prior information gained from a set
moisture [52], [53], [54]. of thoroughly characterized sensors termed prior-generating
6978 IEEE SENSORS JOURNAL, VOL. 23, NO. 7, 1 APRIL 2023

specimens. The prior information is obtained by thoroughly specimens, which constitute a sample of the ensemble.
characterizing those specimens. From their estimated sensor In the present case, Q = 35. Therefore, the so-called
model parameters, one infers the sensor model parameter prior probability distribution p0 (w) is constructed.
distribution of the ensemble in the form of a prior mean 2) The second step consists of two substeps, namely,
and a prior covariance matrix. This prior information is a Bayesian update and a Bayesian design of experiment.
combined with the limited evidence provided by the calibration a) The Bayesian update considers exposing any pre-
of other sensors about their individual responses. Updated viously uncalibrated specimen to a small set of
model parameters and an updated covariance matrix of each N calibration conditions, whereby its output sig-
individual sensor are thereby obtained. nals V and the corresponding applied magnetic
Section II recalls the central mathematical elements of induction values B = (B1 , . . . , B N )⊤ are col-
Bayesian sensor calibration. In Section III, we describe the lected. From p0 (w), V , and B, one obtains the
mentioned Hall–stress–temperature multisensor system and updated probability distribution p1 (w). This allows
the experimental infrastructure and procedures to calibrate to turn the previously uncalibrated specimen into
it. The results in Section IV demonstrate that these sensors a measuring device allowing to translate any com-
can be effectively compensated against mechanical stress and bination of its output signals v into a prediction
temperature variations with five or fewer thermomechanical of the applied B. This inference is achieved by
calibration conditions while offering a high sensor accuracy. the posterior predictive response B1 (v, V , B) and
Then, Section V discusses the results and is followed by the its accuracy expressed by the posterior predictive
conclusions.
variance σ21 (v, V ).
b) The Bayesian design of experiment then aims for
II. B AYESIAN CALIBRATION optimizing the postcalibration measurement accu-
Since this work applies the method described in detail in [4], racy of previously uncalibrated specimens. It min-
here, we only summarize the most important definitions and imizes σ21 (v, V ) subject to a criterion of one’s
results. choice. For this purpose, a numerical search within
We use the term sensor system as a synonym for the the range  of output signals arising from the range
CMOS-integrated, packaged Hall–stress–temperature sensor of expected operating conditions of the specimen
system providing the application and test case of this article. is carried out. As a result, one identifies the com-
An individual sensor system is considered as a specimen bination of sensor output signals V min achieving
of a sensor ensemble with statistically distributed properties. minimality.
A specimen or a group of specimens, therefore, constitute
3) In the third step, previously uncalibrated specimens
samples of the ensemble. We reserve bold symbols for vectors
of the ensemble are calibrated. Any such device is
and matrices and roman symbols for scalar values.
thereby exposed to calibration conditions that are known
In the present case, the magnetic induction B is the measur-
to produce sensor outputs V cal near V min . Based on
and of interest. It plays the role of the dependent variable. The
its individual calibration results, V cal , the specimen is
output signals of the Hall sensor and the stress and temperature
turned into a measuring device whose output signals v
sensing elements of the sensor system are summarized as
allow to infer B. This inference uses the posterior
v = (VH , VS , VT ). The components of v provide the inde-
predictive response B1 (v, V cal , B), while the accuracy
pendent variables of the Bayesian analysis. The v-dependent
of the inference is quantified by σ21 (v, V cal ).
measurand B is modeled by φ(v)⊤ w using a set of basis
functions φ(v)⊤ = (φ1 (v), . . . , φ M (v)) and the corresponding These steps are now presented in the mathematical detail
model parameters w = (w1 , . . . , w M )⊤ , where (·)⊤ denotes required by the experimental study in Sections III and IV.
the matrix transposition and M is the model dimension. In the First, for prior generation, each prior-generating specimen
present case, M = 11. The goal of optimal calibration is to is exposed to 1110 characterization conditions. For each
find w such that B can be inferred from v with the highest specimen numbered i = 1, . . . , Q, one thereby records
possible accuracy. A crucial role in this process is played by sensor output signals V i = (v i1 , . . . , v i1110 ). The B values
the design matrix 8(V ) defined as [77, Ch. 3] applied during characterization are listed in the column vector
B = (B1 , . . . , B1110 )⊤ .
φ1 (v 1 ) · · · φ M (v 1 )
 
Using standard linear regression, the model parameter vec-
8(V ) =  .. .. .. tor wi of prior-generating specimen no. i is then obtained by
. . . (1)
 

φ1 (v N ) · · · φ M (v N ) n o−1
for any list V = (v 1 , . . . , v N ) of N sensor output signal wi = 8(V i )⊤ 8(V i ) 8(V i )⊤ B (2)
vectors.
The Bayesian sensor calibration method consists of three where the term in front of B denotes the well-known
steps. Moore–Penrose pseudoinverse [81] of the design matrix evalu-
1) Prior information about the sensor parameter vectors ated at V i . From the distribution of wi in w space, one derives
w of a considered ensemble is gathered by thoroughly the prior probability distribution p0 (w) of the ensemble and
characterizing a group of Q so-called prior-generating approximates it by a multivariate normal distribution with
BERGER et al.: BAYESIAN SENSOR CALIBRATION OF A CMOS-INTEGRATED HALL SENSOR 6979

mean The G-optimality criterion [82] considers the objective

1 function
w0 = (w1 + · · · + w Q ) (3)
Q f G (V ) = max σ21 (v, V ) (11)
v∈
and covariance matrix
with optimum defined as
Q+1
60 = 1W ⊤ 1W (4) V min = min f G (V ). (12)
Q(Q − M − 2)
V ∈ N
where 1W = (w1 − w 0 , . . . , w Q − w 0 )⊤ lists the mean- In contrast, I-optimality [82], [83] relies on the objective
centered wi vectors. With w 0 and 6 0 , one is able to infer B function
from an output signal v of any uncalibrated specimen of the Z
1
ensemble. This B value is given by f I (V ) = σ2 (v, V )dv (13)
V  1
B0 (v) = φ(v)⊤ w0 (5) and identifies the optimum at
while the accuracy of the prediction is quantified by the V min = min f I (V ). (14)
variance V ∈ N

Finally, the Bayesian sensor calibration is ready to be

σ20 (v) = σ2 + φ(v)⊤ 6 0 φ(v) (6)
applied to uncalibrated specimens. In this work, such uncal-
where σ2 denotes the variance of the individual B measure- ibrated specimens serve to experimentally demonstrate the
ment of thoroughly characterized specimens. The quantities validity of the outlined method. They are therefore termed val-
B0 (v) and σ20 (v) are the mean and variance of the distribution idation specimens. Since the optimal V min lists sensor output
of B values inferrable from output signals v of any uncali- signals rather than calibration conditions (B, s, and T ), it can
brated specimen, given only the prior information. The prior be challenging to perform the calibration under the optimal
predictive probability distribution of B is in fact the normal conditions. Nevertheless, by analyzing the responses of the
distribution defined by these parameters. This is a consequence prior-generating specimens, one is able to identify calibration
of the assumed multivariate normality of p0 (w) and of the conditions that within the group of prior-generating specimens
Gaussian statistics of the individual measurement. have yield output signals close to the optimum. When these
When evidence about a specimen’s individual response near-optimal calibration conditions are applied to a validation
becomes available in the form of calibration data V = specimen, they produce the specimen’s individual near-optimal
(v 1 , . . . , v N ) and B = (B1 , . . . , B N )⊤ , Bayes’ theorem allows output signals V cal . The specimen’s individual w 1 (V cal , B)
to determine the posterior probability distribution p1 (w) valid and 6 1 (V cal ) is then updated from w 0 and 6 0 according to (7)
for the specimen. This is again a multivariate normal distribu- and (8), respectively. Using (9) and (10), these allow to infer
tion, with posterior mean [77, Sec. 3.3.1] B from the validation specimen’s measured output signals v
  using the posterior predictive response B1 (v, V cal , B) with
1 accuracy quantified by σ1 (v, V cal ).
w1 (V , B) = 6 1 (V ) 6 0 w0 + 2 8(V ) B
−1 ⊤
(7)
σ
and posterior covariance matrix III. E XPERIMENT
 −1 The Bayesian sensor calibration methodology is now
1 applied to the sensor system mentioned above. Details of
6 1 (V ) = 6 0 + 2 8(V ) 8(V )
−1 ⊤
. (8)
σ the system are presented in Section III-A. The experi-
Equations (7) and (8) describe the Bayesian updates of w 0 and mental setup and data acquisition are then described in
6 0 mediated by the new evidence V and B complementing Sections III-B and III-C, respectively.
the evidence available from the prior characterization.
Similar to the prior case, w 1 (V , B) and 6 1 (V ) enable B A. Sensor System
to be inferred from output signals v of the specimen that has The Hall sensor microsystem under study was fabricated
yielded (V , B) during calibration. This posterior predictive in an industrial 180-nm CMOS technology. An optical micro-
value is graph of an unpackaged sensor chip in Fig. 1 highlights the
Hall sensor, the mechanical stress sensor, and the temperature
B1 (v, V , B) = φ(v)⊤ w1 (V , B) (9)
sensor. The Hall sensor consists of two interconnected Hall
and the corresponding posterior predictive variance is plates sensitive to the out-of-plane magnetic induction com-
ponent B and providing the Hall voltage VH . The respective
σ21 (v, V ) = σ 2 + φ(v)⊤ 6 1 (V )φ(v). (10) output signals VS and VT of the stress and temperature sensors
reflect s and T near the Hall sensor. Further details about the
The 68.3% confidence interval of the posterior prediction is system and its architecture are reported in [30].
the ±σ1 (v, V ) range surrounding B1 (v, V , B). The Hall voltage VH is described by
For the design of experiment, as in [4], we consider two
optimization criteria, namely, G-optimality and I-optimality. VH (B, s, T ) = SA (s, T )B + Voff (s, T ) (15)
6980 IEEE SENSORS JOURNAL, VOL. 23, NO. 7, 1 APRIL 2023

