1. Introduction
The design of thermoelectric generators
using geometric optimization techniques is the topic of interest in this work. Here, we describe some advantages of this type of design. First, it allows us to make better use of the thermoelectric material; to date, the best materials have been inorganic compounds (such as
), which have a relatively low abundance on Earth, and their manufacture requires a highly complex vacuum process. It is also possible to gradually adjust the dimensions of the system according to the space available for its coupling to a heat source. One parameter used in the geometric method is the cross-sectional area of the leg [
1]. For example, it is possible to maximize the output power of the
as a function of the variable cross-section; in fact, in [
2], it was stated that “The geometry of the
has a vital impact on the thermal resistance and the electrical resistance, influencing its integral performance”.
Another aspect that motivated the development of this work is the emergence of new manufacturing techniques, such as additive manufacturing, which allow customizing thermoelectric systems according to the requirements of the field of application (from low-power applications, medical and wearable devices, Internet of Things, and wireless sensor networks; to high-power applications, industrial electronics, automotive engines, and aerospace). However, as rightly mentioned in [
3], “little knowledge exists about which shapes are beneficial in applications with different thermal conditions”. This author analyzed the effect of different thermoelectric leg designs on device performance. The authors in [
4] proposed an optimized design of legs with special geometric shapes and their manufacture using 3D printing to increase the output power of
.
In [
2], two relevant aspects were mentioned: (a) in the current literature, there are still few studies on the optimization of the geometry of
, especially for the shape of the legs; (b) recently, algorithms have been used to design optimal devices by simultaneously analyzing two or more geometric parameters.
The idea of geometric optimization comes from the structural aspect of the thermoelectric generator, which in its most elementary form (thermocouple) is composed of two legs of semiconductor material (for example, BiTe or SiGe), one
n-type and the other
p-type, which are electrically connected in series by means of a metallic bridge (for example, copper) (
Figure 1); these legs have a rectangular prism shape. Then, two geometric parameters present in each leg are identified, the cross-sectional area
and the length
. These geometric parameters are linked to two properties of thermoelectric materials, thermal conductance (K) and electrical resistance (R). This relationship between the geometric parameters and the thermoelectric properties is useful for adjusting the shape and size of the thermocouple. To achieve this dimensioning, a useful technique is the analysis of the electrical power produced by the TEG, in terms of the geometric parameters.
The main motivation for carrying out this work arose from identifying that currently, in the thermoelectricity field, there is a growth in the amount of researchers interested in developing
devices for harvesting energy from waste heat sources [
5,
6]. To achieve the maximum use of this heat, it is essential to analyze the geometric characteristics of the device. As previously mentioned, geometric optimization methods require knowledge of thermoelectric properties, which are linked to the dimensional parameters and that also vary with temperature. These conditions require knowledge of property measurements to achieve the best adjusted
design for the requirements imposed by the heat source, the available space, and the load resistance of the system that will use the power produced by the
. However, not all researchers attempting to design
have a materials laboratory or the equipment required to prepare samples and measurements. At present, a useful resource is simulation software to study the properties of materials and design
; even so, depending on the type of software, investment in licensing is required, in addition to high-end computer equipment.
Seeking to develop an affordable alternative for the community intending to design , in this work, we propose developing a methodology built on three principles:
(I) Use data from experimental measurements found in publications by specialists in thermoelectric materials; (II) Use a formalism or an approximation that allows the immediate use of the data obtained from the literature to obtain the leg dimensions ; (III) Merge the selected formalism or approximation with a prediction algorithm that allows us to generate the design best adjusted to the operating conditions of the environment.
The three previous principles guided us toward combining the reduced variables technique with supervised machine learning in a design process that consists of the following steps: (a) take advantage of the qualities of the reduced current approximation to obtain the architecture of the thermoelectric generator by calculating the parameters (cross-sectional area and length of its legs); (b) with the help of an algorithm , the values of the thermoelectric properties and the maximum electrical power are predicted for any temperature value; and (c) the values obtained in the prediction are useful for adjusting the design for the operating conditions. In addition, the algorithm has the outstanding feature of simultaneously analyzing and determining various parameters related to the geometry of the legs that maximize . These parameters are the cross-sectional area ratio , length , length n ratio , length p ratio (in the case of a segmented ), and the space between legs .
The results provide useful information for the construction of optimal devices and their possible applications. The scope of this work was extended to a new training phase, in which it is possible to introduce values of the parameter H and of the temperature, managing to predict the corresponding value of the maximum power. Of course, this study has limitations, which are discussed in later sections; however, these first results allowed us to determine that it is possible to obtain an algorithm for the design of conventional and segmented thermocouples based on a reduced variables approach fused with a supervised machine learning calculation model, trained for various thermoelectric materials. The utility and scope of the method are shown when confronted with a computerized and experimental model in
Section 6.
Notes on the State of the Art
The reduced variables approach is a strategy that can be applied for the optimization of power, efficiency, and even the coefficient of performance, modifying the current flow in the legs by adjusting the cross-sectional area. This feature allows us to model the architecture of the TEG, managing the size of the cross-sectional area and length of the legs. The scope of this tool was reported in the works of [
7,
8,
9,
10].
On the other hand, in the field of thermoelectricity, machine learning has been applied for the prediction of material properties and for the design of new materials. As [
11] rightly mentioned, “the machine learning technique can provide a powerful discovery tool for thermoelectric materials with respect to the new chemical composition, nanostructural design, stoichiometry optimization, etc.”. Specifically, supervised machine learning, which is based on algorithms that learn from an input dataset and a training dataset and manage to predict unseen data or future values—divided into two categories of classification and regression [
12]—has been applied to carry out the synthesis of a new material spin-driven thermoelectric effect (STE) [
13].
