We are IntechOpen,

the world’s leading publisher of

Open Access books
Built by scientists, for scientists

6,200
Open access books available
169,000
International authors and editors
185M
Downloads

Our authors are among the

154
Countries delivered to
TOP 1%
most cited scientists
12.2%
Contributors from top 500 universities

Selection of our books indexed in the Book Citation Index

in Web of Science™ Core Collection (BKCI)

Interested in publishing with us?

Contact [email protected]
Numbers displayed above are based on latest data collected.
For more information visit www.intechopen.com
Chapter

Applied Mathematics Tools in

Digital Transformation
Francesco Calabrò, Maurizio Ceseri and Roberto Natalini

Abstract

Digital transformation is a process that companies start with different purposes.

Once an enterprise embarks on a digital transformation process it translates all its
business processes (or, at least, part of them) into a digital replica. Such a digital
replica, the so-called digital twin, can be described by Mathematical Science tools
allowing cost reduction on industrial processes, faster time-to-market of new products
and, in general, an increase of competitive advantage for the company. Digital twin is
a descriptive or predictive model of a given industrial process or product that is a
valuable tool for business management, both in planning—because it can give differ-
ent scenario analysis—and in managing the daily operations; moreover, it permits
optimization of product and process operations. We present widespread applied
mathematics tools that can help this modeling process, along with some successful
cases.

Keywords: data mining, digital twin, modeling simulation optimization (MSO),

numerical linear algebra, scientific machine learning

1. Introduction

What is digital transformation? According to Ebert and Duarte [1], Digital Trans-
formation (DX) is “about adopting disruptive technologies to increase productivity,
value creation, and the social welfare”. The above definition is quite general meaning
that DX can be adopted in many circumstances and by many actors: governments
(from local to national level), multilateral organizations, industries. Digitalization,
and in particular the concept of Digital Twin, has applications in many different fields
[2]: Health, Meteorology, Manufacturing, Education, Cities and Transportation, and
Energy.
Indeed, DX is having a huge impact on society at large. For instance, concerning
labour force, a report from the World Economic Forum [3] states that by 2025, as a
consequence of digitalization, 85 million jobs will be destroyed, while 97 million new
jobs will be created worldwide. Thus, new competencies and skills will be necessary in
a digitally transformed world. Among the top skills listed in the WEF report are the
following: Critical thinking and analysis, Active learning and learning strategies,

1
Digital Transformation - Towards New Frontiers and Business Opportunities

Analytical thinking and innovation, Complex problem-solving, Systems analysis and

evaluation. We can observe that a person with a background in mathematics possesses
the above skills. More generally, STEM education will offer great opportunities in DX.
A recent study [4] shows that in the US the automation and the use of robots in the
productive environment has increased the enrolment at Universities and higher
education levels in the field like Computer Science and Engineering. This shows an
increasing awareness of the workforce of the importance of updating labour
competencies.
This paper focuses in particular on the digital transformation of Industry.
Throughout this paper, the term “Industry” means business and commerce, public
and private research, development, and production facilities; in practice, all the
activities that lie outside the field of academic research and education. For the
industrial sector, DX leads to huge changes in how a company is managed and can also
significantly affect customer satisfaction and product quality. The effects of digital
transformation can be seen at many levels of an organization: the way employees
work, how business processes operate, and how to collect, analyze, and use data. All
the above considerations imply that DX does not mean simply the digitization of
information. This is just a first indispensable step of a bigger transformation of the
way a company is managed: in short, DX requires a digital culture within an
enterprise [5].
Digitalization is enabled by a series of technologies:

• IoT: different objects in a system (a product or a manufacturing environment)

are connected by sensing devices to the Internet and can interact with one
another; this allows the control of the system throughout its life cycle.

• Big Data: the huge amount of data collected every moment on the system are a
powerful tool to understand the process, and extract information.

• Cloud: the data collected has to be stored properly and safely and made available
to the users; the user, in turn, may operate on the data through, for instance,
simulations to understand what is happening on the system under control.