Fig. 1. Optical micrograph of an unpackaged CMOS sensor chip

comprising a pair of n-doped silicon Hall plates, a temperature sensor
based on n- and p-doped resistors [2], and a piezoresistive stress
sensor (orange dashed line), realized by 16 n- and p-doped resistors
placed around the two Hall plates [8]. The stress sensor is electrically
connected in a Wheatstone bridge configuration. Output signals are con-
ditioned and digitized by cointegrated analog and digital circuitry [30].

where SA (s, T ) and Voff (s, T ) denote the stress- and

temperature-dependent absolute Hall sensitivity and the resid-
ual offset voltage at B = 0, respectively.
The purpose of the sensor system is to allow to infer B from
the sensor signals v = (VH , VS , VT ). By rearranging (15), this
Fig. 2. (a) 1/SA and (b) Boff of a representative specimen as a function
inference is achieved by
of rescaled VS and VT . The blue surfaces with white level lines are
VH the result of a polynomial fit of the data. Respective residuals, i.e., the
B(v) = − Boff (VS , VT ) (16) deviations of (c) and (e) 1/SA and (d) and (f) Boff data from their fit
SA (VS , VT ) surfaces, projected along (c) and (d) VS - and (e) and (f) VT -axes.
where 1/SA (VS , VT ) and Boff (VS , VT ) = Voff /SA denote the
inverse Hall sensitivity and the equivalent offset field, respec- The first group is numbered i = 1, . . . , 35, while the second
tively. Note that the arguments of SA and Boff are chosen is assigned the numbers i = 36, . . . , 50.
to be the sensor signals VS and VT reflecting the parasitic
influences s and T . B. Experimental Setup
The stress sensor is designed to be sensitive [8] to mechan-
Fig. 3 shows a schematic of the experimental setup with
ical stress exerted on the Hall sensor, by external constraints,
a close-up photograph of a sensor system soldered to a
e.g., compressive forces F acting on the sensor package [80]
printed circuit board (PCB). An air streamer (Dragon Air
in addition to thermomechanical loads [48]. Values of s caused
Streamer, Froilabo, France) connected to the thermal chamber
by perpendicular forces up to 20 N are expected to be
enables to vary T of the specimen hosted by the chamber.
about −68 MPa at 25 ◦ C [3], [80].
An x yz-table allows to align the thermal chamber within
Fig. 2 shows the measured values of 1/SA and Boff of a
a Helmholtz coil, which serves to apply B between
representative specimen as a function of VS and VT . The data
−25 and 25 mT. The Helmholtz coil was calibrated using
were obtained by varying the temperature from −40 to 100 ◦ C
a Tesla meter (Gauss/Tesla Meter Series 8000, F.W. BELL,
and applying perpendicular compressive forces F to the sensor
Milwaukie, OR, USA). The custom-built thermal chamber
package between 0 and 20 N, as detailed in Section III-B. The
rests on a motorized test stand (Test Stand ESM303, Mark-10,
sensor signals VH , VS , and VT were shifted to reduce their
Copiague, NY, USA). A movable, customized spring load
values at the reference conditions B = 0 mT, T = 30 ◦ C, and
system equipped with a reference force sensor (Mark-10,
F = 0 N to zero. In addition, they were rescaled.
Series 5 Force Gauge, Copiague, NY, USA) allows to expose
Inspection of the data in Fig. 2 shows that in comparison
the specimen packages to perpendicular forces F by pushing
with F = 0 N, the applied forces of 20 N reduce SA by
a rod against their surface. The rod penetrates the thermal
between 3.6% and 4.5%, depending on temperature. Similarly,
chamber through an opening in its ceiling. Under the applied
at F = 0 N, SA is reduced from the room temperature value by
forces, the output signals of the stress sensors cover a compa-
about 27% at 100 ◦ C and increased by about 42.6% at −40 ◦ C.
rable range of values as when specimens are exposed to other
It turns out that 1/SA is well fit by a polynomial model of
test procedures such as high-temperature operating lifetime
degree 4 in VS and VT , while Boff is well modeled by a
(HTOL) testing. The setup is controlled by a LabView routine.
second-degree polynomial. An appropriate polynomial model
is selected in Section IV-A.
The present study covers 50 specimens assembled as pairs C. Characterization
in 25 dual-die TSSOP-16 packages [4], [80]. The speci- The 35 prior-generating specimens were exposed to the
mens are split into two groups, the first of which comprises following characterization conditions:
Q = 35 randomly selected specimens for the generation of 1) T was varied from −40 to 100 ◦ C in steps of nominally
the prior and the remaining 15 specimens serve for validation. 10 ◦ C;
BERGER et al.: BAYESIAN SENSOR CALIBRATION OF A CMOS-INTEGRATED HALL SENSOR 6981

Fig. 3. Schematic of the automated characterization setup for the mag-

netothermomechanical calibration of Hall sensor systems. The magnetic
induction B, mechanical stress s, and temperature T are applied using a
Helmholtz coil, a motorized compression test stand, and an air streamer,
respectively. The photograph shows a packaged sensor soldered to a
PCB with the tip of the movable rod pressed against the sensor surface
as well as a temperature reference sensor.

2) At each T < 100 ◦ C, F was varied from 0 to 20 N in

steps of nominally 5 N, while for T = 100 ◦ C, F was
varied in the same steps up to 15 N;
3) At each T and F combination, B was set to −25, 0,
and 25 mT;
4) At each condition, five successive readings of the three
sensor signals were recorded.
Fig. 4 shows the measurement history of a representative
specimen. The first three graphs show T , as measured by
the temperature reference sensor, F as measured by the force
reference sensor, and the applied B values. The last three
graphs show the resulting output signals VH , VS , and VT .
Fig. 4(g) shows the range  in v space enveloping 97% of
all 35 × 1110 data points of the prior-generating specimens.
For the Bayesian data analysis in Section IV, the output
data of each prior-generating specimen are formatted as the
list of independent variables V i = (v i1 , . . . , v i1110 ) with
i = 1, . . . , 35. The dependent variable vector of the Bayesian
analysis common to all prior-generating specimens forms the
column vector B = (B1 , . . . , B1110 )⊤ .

IV. R ESULTS Fig. 4. Exemplary characterization history of a specimen with

1110 measurements consisting of three B values applied each at
Section IV-A is dedicated to the selection of a polyno- 15 nominal T values and five F values for T < 100 ◦ C and four F
mial model able to adequately describe the observed sensor values for T = 100 ◦ C. At each condition, VH , VS , and VT were recorded
five times. (a) Nominal T values, (b) force values F, and (c) B values.
responses. Thereafter, we closely follow the procedure laid Resulting sensor output signals (d) VH , (e) VS , and (f) VT . (g) Range Ω
out in Section II. in signal space covered by the applied characterization conditions. The
top and bottom surfaces correspond to B = ± 25 mT.
A. Model Selection
It is guided by three requirements. First, the model needs to by M should be large enough to allow the experimental data
be linear in its parameters w. Second, its complexity described to be accurately fit. Third, M should at the same time be as
6982 IEEE SENSORS JOURNAL, VOL. 23, NO. 7, 1 APRIL 2023

calibration conditions required to determine w of a specimen

without prior knowledge. Note that in Fig. 2, the fit surface for
1/SA was obtained with the last seven basis functions in (18),
whereas that of Boff relied on the first four.