But one of the most important technologies enabling the Digital Transformation is
Mathematics. Modeling, Simulation, and Optimization methods (MSO) have demon-
strated their usefulness for solving problems in real life: forecast of air pollution,
image processing, filtration processes, cultural heritage conservation, just to mention
a few. MSO are becoming very important in a digitalized world, since they make it
possible to extract information and knowledge from the data collected. In fact, in the
last years research have developed new mathematical tools to be applied, for example,
in Digital Twins [6, 7]. And some authors have talked explicitly of an “Era of Mathe-
matics” [8].
This paper will give an overview of the mathematical tools that can be applied in a
digitally transformed enterprise. It is organized as follows. Section (2) will describe
the two main approaches of applied mathematics to digitalization: Physics-Based and
Data-Driven. We will depict the main differences and how they can be combined to
increase their effectiveness, with an example from an Industrial case. Section 3 will
introduce some tools from linear algebra that can be applied to process collected data.
Finally, in a concluding section, we will summarize our points and stress the impor-
tance of promoting Mathematics towards Industry.
2
Applied Mathematics Tools in Digital Transformation
DOI: http://dx.doi.org/10.5772/intechopen.103806

2. Physics-based versus data-driven approach: competition or

collaboration?

The Digital Transformation implies the use of mathematical models and methods
to take advantage from the (possibly huge) amount of data collected by an enterprise
about its own processes. When an enterprise wants to digitalize its processes it may
apply a physics-based approach or a data-driven approach. The above approaches
imply different methods to deal with a problem, each one with its pros and cons [9].
In recent years, however, Hybrid methods have been developed building on the two
approaches. The hybrids have been developed in such a way to sum up the advantages
of both model-based and data-driven approaches while reducing the disadvantages.

2.1 Physics-based

With the Physics-Based (PB) approach, we mean a description of the process or

device based on first principles. Thus, the digital counterpart of the process is a
mathematical model that describes the physics of the system taking into account all
the relevant scales. Such an approach is also named Model-Based to emphasize the role
of mathematical modeling.
Thus given a system to be digitized, the enterprise has to translate it in a mathe-
matical model with a variable degree of complexity. The complexity of the model
depends on the problem at hand and on the business objectives. The model can be
used to monitor the performances of the system and then decide how to manage
possible anomalies. How possible anomalies can be detected? Sensors are positioned
on the system to collect data about its functions. The data are then compared with
model output evaluating the residual, i.e.

ri ¼ ∣ui di ∣, for i ¼ 1, … , N (1)

where ui is the outcome of the mathematical model, di is the data collected by the
sensors, and N is the number of data collected. If the residual exceeds a given
threshold, then the system does not “behave properly” and some countermeasure has
to be implemented.
Physics-based models can be powerful tools to create a Digital Twin. They can give
great support in the first design of the digital system; in fact, they can give useful
information on the processes to be described without great moles of data. As an
example, physics-based modeling can give a first idea on what variables shall be
monitored.
Physics-based models are characterized by transparency: this makes straightfor-
ward the interpretation of results.
Generalization is another important characteristic of PB models. The underline
assumptions, as well as the approximations due to the development of the model
determine the extent to which a model, can be generalized to other cases not previ-
ously considered. In general, in a PB approach, the limits of applications of a model are
known in advance and this makes clear the range of problems it can describe.
One of the main problems of such an approach is the complexity of the resulting
system of equations; in the presence of multiple processes interacting with one
another (multi-physics systems) and at multiple spatial and/or temporal scales (multi-
scale systems), the difficulty in solving the equations will increase very rapidly. Some
techniques have been developed, such as Reduced Order Modeling, to diminish model
3
Digital Transformation - Towards New Frontiers and Business Opportunities

and computational complexity and end up with a more manageable system of equa-
tions, while maintaining the accuracy of results [10].
Nevertheless, such an approach is the one that permits a automotive industry to do
crash test only via numerical simulations and new applications of the PB are growing
up. For example, we are now able to simulate the whole heartbeat where all the
physical parts (electrophysiology, the passive and active mechanics of the cardiac
muscle, the microscopic force generation in sarcomeres, the blood flow in the heart
chambers, and the valve dynamic) concur to the final simulation, see [11].
PB approaches have been applied also in the field of building automation: in [12]
the authors focused on predictive maintenance of biomass boilers trying to minimize
user discomfort.