B. Prior Generation
The model parameter vectors wi obtained with model no. 9
for the 35 prior-generating specimens constitute the database
for determining p0 (w) of the sensor ensemble, of which they
constitute a sample. By applying (3) and (4), we compute the
mean w0 and the covariance matrix 6 0 . Fig. 6(a1) shows 6 0
by a heat plot.
Fig. 5. Fit quality of the 15 investigated polynomial models listed in
From w0 and 6 0 , using (5) and (6), we next infer the prior
Table I in the Appendix. Each box plot captures the distribution of the predictive mean B0 (v) as a function of v = (VH , VS , VT ) and
35 rms residuals ∆Bi [cf. (17)] of the fits of the prior-generating sensor similarly the prior predictive standard deviation σ0 (v). The
data with the corresponding model by the method of least squares.
value of σ required in (6) is taken to be 57.7 µT. This value
is determinedPfrom the 35 × 1110 prior-generating specimen
data as σ = ( i 1Bi 2 /35)1/2 . The prior predictive confidence
small as possible to avoid overfitting [84], [85], [86], [87].
range of the inferred B0 (v) value, quantified by σ0 (v), is plot-
With the expectation that response surfaces such as those in
ted in Fig. 6(b1). In fact, the plot shows the values of σ0 (v)
Fig. 2 lend themselves to Taylor series expansion, we focus
on the top surface of  projected onto the (VS , VT )-plane.
on polynomial models in the variables VH , VS , and VT of
This choice is justified by the observation that, for given VS
increasing complexity. A selection of 15 such models is
and VT , σ0 assumes its maximum in the VH -direction either at
proposed in the Appendix. For each model, we proceed as
B = 25 mT or B = −25 mT, i.e., on the top or bottom surfaces
follows. For the prior-generating specimens, we carry out the
of  in Fig. 4(g), respectively. Furthermore, the two values of
linear regression of their data V i with i = 1, . . . , 35, and
σ0 on these two surfaces differ little due to the modest Hall
B according to (2) and hence obtain the list of parameter
sensor offset. The plot also shows the (VS , VT ) data provided
vectors wi . The rms deviation 1Bi between the data B and the
by the characterization of a representative prior-generating
fit function of specimen no. i evaluated at V i , i.e., 8(V i )wi ,
specimen. The white dashed rectangle defines the extent of 
is then given by
in the (VS , VT )-plane. It embraces 97% of the characterization
1 data acquired with the prior-generating specimens for 0 N ≤
1Bi = √ |B − 8(V i )wi | (17)
1110 F ≤ 20 N and −40 ◦ C ≤ T ≤ 100 ◦ C. It covers the ranges
−1 ≤ VS ≤ 3 and −1.6 ≤ VT ≤ 1.8.
where |·| denotes the Euclidean norm. The results are compiled
in Fig. 5. For each model, the 35 resulting rms deviations are
summarized as a box plot. The box represents the interquartile C. Calibration at near-optimal stress–temperature
range (IQR), with the median indicated by the solid line in conditions
the box; the whiskers embrace all data points lying within The next goal is to identify combinations of N calibration
1.5 IQRs below the first quartile and above the third quartile. conditions such that the corresponding sensor outputs of any
Values beyond the whiskers are considered as outliers and uncalibrated specimen minimize its posterior predictive uncer-
plotted as diamond symbols. Overall, models nos. 9 and 15 are tainty (σ21 ) with respect to the chosen optimality condition.
found to achieve the best fits, as highlighted also by the inset in The search for optimal conditions is carried out within the
Fig. 5. Since the performance of the simpler model no. 9 with domain . Thereby, one ensures that the search reasonably
M = 11 equals that of model no. 15 with M = 12, model covers the range of operating conditions to which the specimen
no. 9 is adopted for further data analysis. A formal model will later be exposed and which were consequently covered
comparison and selection, as discussed, e.g., in [77, Sec. 3.4], during the characterization of the prior-generating specimens.
[78, Ch. 5], and [88, Ch. 7], is beyond the scope of this work We perform the calibration measurements exclusively with
and delayed to a future study. In conclusion, the result of the B = ±25 mT. This is justified by the fact that for given
model selection is the vector of basis functions (VS , VT ), the relationship between B and VH is highly linear
and thus well determined by a pair of measurements. More-
φ(v) over, these measurements should ideally lie as far apart as
= 1, VT , VS , VS VT , possible in the VH -direction. Within the operating range, this
VH , VH VT , VH VT2 , VH VT3 , VH VT4 , VH VS , VH VS VT

⊤
. (18) is the case when B = ±25 mT. Therefore, the corresponding
v search domain by definition consists of the top and bottom
Note that the first four terms are independent of VH and surfaces of . Since the VH values assumed at B = ±25 mT
are thus well-suited for modeling Boff , while the other seven depend on s and T and thus on VS and VT , these two
terms are proportional to VH and thus aptly model VH /SA . surfaces can be parameterized as v ± (VS , VT ) = (VH (B =
The model dimension M = 11 sets the minimum number of ±25 mT, VS , VT ), VS , VT ), where + and − denote the top and
BERGER et al.: BAYESIAN SENSOR CALIBRATION OF A CMOS-INTEGRATED HALL SENSOR 6983

Fig. 6. (a) Covariance matrices, G-optimality. (b) Confidence intervals, G-optimality. (c) Confidence intervals, I-optimality. (a1) Elements 6 0ij of the
prior covariance matrix 6 0 and (b1) and (c1) prior confidence interval quantified by σ0 . White semitransparent dots show the characterization data
of a representative prior-generating specimen; white dashed line delimits Ω in the (VS , VT )-plane. The improved sensor accuracy after calibration
with (a2) N = 2, (a3) N = 4, (a4) N = 6, and (a5) N = 8 calibration conditions is visualized by the progressive shrinkage of the posterior covariance
matrix 6 1 . Posterior confidence intervals inferred for a representative validation specimen, as quantified by σ1 , after near-optimal calibrations with
N = 2, 4, 6, and 8 subject to (b2)–(b5) G-optimality and (c2)–(c5) I-optimality. Gray dots denote the ideal calibration conditions (VSmin , VT min ),
while the black dots show the representative validation specimen’s output signals Vcal obtained under the near-optimal calibration conditions. The
white semitransparent dots show the entire set of validation data for the representative specimen. The heat plots [(b1)–(c5)] share the same color
scale.

bottom surfaces, respectively. The search is therefore carried white dashed line delimits the search domain. These values are
out in the 2-D region delimited by −1 ≤ VS ≤ 3 and −1.6 ≤ (3, −0.3) for G-optimality and (2.2, −0.14) for I-optimality.
VT ≤ 1.8. One has to be aware that for each combination The next task is to identify calibration loads F and T
of VS and VT identified as an adequate calibration condition, able to elicit response signals (VScal , VT cal ) from any spec-
a pair of calibration measurements is carried out, namely, imen near the ideal values (VSmin , VTmin ). By analysis of the
at B = ±25 mT. Switching B requires only the Helmholtz prior-generation data and the loads applied there, F = 20 N
coil current to be inverted, which is fast and thus efficient. and T = 20 ◦ C are concluded to be a reasonable choice for
In what follows, the procedure is illustrated in detail for G-optimality, whereas F = 15 N and T = 20 ◦ C are for
N = 2. In other words, a single optimal calibration combi- I-optimality. Under these near-optimal loads, calibration output
nation of (VSmin , VTmin ) is to be identified. For this purpose, signals V cal are recorded from a specimen being calibrated.
we define V = (v − (VS , VT ), v + (VS , VT )). Therefore, σ1 (v, V ) Based on V cal and B cal = (−25 mT, 25 mT)⊤ , one deduces
is obtained using (8) and (10). This then serves to evaluate B1 (v, V cal , B cal ) = φ(v)⊤ w 1 (V cal , B cal ) and σ21 (v, V cal ) of
f G (V ) and f I (V ) [cf. (11) and (13)]. The two objective the sensor, using (7)–(10).
functions are shown in Fig. 7 as a function of the two variables Fig. 6(a2) symbolizes the posterior covariance matrix 6 1
VS and VT of V . Their minima were identified numerically in of a representative validation specimen after near-G-optimal
Python using the SciPy library [89]. The optimal calibration calibration at V cal . The corresponding confidence interval,
conditions (VSmin , VTmin ) are shown as gray dots, while the as quantified by σ1 (v, V cal ) on the top surface of , is
6984 IEEE SENSORS JOURNAL, VOL. 23, NO. 7, 1 APRIL 2023

Fig. 7. (VS , VT ) dependence of the objective functions for

(a) G-optimality [fG (mT2 )] and (b) I-optimality [fI (mT2 )] for calibration at
a single stress–temperature condition. The white dashed line delimits Ω.
The minima, indicated by gray dots, define the respective calibration
conditions (VSmin , VT min ).

shown in Fig. 6(b2), again as a function of VS and VT . The

optimal calibration condition (VSmin , VTmin ) is indicated by
the gray dot, while the near-optimal condition is indicated
by the black dot. The white dashed boarder again delimits
. The semitransparent white dots show sensor output signals Fig. 8. Top half: rms residual values (light-colored dots) and cor-
(VS , VT ) of the representative specimen under the same set of responding box plots for the 15 validation specimens after calibra-
load conditions as applied during characterization of the prior- tion based on (a) G-optimality and (b) I-optimality for calibration with
N = 2, 4, 6, and 8 near-optimal measurements and characterization
generating specimens. These actual output data, including using 1110 measurements. Distributions of the corresponding maximum
corresponding VH values for B = −25, 0, and 25 mT, are used absolute residuals are shown in the lower half. For N = 0, the plots also
for the comparison of actual data with the predictive response show the corresponding results for the 35 prior-generating specimens.
B1 (v, V cal , B cal ) and thus for the validation of the method
in Section IV-D. Similar to Fig. 6(b2), Fig. 6(c2) shows σ1 as
inferred from near-I-optimal calibration of the same validation procedure as the prior-generating specimens, as described in
specimen. Section III-C. Per validation specimen, this results in a set
Finally, optimal and near-optimal calibration strategies with of 1110 data triples v = (VH , VS , VT ) with the corresponding
N > 2 load conditions are analyzed. We considered N = 4, 6, applied B values. The set of validation data of a representative
and 8, implying combinations of 2, 3, and 4 (F, T ) conditions, validation specimen is shown in Fig. 6(b2)–(b5) and (c2)–(c5).
respectively. At each condition, measurements are carried out For each validation specimen, we compute the 1110 resid-
with applied B = ±25 mT. Results are shown in the last uals B0 − B for N = 0 and B1 − B for the cases N = 2, 4, 6,
three columns of Fig. 6. The heat plots in Fig. 6(a3)–(a5) and 8 discussed in Section IV-C and for N = 1110. The latter
again symbolize the resulting 6 1 of the updated probability case means that the Bayesian update is performed with all
distribution of w. These matrices are obtained for measure- available data.
ments performed at the optimal calibration conditions indi- Accuracies of the validation specimens before and after
cated by the two [Fig. 6(a3)], three [Fig. 6(a4)], and four calibration are shown by orange box plots in Fig. 8. After
[Fig. 6(a5)] gray dots in Fig. 6(b3)–(b5), respectively. Like for near-G-optimal calibration, the distribution of the rms resid-
N = 2, these optimal conditions were determined numerically uals of the 15 validation specimens is shown in Fig. 8(a1),
using a search algorithm programmed in Python and taking while the distribution of the maximum absolute residuals
advantage of the SciPy library [89]. Actual calibration data is reported in Fig. 8(a2). Fig. 8(b1) and (b2) summarizes
of a representative validation specimen were obtained under the corresponding observed accuracies after near-I-optimal
near-optimal conditions indicated by the black dots. These calibrations of the validation specimens. For N = 0, the
substitute conditions were again selected based on the prior distributions of the rms and maximum absolute residuals
characterization data, as described for N = 2. The contour computed for the prior-generating specimens are reported as
plots show the resulting (VS , VT )-dependent updated predictive well by blue-colored box plots.
confidence interval defined by σ1 derived from near-optimal Furthermore, the fraction of residuals lying within the
calibration conditions V cal . These are again the values of σ1 ±σ1 -interval of the inferred B values, i.e., B0 for N = 0 and
on the top boundary of . B1 for all other N , has been computed. In its top section,
The results in Fig. 6(c3)–(c5) are similar to those in Fig. 9 shows these fractions for G- and I-optimality. Orange
Fig. 6(b3)–(b5). However, like those in (c2), they were bars represent again the validation specimens, whereas blue
obtained with the aim of I-optimal calibration. bars result from the prior-generating specimens.
In addition, for each specimen and each abovementioned
1/2 1/2
N value, f G (V cal ) and f I (V cal ) are computed using (11)
D. Validation and (13). These quantify the maximum and rms predictive
For the purpose of calibration strategy validation, the 15 val- uncertainties resulting from the prior (N = 0) and updated
idation specimens were submitted to the same characterization (N = 2, 4, 6, 8, and 1110) knowledge about the specimens.
BERGER et al.: BAYESIAN SENSOR CALIBRATION OF A CMOS-INTEGRATED HALL SENSOR 6985

B value. It does so by providing the standard deviation σ1

within  and even beyond its boundaries.
Figs. 6(b2) and (c2) and 7 show the G- and I-optimal
calibration conditions found for a single (F, T ) calibration
condition as gray dots. It is noteworthy in both cases that
the near-optimal calibration within the grid of (F, T ) condi-
tions defined in Section III-C is performed at T = 20 ◦ C.
Calibration close to room temperature is favorable from an
economic point of view. It is interesting to know whether
the near-optimal conditions F = 20 N (G-optimality) and
F = 15 N (I-optimality) can be replaced without significant
loss by the more economical F = 0 N. This can be assessed
using the validation data. We conclude that the maximum
Fig. 9. N-dependence of the objective functions (a) fG (G-optimality) σ1 increases to about 428 µT from about 283 µT in the
and (b) fI (I-optimality). For N = 0, values obtained with σ0 are shown near-optimal case. Similarly, the rms σ1 increases to 160 µT
as olive-colored dots. Green dots show the values obtained under 2, 4, from 113 µT. In conclusion, it is advisable in this case to apply
6, and 8 ideal calibration conditions. The box plots show the distributions
of values obtained by corresponding near-optimal calibrations of the a nonzero force to ensure proper stress compensation.
15 validation specimens. The upper two plots show the fractions of the The posterior accuracy is now evaluated in terms of the
residuals of the entire datasets of the specimens within the respective residuals of the 15 validation specimens. In the following, the
± σ0 (N = 0) and ± σ1 (N = 2, 4, 6, and 8) confidence intervals (blue:
prior generation and orange: validation). focus is on rms residuals after I-optimal calibrations reported
in Fig. 8(b1). Two outliers among the 15 × 6 = 90 rms
values, indicated by the diamond symbols, are neglected. The
The lower part of Fig. 9 shows the results. The olive-colored figure shows the distribution of the rms residuals and describes
dots show the values derived from σ20 (v), while the green dots the tradeoff between absolute accuracy and calibration effort
show the corresponding values obtained with σ21 (v, V min ) after as a function of N . Before calibration and based on the
optimal calibration. The box plots show the distributions of prior knowledge alone, the rms residuals were found to be
1/2 1/2
f G and f I after individual near-optimal calibrations of the 74–383 µT, corresponding to relative errors of 0.3%–1.5% in
validation specimens, using σ1 (v, V cal ). comparison with the maximum B value of 25 mT. After cali-
bration at the near-I-optimal single (F, T ) condition, the rms
residuals were reduced to 65–129 µT, i.e., 0.26%–0.52%. With
V. D ISCUSSION increasing (F, T ) conditions, only small further improvements
For the first time, a Hall–stress–temperature sensor system to 63–89 µT (N = 4), 53–90 µT (N = 6), and 58–99 µT
was calibrated against two cross-sensitivities by means of (N = 8) are achieved. It is noteworthy that near-I-optimal
Bayesian calibration. The calibration was carried out with calibration with N = 2 allowed to significantly reduce the
fewer measurements than model parameters. For a sensor inaccuracy of the outlier in Fig. 8(b1) for N = 0. It is
model with M = 11 parameters, the method was demonstrated captured by the upper whisker. By a thorough calibration with
by calibrations using N = 2, 4, 6, and 8 measurements N = 1110, the residuals are further reduced to 47–63 µT
requiring only one, two, three, and four force–temperature (0.19%–0.25%). This is only a minor further improvement
(F, T ) pairs as calibration conditions. All identified G- and considering the additional effort.
I-optimal calibration strategies allow an unknown B pervading In [80], a non-Bayesian calibration of 20 Hall–stress–
calibrated sensors to be inferred from the sensor system’s temperature sensor systems provided an accuracy of
output signals VH , VS , and VT with known accuracy. The 149 (0.5%) and 236 µT (0.79%) using N = 570 and N = 6
accuracy increases with N , as shown in Figs. 6, 8, and 9. measurements, respectively, for B values in the range of
Near-G-optimal calibrations lead to a progressive refinement ±30 mT. A polynomial model comprising six parameters
of the covariance matrix with each additional (F, T ) cali- was used to infer the magnetic induction B for temperatures
bration condition [cf. Fig. 6(a1)–(a5)]. The refinement of 6 1 between −40 ◦ C and 125 ◦ C and mechanical stress values
entails a corresponding shrinking of the uncertainty σ1 within comparable to those applied here. In conclusion, the accuracy
the operating range  of accordingly calibrated specimens and achieved in the present Bayesian study after calibration with
hence allows a more accurate inference of B from the signals. N = 2 exceeds that of the non-Bayesian case in [80] with
For example, the maxima σ1 on the top surface of  after N = 6 by a factor of 1.5–3.
a single (F, T ) calibration under G-optimal [cf. Fig. 6(b2)] The continuous reduction of both objective functions, f G
and I-optimal [cf. Fig. 6(c2)] conditions are smaller than the and f I , with increasing N is apparent from Fig. 9. It confirms
minimum of σ0 before calibration [cf. Fig. 6(b1) and (c1)]. that the Bayesian optimal calibration design is effective and
With increasing N , the uncertainty shrinks further, as shown works with smaller numbers of calibration measurements
in Fig. 6(b1)–(c5). A salient feature of Bayesian sensor cali- than model parameters. In the following, we discuss the
bration in comparison with nonprobabilistic approaches is that N -dependent uncertainty left by the calibration based on
1/2
it not only allows to infer B from the sensor signals but the 15 f I values (i.e., the 15 values of the rms σ1 in ) for
simultaneously also provides the uncertainty of the inferred the near-I-optimal calibrations in Fig. 9(b) again neglecting
6986 IEEE SENSORS JOURNAL, VOL. 23, NO. 7, 1 APRIL 2023