2.2 Data-Driven

Data-Driven (DD) approach implies the use of methods from Machine Learning or
even Deep Learning to exploit directly the data collected by sensors on system per-
formance. Artificial Intelligence tools are suitable when very high amounts of data are
available and can be used to find hidden patterns in the sample that cannot be
discovered otherwise. Such a pattern can be refined whenever new data are collected.
In such a way, we can say that the tool “learns from experiences”.
The methods applied are diverse and comprises the following:

• Support Vector Machines

• Artificial Neural Network

• Convolutional Neural Network

• Recurrent Neural Network

• Generative Adversarial Network

The main advantage of the DD approach is the ease with which the solutions can be
found, compared with the first principle approach. They have proven to be a very
good alternative to PB models. If enough data are available, tools like neural networks
can find hidden structures in very complex problems, where it is difficult to describe
exactly the underlined physical structure. To give an example in the field of Material
Sciences, in [13] the authors developed a Neural Network to simulate a Cellular
Mechanical Metamaterial (CMM) and compared the performances with Direct
Numerical Simulations (DNS), usually applied in this kind of problem. While the DNS
needs about 105 Degrees of Freedom and 5
102 s to describe the mechanical proper-
ties of CMM, the Neural Network scheme needs only 290 DoF and about 6 to 8 s to
provide a solution.
The ability to generalize a DD model is very limited. A DD model can describe
circumstances only spanned by the data already available. It will need a large
number of new data to take into account a more general setting. The need for
(possibly large) amount of information is an overall limit to the development of the
DD approach; sometimes, one does not have sufficient data to describe properly a
given process.
4
Applied Mathematics Tools in Digital Transformation
DOI: http://dx.doi.org/10.5772/intechopen.103806

Another problem is the lack of transparency. DD model builds in general so-called

black-boxes that do not permit a clear interpretation of results. This led to the devel-
opment of explainable AI [14].
According to Gartner [15], thirty-seven percent of companies have
implemented AI in some form, translating to a 270 percent increase over the last
four years. Customers are accustomed to bots and AI mechanisms that provide,
among others, recommendations on services such as Netflix or Spotify. Not
surprisingly, 31 percent of companies plan to increase the share of artificial
intelligence in their business.

2.3 Hybrid approach

Recently, some researchers attempted to compare the two approaches in order to

understand the advantages and the disadvantages of both. For example, in [16] the
authors compared PB and DD approaches to the problem of fault detection and
isolation in an automotive application. They found the same performances for both
approaches in terms of detection and robustness and described the main shortcomings
of both. However, Moallemi and coauthors [17] applied both Model Based and DD
models to the problem of Structural Monitoring of a building; they found that Model
Based approach results in best performances and point out that scarcity of data limits
the behavior of DD algorithms. Thus, the debate is still open on what is the best
approach for digitalisation.
The recent tendency, however, is to take advantage of the benefits of both by
building hybrid methods that mix together. First principles with data-intensive
modeling [18–20]. In [21] the authors survey the research literature on this topic and
propose a taxonomy of the developed methods. The integration of PB and DD
approaches has several applications. PB models of complex processes require approx-
imations to have a usable set of equations. Approximations are necessary whenever
the process is not completely understood in all of its elements. This will introduce bias
in model results. Another issue is due to the presence of physical parameters that have
to be estimated some time with a small amount of data. On the other hand, Machine
learning tools need a large amount of data to reproduce with precision a given process
and have limited generalization capabilities. The use of DD techniques in combination
with first principle modeling may help to overcome such limitations and this has been
shown to be very useful in practice.
One attempt is the so-called Physics-Informed Neural Networks (PINN) [22]: a deep
learning framework for the resolution of PDE in several application (fluid dynamics,
quantum physics, reaction–diffusion systems, etc.). Neural networks are defined by
minimizing a loss function incorporating physical constraints. If one wish to solve an
equation of the type AðuÞ ¼ 0, the solution is given by the minimum of the following
function
N M   
1X 2 1 X 2
jAðuðxi ÞÞj þ u x j u j (2)
N i¼1 M j¼1
 
where x j , u j for j ¼ 1, … , M are training data (including initial and boundary
conditions), and ðxi Þ for i ¼ 1, … , N are points where the solution is defined.
The interplay of Pb and DD approaches has been employed in Unmanned Aerial
Vehicle (UAV) management; in [23], the authors present a Component Based
5
Digital Transformation - Towards New Frontiers and Business Opportunities

Reduced Order Model coupled with Baysian state estimation to allow a small aircraft
to adjust dynamically its trajectory.
Hybrid methods have been applied to describe a spring-mass system subjected to
damping that evolves according to multiple time scales [24]; such an application
is relevant for many industrial settings and is generally very difficult to solve
numerically.