the outliers. For comparison, the rms σ0 in  is 267 µT. response function of the form φ(x)⊤ w linear in the model
1/2
After a single (F, T ) calibration (N = 2), the 15 f I values parameters w, where x denotes the independent variables.
are reduced to 112–114 µT. Thereafter, like the residuals There is no fundamental restriction regarding the set of basis
(cf. Fig. 8), they are only slightly further reduced to 87–89 µT functions φ(x). Sensor types that may benefit from Bayesian
(N = 4), 73–83 µT (N = 6), and 68–77 µT (N = 8). When sensor calibration in the present form include ion-selective
all available data (N = 1110) are used, a further reduction chemical sensors [90], [91], inertial sensors [22], [23], [92],
to 57.9–58 µT close to σ = 57.7 µT is achieved. A similar and mechanical sensors [5], [8], [10], [17], [18], [93]. By using
observation was found in [4]. After 14 Hall–temperature the method of multivariate Bayesian regression and infer-
sensor systems were calibrated at two temperatures near the ence [77], [88], we expect the method to be generalizable
I-optimum, their uncertainty was reduced from 203 (rms σ0 ) to multisensor systems designed to provide values of more
to 41 µT (rms σ1 ). In comparison, the more complex model than a single measurand. Sensors modeled by more com-
in the present study additionally ensures the stress compen- plex response functions nonlinearly involving some model
sation and therefore demands a more substantial calibration parameters, such as chemical sensors, do not preclude the
effort in order to achieve a similar accuracy. In [6], similar application of Bayesian methods. However, the mathemat-
observations were made regarding the accuracy improvement ics will no longer boil down to matrix calculus and likely
after a Bayesian calibration of temperature-sensitive pressure entail heavier computations [76], [77], [88]. In these cases
sensors. This study used a polynomial model with five model and others without available explicit models, ANNs may be
parameters and determined calibration conditions in terms of helpful [67], [68].
I-optimality.
The application of near-optimal (F, T ) calibration condi- VI. C ONCLUSION
tions produces sensor signals V cal differing from the identified In this article, the method of Bayesian sensor calibration
optimum V min minimizing f G or f I . The optimal f G and f I was successfully applied to a multisensor system affected
values are represented by green dots in Fig. 9, while the boxes by two parasitic sensitivities. Bayesian calibration of the
include the corresponding values obtained with near-optimal investigated Hall–stress–temperature sensor system guarantees
V cal values. All experimental values in Fig. 9 closely follow a satisfying accuracy even when relying on fewer calibration
the theoretical optimal values. Consequently, near-optimal measurements (N = 2, 4, 6, and 8) than model parameters
calibration strategies were obviously identified. Nevertheless, (M = 11). For comparison, a thorough calibration with a
how to select a near-optimal (F, T ) pair of conditions for a set of N = 1110 conditions leads to a median residual
given (VSmin , VTmin ) will likely depend on the required sensor of 0.21% referred to B = 25 mT. This is only 0.07%
specification in view of its application and may also be subject better than a calibration using six measurements. A second
to the question of cost-effectiveness. strength of the Bayesian approach to sensor calibration is that
The decision to rely on Q = 35 specimens for the prior it enables to predict the accuracy resulting from calibration.
generation was taken in view of the complexity of the model The validity of the accuracy predictions was experimentally
with its M = 11 parameters. With M Q = 385 prior-generating verified. The accuracy of the validation specimens after parsi-
data, the information available to determine the symmetric 6 0 monious calibration was indeed found to be as predicted. The
was significantly larger than the number of its independent ability to predict the accuracy of specimens after calibration
entries, namely, M(M + 1)/2 = 66. Compared to the study distinguishes the Bayesian approach from ANN-based ML
in [4], where Q = 14 and M = 7, the ratio of available prior algorithms, where the trustworthiness of the trained ANN,
information to independent entries of 6 0 was chosen here to instead of being confirmed, is established by testing it using
be even larger, namely, 385/66 ≈ 5.8 in comparison with independent data [14], [71], [76], [94], [95].
98/28 = 3.5 in order to ensure a trustworthy prior. However, The successful reduction of calibration conditions still
the question of how far the multivariate normal distribution ensuring competitive accuracy is rooted in the prior distri-
described by (3) and (4) is a trustworthy approximation of the bution of sensor model parameters. A fundamental require-
subjacent multivariate student distribution [4] remains open at ment for establishing such useful prior knowledge is that the
this point and deserves a dedicated thorough study. specimens belong to an ensemble of sensor systems with a
Nevertheless, in the context of this question, it is remarkable reasonably narrow distribution of response parameters. In the
that in all calibration cases, as highlighted by the top of present case, this is ensured at the technology and hardware
Fig. 9, more than 67% of the applied B values lie within the levels by a commercial standard 180-nm CMOS process of
±σ1 -interval of the B value inferred from the sensor signals. X-Fab Silicon Foundries (Erfurt, Germany) for the fabrication
This is close to 68.3%, the well-known cumulative probabil- and by a sophisticated sensor design [30]. It is clear that
ity within the ±1σ1 range of a Gaussian around its mean. the effort needed to acquire the prior database is intense
We interpret this as evidence that the prior probability dis- and represents a weighty factor in the total calibration cost.
tribution of w estimated from the prior-generating specimens However, the parsimony of the reduced calibration schemes
using (3) and (4) provides a reasonable picture of the actual building upon the prior allows to save costs, possibly over
w distribution of the sensor ensemble. entire production volumes. Whether the Bayesian approach
The formalism described in Section II and applied here is able to offer a net cost saving in some calibration task
to the special case of Hall sensors is applicable whenever will depend on aspects extending beyond the limits of purely
the measurand of a sensor system is well modeled by a scientific questions.
BERGER et al.: BAYESIAN SENSOR CALIBRATION OF A CMOS-INTEGRATED HALL SENSOR 6987

TABLE I [2] S. Huber, A. Laville, C. Schott, and O. Paul, “A bridge-type resistive