2.4 An example in predictive maintenance

Through the technologies described it is possible to solve usual problems of engi-

neering and computational mechanics but also to explore new fields in which up to
now mathematical modeling had not been able to replace the experience or the direct
control of the mechanism. One of the most interesting areas is that of predictive
maintenance [25], already mentioned above. In some sectors, such as in aircraft
maintenance, the benefits of having guidelines available to ensure the correct func-
tioning of circulating machines are well known: in civil aeronautics, periodic checks
and periodic replacements, for example, of bolts or panels subjected to greater stress is
an established practice, but this practice comes from the combination of experience
and modeling. On the other hand, it is well known that it is not at all easy in a complex
system to keep all the components under control to have guarantees on correct
functioning and to prevent possible failures and/or breakages. Machines are often
equipped with many sensors that monitor correct operation, signaling anomalies on
what each sensor is capable of measuring. Unfortunately, it is often only from the
combination of information that the general picture is obtained, and therefore if only
the single sensor is monitored, this information collected is not really helpful in
decisions. Some industrial sectors, such as boating or highly automated production,
would greatly benefit from the possibility of preventing critical situations before
failures occur that can put in danger the correct functioning of the machinery. The
modeling available to complex tools—which work on different scales and interact
with each other—often has to overlook some effects and does not allow predictive
models so accurate as to be able to help decision makers determine if or when to
intervene with substitutions or repairs. Most of the time one can notice a problem
when it is already too late.
The recent literature on predictive maintenance foresees the possibility of com-
bining the effect of a control on the physics of phenomena, possibly linked to the
information of the sensors, and of an experience given by the analysis of the avail-
able data processed through artificial intelligence. For example, in [26] the authors
describe the use of hybrid techniques for controlling the operation of Computer
Numerical Control (CNC) machines. These are fundamental briks in modern indus-
tries (from Wikipedia): “CNC is the automated control of machining tools (such as
drills, lathes, mills and 3D printers) by means of a computer. A CNC machine
processes a piece of material (metal, plastic, wood, ceramic, or composite) to meet
specifications by following coded programmed instructions and without a manual
operator directly controlling the machining operation.” Authors in [26] present how
it is possible to build a digital twin that provides information on the expected
behavior of the machine and at the same time, through the data collected by the
sensors inside the machine, the on-situ behavior of the machine. Then, by taking a
comparison of the two, this reveals a good ability to highlight anomalies, based on
the physical modeling. At the same time, however, the acquisition of a large amount
6
Applied Mathematics Tools in Digital Transformation
DOI: http://dx.doi.org/10.5772/intechopen.103806

of data through sensors allows the elaboration of a database that also includes a sort
of forecasting model: this increases the knowledge of the ongoing process and helps
in prediction.
The use of hybrid approach helps to meet the targets of predictive maintenance:

• Fault diagnosis

• Fault forecast

• Intelligent decision

• Intelligent maintenance.

3. Applied linear algebra tools

Now we consider how the tools coming from linear algebra can manage to solve
the problems considered in the previous section. We consider both problems, namely
the search of a model that can justify some available data and the case where the
model is given and some prediction is done via numerical computations. What we
present next is inspired by [27, 28], where readers can find full details of the topics
introduced in this section.

3.1 Modeling with linear functions

First, we consider that a dataset is given; instances xðiÞ , yðiÞ are known ∀i ¼
 

1, … , M. As previously discussed, such dataset is described with numbers, and we

identify it as input x and y is the output. Each of the inputs is vectors of numbers
xðiÞ ∈ Rd while we consider the output to be a number yðiÞ ∈ R. The dataset then can be
described as an input matrix D ∈ RMd whose rows are the instances and the columns
are the features, and a column vector Y that stores the output.
Our modeling problem is to describe the dependence of the output from the input,
so to identify a mathematical law F : Rd ! R in good agreement with the available
dataset F xðiÞ ≈ yðiÞ . The most simple mathematical law that we can introduce would
 

be the linear function. This is completely described by the coefficients w ∈ Rd : F ðxÞ ¼

Pd
j¼1 w j x j . If we impose the agreement with the data, we have that the unknown
coefficients solve the problem:

Find w ∈ Rd such that Dw Y ≈ 0
 
Linear System PB

Such a problem, if settled out in the square (d ¼ M) and exact framework, is

the well-known resolution of linear systems. We prefer to write the search problem
in an approximate way, so we use the ≈ symbol, because usually we consider
inexact data or non-square datasets, so that existence of the solution is not
guaranteed. In the next section we discuss the resolution of the linear (approximate)
problem. Finally, when the w are calculated, for a new instance x ∈ Rd the model
would predict y ¼ dj¼1 w j x j .
P

7
Digital Transformation - Towards New Frontiers and Business Opportunities

3.2 Model approximation

When a model is given, this is used to predict the behavior of the quantity of
interest for forecasting, i.e., in unseen cases. Mathematical models are relations
involving, in many cases, operations that are not tabled or easily computable. In order
to understand how the tools of applied mathematics apply in such cases, we can start
from a general formulation where the quantity of interest is a u ∗ (an unknown
function in the general case or a vector/scalar) that is the solution of a problem:
P ðu ∗ , δÞ ¼ 0 where δ are the given data.

N
X  
N
Find α ∈ R such that uðxi Þ  α j f j xðiÞ yðiÞ ≈ 0 ∀i ¼ 1 … M
j¼1

To begin with a simple example, consider the approximation/extrapolation of data.

As in the previous section, the inputs are made of vectors of numbers xðiÞ ∈ Rd while
we consider the output to be a number yðiÞ ∈ R so that the known data are δ ¼
 ðiÞ ðiÞ 
x , y i ∀i ¼ 1, … , M. Eventually, the outputs yðiÞ are evaluations of a “black box”
function that we want to describe in a different way. In the previous section, we
considered the resolution of a liner problem in order to look for a linear model that
could justify the dependence of the output by the input data. In this case, we start
from different knowledge and look for a non-linear model that gives “exactly” the
dependence of y from x. The approximate solution will be a function u ≈ u ∗ that can
be described by a finite set of coefficients. We look for u such that u xðiÞ ≈ yðiÞ , where
 

uðxÞ ¼ Nj¼1 α j f j ðxÞ, i.e., u is a linear combination of N fixed functions f j ðxÞ. The
P

coefficients of the linear combination, α ∈ RN determinate the solution of our approx-

imated problem u. One way to fix such coefficients is to impose accordance with the
available data: Nj¼1 α j f j xðiÞ ¼ yðiÞ .
P  

This problem is exactly in the same framework as the ones seen in the previous
section: a matrix problem where the matrix is the so-called collocation matrix
Φ ∈ RMN , Φ j,i ¼ f j xðiÞ .
 

Remark. The choice of the functions f j strongly affects both the quality of the
solution and the ability to solve the discrete problem easily. Usual choices are poly-
nomials or piece-wise polynomial functions. Nevertheless many systems of functions
can be used instead and the results on the capability to approximate general functions,
known as the Universal Approximation Theorem (UAT), are available. One of the
reasons why the Neural Network (NN) functions are widely used in numerical
approximations is due to the fact that there are available UAT results in this case. We
remark that in the general case of NN also some parameters inside the functions can
be tuned so that the overall determination of the network is a nonlinear optimization
problem.
Moving from the approximation of a given function to different scientific models,
we encounter the case of equilibrium laws based on first principles, as described in the
previous section on Physics-Based approach. Also in this case the unknown is a
function but in this case, the model is written as a solution to a mathematical problem
P ðu ∗ , δÞ ¼ 0 that involves some operators applied to the unknown function: deriva-
tives and/or integrals (that are linear operators) and eventually nonlinear transfor-
mations. (The problem of interpolations described at the beginning of this section can

8
Applied Mathematics Tools in Digital Transformation
DOI: http://dx.doi.org/10.5772/intechopen.103806

be sought as the resolution of a mathematical problem that involves only evaluation of

functions on sites, which is a linear functional.) Also in the case where only linear
operators are considered, approximated problems for integration and derivation are
needed because both operations are intrinsically infinite-dimensional [29, 30]. In
general, two main roads—or combinations of these two—can be taken for the numer-
ical resolutions of problems that involve the use of operators:

• instead of considering general functions, look for a solution written as a linear

combination of simple functions where the operation can be done simply ) u ¼
α i ϕi ≈ u ∗ ;
P

• instead of considering exact modeling approximate the operators and look for
solutions in particular cases, ad example in fixed sites ) P N ≈ P.

Applied mathematics tools aim to introduce methods that translate the original
model in some approximation that is consistent with the original formulation and
gives an accurate approximation of the unknowns. The initial problem P ðu ∗ , δÞ ¼ 0 is
then reformulated in an approximate way (P N ≈ P) and the sought solution is an
approximation of the mathematical solution (u ≈ u ∗ ). The final resolution step, in
most cases revels in the resolution of a linear problem as the ones seen before.

3.3 Numerical resolution of linear systems

Because many problems seen in this excursus are finally modeled via linear systems,
we aim to present some of the ingredients that can be used for the resolution of linear
systems. The problem written in (Linear System PB) is the one that we aim to solve.
As commented, the unknown vector w has to solve the problem of interpolating
ð jÞ
the available data with a linear function, i.e., if Di , j ¼ 1, … , d are the features of
P ð jÞ
the i th individual and Y i is the output, then w j are such that j w j Di should be
close to Y i ∀i ¼ 1, … , M. The first case that we consider, is the one where we have
more parameter to fix than available information, so that the number of features d is
greater of the number of individuals M; the problem is also referred as overpara-
metrized.
First of all we introduce the rank r ≤ d of the matrix D: this is the number of
columns of the matrix that are linearly independent, i.e., columns that cannot be
written as a linear combination of the other ones. Thinking of the matrix as a collec-
tion of features from individuals, the rank would be the number of information that
we collect that are linearly independent. Because our aim is to find the coefficients w j
that solve the linear system, if we call Dð jÞ the j th column of D, the linear system can
w j D ð jÞ
P 
be written as Y ≈ 0 thus only columns that are linearly independent
give useful information for the determination of the w. The real dimensionality of the
problem, the number of degrees of freedom that can be sought, is then the rank of the
matrix.
An important feature when data are collected with some error—or “noise”—is that
the rank of the matrix could be affected by this error: some columns can be linearly
independent only for effect of the randomness of the noise. Then, a different prelim-
inary analysis—the so-called feature selection—that reduces the dimension of the
columns could be necessary. One possible way to proceed is by exploring the Singular
9
Digital Transformation - Towards New Frontiers and Business Opportunities

Value Decompostion of the matrix and, in particular, the order of magnitude of the
singular values. All these procedures that avoid the presence of redundancy on the
feature collection, end up with the so-called active features. When this preprocessing
of the data is done, we need to solve the problem of the determination of d ∗ ≤ d
coefficients w ∗ to be fixed from the M linear equations: a linear problem of the type
D ∗ w ∗ Y ≈ 0. If d ∗ is still greater or equal to M, the problem—which is referred as
square if d ∗ ¼ M or underdeterminated if d ∗ > M- has always an optimal solution,
i.e., a solution to the matrix problem D ∗ w ∗ Y ¼ 0. If the problem is square, the
solution is unique. It has more solutions when d ∗ > M: to select between these solu-
tions we can introduce the square problem ðD ∗ ÞT D ∗ w ~ ¼ ðD ∗ ÞT Y, where ð
ÞT stands
for the transposed matrix. Such equations are referred as “normal equations” for the
non-squared linear problem. The solution w ~ solves a square linear system and is one of
the possible vectors w ∗ .
Overdetermined linear problems are the ones where the number of features (or
active features as described above) are less than the number of individuals, that is, the
number of rows of the matrix M; the problem is also referred as overfitted. To focus on
such an overdeterminated case, we can focus on the scalar case, where d ¼ 1. The
coefficient w ∈ R is, then, the slope of the straight line that we want to construct in order
to give a law of direct-proportionality between x and y: wxi ¼ yi . This problem admits a
 
solution only if M ¼ 1 or if the points xi , yi are aligned (and aligned to ð0, 0Þ).
In the general cases, in order to solve the linear problem D ∗ w ∗ Y ≈ 0 one has to
introduce an optimality condition and try to solve the problem of “best fit”. After
some computation, what turns out to be a good solution is the solution of the normal
equations as introduced in the underdeterminated case.

4. Conclusions

What about the need for companies of such technologies? What we experience
every day is that new technologies are deeply changing markets, as if this is “second
wave of digital transformation”. In fact, Robotics, Machine Learning, Big Data and
Artificial Intelligence are becoming more and more common.
Using properly DX, companies have the opportunity to optimize operating costs
and thus increase process efficiency via new services. Additionally, the digitalization
of company data produces the so-called “digital capital” made of huge amounts of
information from various sources, which increases the possibility to settle out predic-
tive models. This allows companies to better meet customers’ needs. DX is the present
for companies that aim to remain competitive, and the use of simple mathematical
tools opens up great opportunities in unexpected fields.
However, Enterprises are not completely aware of this point and how they would
benefit from the application of Mathematical research in their daily operations. In the
last few years, a new professional figure emerged world wide to promote:

• MSO towards enterpises, and

• Industry-Academic Collaborations.