S ETS OF B ASIS F UNCTIONS OF M ODEL N OS . 1 TO 15 W ITH temperature sensor in CMOS technology with low stress sensitivity,” in
M M ODEL PARAMETERS . M ODEL N O. 9 H AS B EEN Proc. IEEE SENSORS, Valencia, Spain, Nov. 2014, pp. 1455–1458.
S ELECTED IN THE P RESENT S TUDY [3] S. H. Lindenberger, “Active stabilization of the magnetic sensitivity in
CMOS Hall sensors,” Ph.D. thesis, Dept. Microsyst. Eng. (IMTEK),
Univ. Freiburg, Breisgau, Germany, 2017. [Online]. Available:
https://freidok.uni-freiburg.de/fedora/objects/freidok:11568/datastreams/
FILE1/content
[4] M. Berger, C. Schott, and O. Paul, “Bayesian sensor calibration,” IEEE
Sensors J., vol. 22, no. 20, pp. 19384–19399, Oct. 2022.
[5] M. Akbar and M. A. Shanblatt, “Temperature compensation of piezore-
sistive pressure sensors,” Sens. Actuators, A. Phys., vol. 33, no. 3,
pp. 155–162, 1992.
[6] S. B. Crary and Y. Jeong, “Bayesian optimal design of experiments
for sensor calibration,” in Proc. Int. Solid-State Sens. Actuators Conf.,
Stockholm, Sweden, Jun. 1995, pp. 48–51.
[7] M. Aryafar, M. Hamedi, and M. M. Ganjeh, “A novel temperature
compensated piezoresistive pressure sensor,” Measurement, vol. 63,
pp. 25–29, Mar. 2015.
[8] S. Huber, C. Schott, and O. Paul, “Package stress monitor to compensate
for the piezo-Hall effect in CMOS Hall sensors,” IEEE Sensors J.,
vol. 13, no. 8, pp. 2890–2898, Aug. 2013.
[9] J. L. Ramirez and F. Fruett, “Integrated octagonal mechanical stress
sensor with temperature compensation,” IEEE Sensors J., vol. 18, no. 14,
pp. 5707–5714, Jul. 2018.
[10] M. O. Kayed, A. A. Balbola, E. Lou, and W. A. Moussa, “Development
of doped silicon multi-element stress sensor rosette with tempera-
ture compensation,” IEEE Sensors J., vol. 20, no. 3, pp. 1176–1183,
Feb. 2020.
[11] R. S. Popovic, Hall Effect Devices, 2nd ed. London, U.K.: Institute of
Physics Publishing, 2004.
[12] S. Dey, S. Santra, S. K. Ray, and P. K. Guha, “Coral-like Cux Ni(1−x) O-
based resistive sensor for humidity and VOC detection,” IEEE Sensors
J., vol. 18, no. 15, pp. 6078–6084, Aug. 2018.
[13] Y. Vlasov, A. Legin, and A. Rudnitskaya, “Cross-sensitivity evaluation
of chemical sensors for electronic tongue: Determination of heavy metal
ions,” Sens. Actuators B, Chem., vol. 44, nos. 1–3, pp. 532–537, 1997.
[14] T. Hayasaka et al., “An electronic nose using a single graphene FET
and machine learning for water, methanol, and ethanol,” Microsyst.
Nanoeng., vol. 6, no. 1, p. 50, 2020, doi: 10.1038/s41378-020-0161-3.
[15] W. Qu and W. Wlodarski, “A thin-film sensing element for ozone,
A PPENDIX humidity and temperature,” Sens. Actuators B, Chem., vol. 64, nos. 1–3,
pp. 42–48, Jun. 2000.
Table I lists the 15 polynomial basis functions φ(v) used [16] C. Cantalini et al., “Cross sensitivity and stability of NO2 sensors
to model the magnetic induction B as a function of VH , VS , from WO3 thin film,” Sens. Actuators B, Chem., vol. 35, nos. 1–3,
pp. 112–118, 1996.
and VT . The models are systematically arranged in five triplets,
[17] F. Becker and O. Paul, “Efficient cross-sensitivity compensation in
namely (1, 2, 3), (4, 5, 6), . . . , (13, 14, 15). Each triplet multisensor systems by half-blind calibration,” Sens. Actuators A, Phys.,
has the same set of basis functions modeling 1/SA and an vol. 257, pp. 154–164, Apr. 2017.
increasing number of basis functions modeling Boff . The first [18] F. Becker, R. Jäeger, F. Schmidt, B. Lapatki, and O. Paul, “Miniatur-
ized six-degree-of-freedom force/moment transducers for instrumented
model, no. 1, uses the same basis functions as in [4] and has teeth,” IEEE Sensors J., vol. 17, no. 12, pp. 3644–3655, Jun. 2017.
no polynomial terms in VS . For models 2 and 3, the terms VS [19] F. Becker, B. Lapatki, and O. Paul, “Miniaturized six-degree-of-freedom
and VS , VS VT are added to 1 and VT , respectively. The same force/moment transducers for instrumented teeth with single sensor
principle applies to all further triplets of basis functions where chip,” IEEE Sensors J., vol. 18, no. 6, pp. 2268–2277, Mar. 2018.
[20] E. Tatar, S. E. Alper, and T. Akin, “Quadrature-error compensation and
the basis functions modeling 1/SA are progressively expanded. corresponding effects on the performance of fully decoupled MEMS
For example, the triplet (4, 5, and 6) has the additional term gyroscopes,” J. Microelectromech. Syst., vol. 21, no. 3, pp. 656–667,
VH VS in comparison with the triplet (1, 2, and 3), while the Jun. 2012.
[21] B. E. Uzunoglu, D. Erkan, and E. Tatar, “A ring gyroscope with on-
terms for modeling Boff are the same. chip capacitive stress compensation,” J. Microelectromech. Syst., vol. 31,
no. 5, pp. 741–752, Oct. 2022.
ACKNOWLEDGMENT [22] C. Jurgschat, F. Roewer, I. Toth, T. Ohms, and A. Zimmermann,
The authors would like to thank Melexis Technologies, “Integrated stress sensors for humidity performance drift analysis and
compensation in inertial measurement units,” J. Microelectromech. Syst.,
Bevaix, Switzerland, for providing the Hall multisensor sys- vol. 31, no. 6, pp. 918–926, Dec. 2022.
tems and for technical support, especially Hugues Débieux for [23] I. Skog and P. Händel, “Calibration of a MEMS inertial measurement
his professional support of the measurements. unit,” in Proc. XVII IMEKO World Congr., Rio de Janeiro, Brazil, 2006,
pp. 1–6.
[24] W. Fong, S. Ong, and A. Nee, “Methods for in-field user calibration
R EFERENCES of an inertial measurement unit without external equipment,” Meas. Sci.
[1] I. N. Cholakova, T. B. Takov, R. T. Tsankov, and N. Simonne, “Tem- Technol., vol. 19, no. 8, 2008, Art. no. 085202.
perature influence on Hall effect sensors characteristics,” in Proc. 20th [25] D. W. Braudaway, “The costs of calibration,” IEEE Trans. Instrum.
Telecommun. Forum (TELFOR), Nov. 2012, pp. 967–970. Meas., vol. 52, no. 3, pp. 738–741, Jun. 2003.
6988 IEEE SENSORS JOURNAL, VOL. 23, NO. 7, 1 APRIL 2023