The above role is played by the Technology Translator. A Technology Translator has
strong competencies of Mathematical research and on the other hand good knowledge
10
Applied Mathematics Tools in Digital Transformation
DOI: http://dx.doi.org/10.5772/intechopen.103806

of industrial processes; thus, he/she is the right person to talk both with mathemati-
cians and enterprises. Since Academics and Industrialists have different languages,1
the need has emerged for a professional figure that can translate Industrial problems
in mathematical terms enabling the cooperation among the two world. There are
several Institutions in Europe employing Technology Translators. In Italy, the
National Research Council of Italy promotes the project Sportello Matematico per
l’Innovazione e le Imprese.2 The main objectives of Sportello Matematico are: Promoting
Mathematical Technologies as source of Industrial Innovation; Activating
cooperations between Enterprises and Research Centers; Facilitating the employment
of Mathematicians in Industry [31]. The work of Technology Translators contributes
to the change of perspective required by the complex challenges of Digital
Transformation.

1
This is also related to the fact that Industry and Academia have different objectives and timescales.
2
www.sportellomatematico.it.

Author details

Francesco Calabrò1*†, Maurizio Ceseri2† and Roberto Natalini2

1 Università di Napoli “Federico II”, Naples, Italy

2 Istituto per le Applicazioni del Calcolo “Mauro Picone”, Consiglio Nazionale delle
Ricerche, Rome, Italy

*Address all correspondence to: [email protected]

† These authors contributed equally.

© 2022 The Author(s). Licensee IntechOpen. This chapter is distributed under the terms of
the Creative Commons Attribution License (http://creativecommons.org/licenses/by/3.0),
which permits unrestricted use, distribution, and reproduction in any medium, provided
the original work is properly cited.
11
Digital Transformation - Towards New Frontiers and Business Opportunities

References

[1] Ebert C, Duarte CH. Digital Krishnaswamy S, et al. A simulation-

transformation. IEEE Software. 2018; based digital twin for model-driven
35(4):16-21 health monitoring and predictive
maintenance of an automotive braking
[2] Rasheed A, San O, Kvamsdal T. Digital system. In: Proceedings of the 12th
twin: Values, challenges and enablers International Modelica Conference,
from a modeling perspective. Ieee Access. Prague, Czech Republic, May 15-17, 2017.
2020;8:21980-22012 Vol. 132. Linköping: Linköping University
Electronic Press; 2017. pp. 35-46
[3] World Economic Forum. The Future
of Jobs Report. Geneva, Switzerland:
[11] Quarteroni A, editor. Modeling the
World Economic Forum; 2020
Heart and the Circulatory System.
[4] Di Giacomo G, Lerch B. Automation Berlin, Germany: Springer; 2015
and Human Capital Adjustment: The
Effect of Robots on College Enrollment. [12] Cauchi N, Macek K, Abate A. Model-
2021. Available from: SSRN 3920935 based predictive maintenance in
building automation systems with user
[5] Ivančić L, Vukšić VB, Spremić M. discomfort. Energy. 2017;138:306-315
Mastering the digital transformation
process: Business practices and lessons [13] Xue T, Beatson A, Chiaramonte M,
learned. Technology Innovation Roeder G, Ash JT, Menguc Y, et al. A
Management Review. 2019;9(2):36-50 data-driven computational scheme for
the nonlinear mechanical properties of
[6] Hartmann D, Van der Auweraer H.
cellular mechanical metamaterials under
Digital twins. In: Progress in Industrial
large deformation. Soft Matter. 2020;
Mathematics: Success Stories. Berlin,
16(32):7524-7534
Germany: Springer, Springer,
Champions; 2021. pp. 3-17
[14] Longo L, Goebel R, Lecue F,
[7] Hauret P. At the crossroads of
Kieseberg P, Holzinger A. Explainable
simulation and data analytics. European artificial intelligence: Concepts,
Mathematical Society Magazine. 2021; applications, research challenges and
16(121):9-18 visions. In: International Cross-domain
Conference for Machine Learning and
[8] Bond P. The Era of Mathematics; An Knowledge Extraction. Berlin, Germany:
Independent Review of Knowledge Springer, Champions; 2020. pp. 1-16
Exchange in the Mathematics Sciences.
2018. Available from: epsrc.ukri.org/ [15] GARTNER. What Is Artificial
newsevents/pubs/era-of-maths/ Intelligence? Avaiable from: https://
[Accessed: February 9, 2022] www.gartner.com/en/topics/artificial-
intelligence
[9] Rueckert D, Schnabel JA. Model-
based and data-driven strategies in [16] Yang R, Rizzoni G. Comparison of
medical image computing. Proceedings model-based vs. data-driven methods for
of the IEEE. 2019;108(1):110-124 fault detection and isolation in engine
idle speed control system. Annual
[10] Magargle R, Johnson L, Mandloi P, Conference of the PHM Society 2016 8 1
Davoudabadi P, Kesarkar O, PHM Society San Diego, California
12
Applied Mathematics Tools in Digital Transformation
DOI: http://dx.doi.org/10.5772/intechopen.103806