[26] J. Fraden, “Calibration,” in Handbook of Modern Sensors: Physics, [51] J. C. Suhling and R. C. Jaeger, “Silicon piezoresistive stress sensors and
Designs, and Applications, 3rd ed. New York, NY, USA: Springer, 2004, their application in electronic packaging,” IEEE Sensors J., vol. 1, no. 1,
ch. 2, pp. 18–19. pp. 14–30, Jun. 2001.
[27] G. van der Horn and J. H. Huijsing, “Integrated smart sensor calibration,” [52] U. Ausserlechner, M. Motz, and M. Holliber, “Drift of magnetic sensi-
in Smart Sensor Interfaces, J. H. Huijsing and G. C. M. Meijer, Eds. tivity of smart Hall sensors due to moisture absorbed by the IC-package
Boston, MA, USA: Springer, 1997, pp. 45–60, doi: 10.1007/978-1-4615- [automotive applications],” in Proc. IEEE SENSORS, Vienna, Austria,
6061-6_5. Oct. 2004, pp. 455–458.
[28] A. Miquel-Ibarz, J. Burgués, and S. Marco, “Global calibration models [53] J. M. Cesaretti, W. P. Taylor, G. Monreal, and O. Brand, “Effect of
for temperature-modulated metal oxide gas sensors: A strategy to reduce stress due to plastic package moisture absorption in Hall sensors,” IEEE
calibration costs,” Sens. Actuators B, Chem., vol. 350, Jan. 2022, Trans. Magn., vol. 45, no. 10, pp. 4482–4485, Oct. 2009.
Art. no. 130769. [54] M. Motz, U. Ausserlechner, and M. Holliber, “Compensation of mechan-
[29] M. Shoaib, N. H. Hamid, A. F. Malik, N. B. Zain Ali, and M. Tariq ical stress-induced drift of bandgap references with on-chip stress
Jan, “A review on key issues and challenges in devices level MEMS sensor,” IEEE Sensors J., vol. 15, no. 9, pp. 5115–5121, Sep. 2015.
testing,” J. Sensors, vol. 2016, Feb. 2016, Art. no. 1639805. [55] M. Cornils and O. Paul, “Reverse-magnetic-field reciprocity in conduc-
[30] S. Leroy, S. Rigert, A. Laville, A. Ajbl, and G. F. Close, “Integrated tive samples with extended contacts,” J. Appl. Phys., vol. 104, no. 2,
Hall-based magnetic platform for position sensing,” in Proc. 43rd IEEE Jul. 2008, Art. no. 024505, doi: 10.1063/1.2951895.
Eur. Solid State Circuits Conf. (ESSCIRC), Sep. 2017, pp. 360–363. [56] O. Paul and M. Cornils, “Explicit connection between sample geometry
[31] R. S. Popovic, “The vertical Hall-effect device,” IEEE Electron Device and Hall response,” Appl. Phys. Lett., vol. 95, no. 23 Art. no. 232112.
Lett., vol. EDL-5, no. 9, pp. 357–358, Sep. 1984. [57] G. Boero, M. Demierre, P.-A. Besse, and R. Popovic, “Micro-Hall
[32] C. Schott, P.-A. Besse, and R. S. Popovic, “Planar Hall effect in devices: Performance, technologies and applications,” Sens. Actua-
the vertical Hall sensor,” Sens. Actuators A, Phys., vol. 85, nos. 1–3, tors A, Phys., vol. 106, pp. 314–320, Sep. 2003. [Online]. Available:
pp. 111–115, Aug. 2000. https://linkinghub.elsevier.com/retrieve/pii/S0924424703001924
[33] C. Roumenin, K. Dimitrov, and P. Tzvetkov, “Vertical Hall effect devices [58] P. Munter, “A low-offset spinning-current Hall plate,” Sens. Actua-
in the basis of smart silicon sensors,” in Proc. IEEE Intell. Data tors A, Phys., vol. 22, pp. 743–746, Jun. 1989. [Online]. Available:
Acquisition Adv. Comput. Syst., Technol. Appl., Sep. 2005, pp. 55–58. https://www.sciencedirect.com/science/article/pii/092442478980069X
[34] C. Sander, C. Leube, and O. Paul, “Novel compact two-dimensional [59] R. Steiner, C. Maier, M. Mayer, S. Bellekom, and H. Baltes,
CMOS vertical Hall sensor,” in Proc. Transducers-18th Int. Conf. “Influence of mechanical stress on the offset voltage of Hall
Solid-State Sensors, Actuat. Microsyst. (TRANSDUCERS), Jun. 2015, devices operated with spinning current method,” J. Microelectromech.
pp. 1164–1167. Syst., vol. 8, no. 4, pp. 466–472, Dec. 1999. [Online]. Available:
[35] C. Sander et al., “Fully symmetric vertical Hall devices in CMOS http://ieeexplore.ieee.org/document/809062/
technology,” in Proc. IEEE SENSORS, Nov. 2013, pp. 15–18. [60] R. Steiner, A. Haberli, F.-P. Steiner, and H. Baltes, “Offset reduction
[36] R. Racz, C. Schott, and S. Huber, “Electronic compass sensor,” in Proc. in Hall devices by continuous spinning current method,” in Proc. Int.
IEEE Sensors, 2004, pp. 1446–1449. Solid State Sens. Actuators Conf. (Transducers), vol. 1, Jun. 1997,
[37] F. Burger, P.-A. Besse, and R. S. Popovic, “New fully integrated 3-D pp. 381–384.
silicon Hall sensor for precise angular-position measurements,” Sens. [61] S. Bellekom, “CMOS versus bipolar Hall plates regarding off-
Actuators A, Phys., vol. 67, nos. 1–3, pp. 72–76, May 1998. set correction,” Sens. Actuators A, Phys., vol. 76, nos. 1–3,
[38] C. Schott, D. Manic, and R. S. Popovic, “Microsystem for high-accuracy pp. 178–182, Aug. 1999. [Online]. Available: https://linkinghub.elsevier.
3-D magnetic-field measurements,” Sens. Actuators A, Phys., vol. 67, com/retrieve/pii/S0924424799000072
nos. 1–3, pp. 133–137, May 1998. [62] S. Huber, W. Leten, M. Ackermann, C. Schott, and O. Paul, “A fully
[39] C. Sander, C. Leube, T. Aftab, P. Ruther, and O. Paul, “Isotropic 3D integrated analog compensation for the piezo-Hall effect in a CMOS
silicon Hall sensor,” in Proc. 28th IEEE Int. Conf. Micro Electro Mech. single-chip Hall sensor microsystem,” IEEE Sensors J., vol. 15, no. 5,
Syst. (MEMS), Jan. 2015, pp. 893–896. pp. 2924–2933, May 2015, doi: 10.1109/JSEN.2014.2385879.
[40] C. Sander, C. Leube, T. Aftab, P. Ruther, and O. Paul, “Monolithic [63] A. Ajbl, M. Pastre, and M. Kayal, “A fully integrated Hall sensor
isotropic 3D silicon Hall sensor,” Sens. Actuators A, Phys., vol. 247, microsystem for contactless current measurement,” IEEE Sensors J.,
pp. 587–597, Aug. 2016. vol. 13, no. 6, pp. 2271–2278, Jun. 2013.
[41] D. Manic, J. Petr, and R. S. Popovic, “Temperature cross-sensitivity of [64] S. Huber, Z. Lazar, M. Berger, and O. Paul, “A CMOS Hall-based
Hall plate in submicron CMOS technology,” Sens. Actuators A, Phys., magnetic multisensor system free from parasitic effects of temperature
vol. 85, no. 1, pp. 244–248, 2000. and package stress,” in Proc. 20th Int. Conf. Solid-State Sens., Actua-
[42] B. Hälg, “Piezo-Hall coefficients of n-type silicon,” J. Appl. Phys., tors Microsyst. Eurosensors XXXIII (TRANSDUCERS EUROSENSORS
vol. 64, no. 1, pp. 276–282, 1988. XXXIII), Jun. 2019, pp. 134–137.
[43] Y. Kanda and Y. Kanda, “A graphical representation of the piezoresis- [65] M. Motz, U. Ausserlechner, W. Scherr, and B. Schaffer, “An inte-
tance coefficients in silicon,” IEEE Trans. Electron Devices, vol. ED-29, grated magnetic sensor with two continuous-time 16-converters and
no. 1, pp. 64–70, Jan. 1982. stress compensation capability,” in IEEE Int. Solid-State Circuits Conf.
[44] Y. Kanda, “Piezoresistance effect of silicon,” Sens. Actuators A, Phys., (ISSCC) Dig. Tech. Papers, Feb. 2006, pp. 1151–1160.
vol. 28, no. 2, pp. 83–91, 1991. [66] N. Kazemi, M. Abdolrazzaghi, P. Musilek, and M. Daneshmand,
[45] U. Ausserlechner, M. Motz, and M. Holliber, “Compensation of the “A temperature-compensated high-resolution microwave sensor using
Piezo-Hall effect in integrated Hall sensors on (100)-Si,” IEEE Sen- artificial neural Network,” IEEE Microw. Wireless Compon. Lett.,
sors J., vol. 7, no. 11, pp. 1475–1482, Nov. 2007. vol. 30, no. 9, pp. 919–922, Sep. 2020.
[46] Y. Zou, J. C. Suhling, R. C. Jaeger, and H. Ali, “Three dimensional die [67] N. Kazemi, M. Abdolrazzaghi, and P. Musilek, “Comparative analy-
surface stress measurements in delaminated and non-delaminated plastic sis of machine learning techniques for temperature compensation in
packages,” in Proc. 48th Electron. Compon. Technol. Conf., May 1998, microwave sensors,” IEEE Trans. Microw. Theory Techn., vol. 69, no. 9,
pp. 1223–1234. pp. 4223–4236, Sep. 2021.
[47] J. H. Lau, Ed., Thermal Stress and Strain in Microelectronics Packaging. [68] Z. Nenova and G. Dimchev, “Compensation of the impact of disturbing
New York, NY, USA: Springer, 1993, doi: 10.1007/978-1-4684-7767-2. factors on gas sensor characteristics,” Acta Polytech. Hung, vol. 10, no. 3,
[48] S. Fischer, H. Beyer, R. Janke, and J. Wilde, “The influence of package- pp. 97–111, Jan. 2013.
induced stresses on moulded Hall sensors,” Microsyst. Technol., vol. 12, [69] P. Khatri, K. K. Gupta, and R. K. Gupta, “Drift compensation of
nos. 1–2, pp. 69–74, 2005. commercial water quality sensors using machine learning to extend the
[49] S. Fischer, T. Fellner, J. Wilde, H. Beyer, and R. Janke, “Analyzing calibration lifetime,” J. Ambient Intell. Humanized Comput., vol. 12,
parameters influencing stress and drift in moulded Hall sensors,” in Proc. no. 2, pp. 3091–3099, Feb. 2021, doi: 10.1007/s12652-020-02469-y.
1st Electron. Systemintegr. Technol. Conf., Sep. 2006, pp. 1378–1385. [70] J. M. Margarit-Taulé, M. Martín-Ezquerra, R. Escudé-Pujol,
[50] E. Stellrecht, B. Han, and M. G. Pecht, “Characterization of hygro- C. Jiménez-Jorquera, and S.-C. Liu, “Cross-compensation of FET
scopic swelling behavior of mold compounds and plastic packages,” sensor drift and matrix effects in the industrial continuous monitoring
IEEE Trans. Compon. Packag. Technol., vol. 27, no. 3, pp. 499–506, of ion concentrations,” Sens. Actuators B, Chem., vol. 353, Feb. 2022,
Sep. 2004. Art. no. 131123.
BERGER et al.: BAYESIAN SENSOR CALIBRATION OF A CMOS-INTEGRATED HALL SENSOR 6989