[17] Moallemi A, Burrello A, Brunelli D, [24] Chakraborty S, Adhikari S. Machine

Benini L. Model-based vs. data-driven learning based digital twin for dynamical
approaches for anomaly detection in systems with multiple time-scales.
structural health monitoring: A case study. Computers Structures. 2021;243:106410
In 2021 IEEE International
Instrumentation and Measurement [25] Ran Y, Zhou X, Lin P, Wen Y, Deng
Technology Conference (I2MTC) 2021 RA Survey of Predictive Maintenance:
May 17 1 6 IEEEManhattan, New York, US Systems, Purposes and Approaches.
arXiv preprint arXiv:1912.07383. 2019
[18] Liu Z, Meyendorf N, Mrad N. The role
of data fusion in predictive maintenance [26] Luo W, Hu T, Ye Y, Zhang C, Wei Y.
using digital twin. In: AIP Conference A hybrid predictive maintenance
Proceedings. Vol. 1. Melville, NY: AIP approach for CNC machine tool driven
Publishing LLC; 1949. p. 020023 by digital twin. Robotics and Computer-
Integrated Manufacturing. 2020;65:
[19] Shetty P, Mylaraswamy D, 101974
Ekambaram T. A hybrid prognostic
model formulation system identification [27] Deisenroth MP, Faisal AA, Ong CS.
and health estimation of auxiliary power Mathematics for Machine Learning.
units. In: 2006 IEEE Aerospace Cambridge: Cambridge University
Conference. Manhattan, New York, US: Press; 2020. Available from:
IEEE; 2006. p. 10 https://mml-book.github.io/

[20] Slimani A, Ribot P, Chanthery E, [28] Strang G. Linear Algebra and

Rachedi N. Fusion of model-based and Learning from Data. Cambridge:
data-based fault diagnosis approaches. Wellesley-Cambridge Press; 2019
IFAC-PapersOnLine. 2018;51(24):
1205-1211 [29] Calabrò F, Manni C, Pitolli F.
Computation of quadrature rules for
[21] Willard J, Jia X, Xu S, Steinbach M, integration with respect to refinable
Kumar V. Integrating physics-based functions on assigned nodes. Applied
modeling with machine learning: A Numerical Mathematics. 2015;90:
survey. arXiv preprint arXiv: 168-189
2003.04919. 2020;1(1):1-34
[30] Calabrò F et al. Null rules for the
[22] Raissi M, Perdikaris P, detection of lower regularity of
Karniadakis GE. Physics-informed neural functions. Journal of Computational and
networks: A deep learning framework for Applied Mathematics. 2019;361:547-553
solving forward and inverse problems
involving nonlinear partial differential [31] Bertsch M, Ceseri M, Natalini R,
equations. Journal of Computational Santoro M, Sgalambro A, Visconti F. On
Physics. 2019;378:686-707 the Italian network of industrial
mathematics and its future
[23] Kapteyn MG, Knezevic DJ, developments: Sportello Matematico per
Huynh DB, Tran M, Willcox KE. Data- l’industria italiana. In: European
driven physics-based digital twins via a Consortium for Mathematics in
library of component-based reduced- Industry. Berlin, Germany: Springer,
order models. International Journal for Champions; 2014. pp. 173-178
Numerical Methods in Engineering. 2020:
1-18
13