[71] M. Abdolrazzaghi, M. H. Zarifi, W. Pedrycz, and M. Daneshmand, [95] A. P. Singh, S. Kumar, and T. S. Kamal, “Virtual compensator for
“Robust ultra-high resolution microwave planar sensor using fuzzy correcting the disturbing variable effect in transducers,” Sens. Actuators
neural network approach,” IEEE Sensors J., vol. 17, no. 2, pp. 323–332, A, Phys., vol. 116, no. 1, pp. 1–9, Oct. 2004. [Online]. Available:
Jan. 2017. https://www.sciencedirect.com/science/article/pii/S0924424704001955
[72] N. Zimmerman et al., “A machine learning calibration model using
random forests to improve sensor performance for lower-cost air
quality monitoring,” Atmos. Meas. Techn., vol. 11, no. 1, pp. 291–313,
2018. [Online]. Available: https://amt.copernicus.org/articles/11/
291/2018/ Moritz Berger (Member, IEEE) received the
[73] P. Nowack, L. Konstantinovskiy, H. Gardiner, and J. Cant, “Machine M.Sc. degree in microsystems engineering from
learning calibration of low-cost NO2 and PM10 sensors: Non-linear the Department of Microsystems Engineer-
algorithms and their impact on site transferability,” Atmos. Meas. ing (IMTEK), University of Freiburg, Germany,
Techn., vol. 14, no. 8, pp. 5637–5655, 2021. [Online]. Available: in 2017, where he is currently pursuing the Ph.D.
https://amt.copernicus.org/articles/14/5637/2021/ degree in collaboration with Melexis Technolo-
gies SA, Bevaix, Switzerland.
[74] F. M. Heckmeier and C. Breitsamter, “Aerodynamic probe calibration
In 2015, he gained practical experience as
using Gaussian process regression,” Meas. Sci. Technol., vol. 31, no. 12,
a Research and Development Intern at Silicon
Oct. 2020, Art. no. 125301, doi: 10.1088/1361-6501/aba37d.
Microstructures, Milpitas, CA, USA, on pressure
[75] S. Urban, M. Ludersdorfer, and P. van der Smagt, “Sensor calibration sensors. His MEMS-related research interests
and hysteresis compensation with heteroscedastic Gaussian processes,” have been in new methods for the calibration of multisensor systems.
IEEE Sensors J., vol. 15, no. 11, pp. 6498–6506, Nov. 2015. Mr. Berger was awarded the Third Prize of the International Contest on
[76] G. Tancev and F. G. Toro, “Variational Bayesian calibration of Applications of Nano-Micro Technologies (iCan) at Transducers 2015 in
low-cost gas sensor systems in air quality monitoring,” Meas., Anchorage, AK, USA, with a team of students of the University of
Sensors, vol. 19, Feb. 2022, Art. no. 100365. [Online]. Available: Freiburg.
https://www.sciencedirect.com/science/article/pii/S2665917421003287
[77] C. M. Bishop, Pattern Recognition and Machine Learning, 6th ed.
New York, NY, USA: Springer, 2006.
[78] C. E. Rasmussen and C. K. I. Williams, Gaussian Processes for Machine
Christian Schott (Member, IEEE) received the
Learning (Adaptive Computation and Machine Learning). Cambridge,
M.Sc. degree in electrical engineering from
MA, USA: MIT Press, 2006.
the Technical University of Karlsruhe, Germany,
[79] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. Cambridge, in 1992, and the Ph.D. degree in magnetic
MA, USA: MIT Press, 2016, http://www.deeplearningbook.org. sensors from the École polytechnique fédérale
[80] M. Berger, S. Huber, C. Schott, and O. Paul, “Half-blind calibration for de Lausanne (EPFL), Lausanne, Switzerland,
the efficient compensation of parasitic cross-sensitivities in nonlinear in 1999.
multisensor systems,” IEEE Sensors J., vol. 19, no. 16, pp. 7005–7014, After the successful development of integrated
Aug. 2019. Hall sensors with magnetic concentrators at
[81] A. Albert, Regression and the Moore–Penrose Pseudoinverse. New York, Sentron AG, Zug, Switzerland, he joined Melexis
NY, USA: Academic, 1972. Technologies SA, Bevaix, Switzerland, in 2004,
[82] A. C. Atkinson, Optimal Design. Hoboken, NJ, USA: Wiley, 2015, contributing to make the triaxis technology a global success. Following
pp. 1–17, doi: 10.1002/9781118445112.stat04090.pub2. different technical and management positions, he founded the Melexis
[83] B. Smucker, M. Krzywinski, and N. Altman, “Optimal experimental Global Intellectual Property Team in 2013 and successfully led it through
design,” Nature Methods, vol. 15, no. 8, pp. 559–560, Jul. 2018, doi: various court litigations and license negotiations. Since 2018, he has
10.1038/s41592-018-0083-2. been working as a Process and Test Architect and preparing the
[84] J. Lever, M. Krzywinski, and N. Altman, “Points of significance: Model future of magnetic testing at Melexis Technologies SA. He holds over
selection and overfitting,” Nature Methods, vol. 13, no. 9, pp. 703–704, 30 patents and has published over 40 articles.
2016.
[85] M. Krzywinski and N. Altman, “Multiple linear regression,” Nature
Methods, vol. 12, no. 12, pp. 1103–1104, 2015.
[86] J. Lever, M. Krzywinski, and N. Altman, “Points of significance: Oliver Paul (Senior Member, IEEE) received the
Regularization,” Nature Methods, vol. 13, no. 10, pp. 803–804, 2016, Diploma degree in physics and the D.Sc. degree
doi: 10.1038/nmeth.4014. from ETH Zürich, Switzerland, in 1986 and 1990,
[87] M. Berger, F. Becker, and O. Paul, “Guidelines for multisensor system respectively.
calibration with and without regularization,” in Proc. 20th Int. Conf. After postdoctoral work at the Fraunhofer
Solid-State Sens., Actuators Microsyst. Eurosensors XXXIII (TRANS- Institute for Solar Energy Systems, Freiburg,
DUCERS EUROSENSORS XXXIII), Jun. 2019, pp. 2103–2106. Germany, he joined the Physical Electronics
[88] A. Gelman, J. B. Carlin, H. S. Stern, and D. B. Rubin, Bayesian Data Laboratory, ETH Zürich, as a Lecturer and a
Analysis, 3rd ed. New York, NY, USA: CRC Press, 2013. Group Leader in 1992. Since 1998, he has
[89] P. Virtanen et al., “SciPy 1.0: Fundamental algorithms for scientific been a Full Professor with the University of
computing in Python,” Nature Methods, vol. 17, pp. 261–272, Feb. 2020. Freiburg, Germany, where he heads the Labora-
tory for Microsystem Materials, Department of Microsystems Engineer-
[90] J. Wang, “Electrochemical glucose biosensors,” Chem. Rev., vol. 108,
ing (IMTEK), Faculty of Engineering. He is the Co-Founder of Sensirion
no. 2, pp. 814–825, Dec. 2007.
AG, Stäfa, Switzerland, and Atlas Neuroengineering, Leuven, Belgium.
[91] A. Weltin, J. Kieninger, and G. A. Urban, “Microfabricated, amperomet- He was the Director of IMTEK and the Dean of the Faculty of Engineer-
ric, enzyme-based biosensors for in vivo applications,” Anal. Bioanal. ing, University of Freiburg, from 2006 to 2008 and from 2016 to 2018,
Chem., vol. 408, no. 17, pp. 4503–4521, Jul. 2016. respectively. He was the Founding Director of the German Cluster of
[92] E. Foxlin and L. Naimark, “Miniaturization, calibration & accuracy Excellence BrainLinks-BrainTools, University of Freiburg. He holds vari-
evaluation of a hybrid self-tracker,” in Proc. 2nd IEEE ACM Int. Symp. ous advisory positions at the University of Freiburg. He is a coauthor of
Mixed Augmented Reality, Tokyo, Japan, 2003, pp. 151–160. [Online]. more than 400 technical publications, patents, and books. The research
Available: http://ieeexplore.ieee.org/document/1240698/ of his group focuses on MEMS materials and fabrication technologies,
[93] Q. Yu and X. Zhou, “Pressure sensor based on the fiber-optic extrinsic physical microtransducers, and microstructures for industrial and life
Fabry–Pérot interferometer,” Photonic Sensors, vol. 1, no. 1, pp. 72–83, science applications.
Mar. 2011, doi: 10.1007/s13320-010-0017-9. Dr. Paul has been a member of the Editorial Board of Sensors and
[94] V. Cimini et al., “Calibration of multiparameter sensors via machine Actuators A: Physical and Journal of Micromechanics and Microengi-
learning at the single-photon level,” Phys. Rev. A, Gen. Phys., neering and the Editorial Advisory Board of the IEEJ Transactions
vol. 15, Apr. 2021, Art. no. 044003, doi: 10.1103/PhysRevApplied. on Electrical and Electronic Engineering. He co-chaired the IEEE
15.044003. MEMS 2004 Conference.