4. Temperature-Dependent Thermoelectric Properties and Supervised Machine Learning
Integrals [
5,
6] could be calculated thanks to the fact that we know the measurements of the thermoelectric properties at the temperatures in
Table 2, generating a first
design with their respective parameters
. However, the system obtained is specific for that temperature range, in such a way that it is not possible to adjust the design for a wider or narrower range. Trying to make a new design may require preparing a new material sample for new measurements or perhaps using advanced software for numerical design and simulation; there is a possibility that researchers do not have some of these resources. This difficulty can be overcome with the implementation of a supervised machine learning algorithm that allows predicting the values of the Seebeck coefficient,
; electrical resistivity,
; and thermal conductivity,
, properties at any temperature value.
Table 3 shows the measurements of these properties for the
material.
The data in
Table 3 were used to train the prediction code, for which an 80/20 rule was used; that is,
of the data were used for the training set and
for the test set. The code was generated using the Wolfram Mathematica 13.2 software using a predictive function that has the advantage of automatically selecting the most appropriate prediction model according to the behavior of the experimental data used for training. Specifically, the algorithm was applied to obtain a set of 50 data points for each thermoelectric property; however, more data can be predicted for any established temperature range. The distance between the data was reduced to
K.
Figure 2 shows a spreadsheet containing the prediction results of the thermoelectric properties generated by the algorithm for each of the materials selected for the design of the conventional generator.
This algorithm helps predict a material’s thermoelectric property value for any operating temperature value. As shown below, it is possible to calculate the maximum electrical power of a conventional system for any T value within this range.
4.1. Segmented Thermocouple
As a result of efforts to take advantage of heat in a wide temperature range, the segmentation technique was developed. It consists of joining segments of various thermoelectric materials and allowing thermal and electrical contact between them; its principle is based on the fact that each of the segments will be subjected to the temperature range in which it reaches its highest figure of merit value.
4.1.1. Design of a Segmented Thermocouple
The materials selected were
and
(
p-type) and
and
(
n-type). The operating temperatures selected were
K and
= 573 K.
Table 4 and
Table 5 show the values of the product
of the materials
and
(
p-type), respectively.
Table 6 and
Table 7 show the values of the product
of the materials
and
(
n-type), respectively.
Applying the treatment defined in the previous sections, the following parameters of the segmented thermocouple were obtained:
Table 8 contains the design parameters of the segmented thermoelectric generator. A design sketch is shown in
Figure 3.
Figure 4 shows a diagram that represents the interaction between the components in the system.
Again, the prediction algorithm was applied to obtain 50 data points for each of the thermoelectric properties. For each of the materials selected in the design, the distance between the data points was reduced to
K.
Figure 5 shows a spreadsheet containing the prediction results of the thermoelectric properties generated by the algorithm for each of the materials selected for segmented generator design.
This algorithm helps predict the values of the materials’ thermoelectric properties for any operating temperature value within the range of 398–573 K. As shown below, it is possible to calculate the maximum electrical power of the segmented system for any T value within this range.
4.1.2. Evaluation of the Feasibility of the Segmented Thermocouple by Means of the Compatibility Factor
An important aspect to remember in the design of segmented thermocouples is that only certain combinations of materials are appropriate, because there is a risk of building a thermocouple with a low efficiency, even lower than that of a conventional thermocouple. A useful resource for evaluating combinations of materials is the compatibility factor
[
10], defined as
Applying Equation (
15), it can be confirmed that the materials selected in the system studied in this work were correct for the segmentation; see
Table 9. In this table, it can be observed that the values of
S between
p-type and
n-type materials differed by a factor not greater than 2, as indicated by the rule.
6. Model Building and Experimental Setup
In order to verify the validity of the proposed methodology, in this section, we consider the results of the work of Crane et al. [
17], who created a design using a computerized model and performed an experiment with the built prototype of a system called a three-couple
engine; see
Figure 14 [
17].
The test carried out for this work consisted of using the information from the experimental measurements to design a conventional
; however, unknown data were necessary to correctly develop the design. The source of the known information was the data provided by Crane’s paper [
17], shown in
Table 12.
Figure 15 shows the power curves against the electric current for certain values of
, obtained through the computerized model made by authors in [
17]. It is observed that they coincide with the experimental curves.
Next, we sought to design a three-couple
engine system applying the reduced variables methodology with supervised machine learning, to obtain a graph like the one shown in
Figure 15. In addition to the data shown in
Table 12, it was required to know the properties and dimensions
,
,
,
,
,
,
,
, and
not provided by [
17], so it was necessary to calculate them.For this purpose, a first attempt was made, in which the Seebeck coefficient and electrical resistivity data from
Table 3 were used to obtain the graph of power at temperature
= 150 °C = 423.15 K, presented as the yellow curve in
Figure 15. In this case, the six temperature values that are provided in
Table 3 were the real values taken from the measurement and, as can be seen, the extreme values
= 298 K = 24.85 °C and
= 423 K = 149.85 °C were only approximate to the values actually required for the design (
= 293.15 K = 20 °C and
= 423.15 K = 150 °C). With this information, the reduced variables technique was applied to obtain a first design for the three-couple TEG engine, for which the graph of electrical power as a function of electrical current is shown in
Figure 16.
A comparison between
Figure 16 and the yellow curve in
Figure 15 shows that the design obtained in this first attempt deviated from Crane’s model by approximately
, which is a very high margin. It was therefore necessary to adjust the six temperature values from
Table 3 to the correct range of
= 298 K = 24.85 °C and
= 423 K = 149.85 °C. The following table shows the adjustment.
The values
,
,
,
,
, and
(
Table 2 and
Table 3) were used as training data for the supervised machine learning code implemented in this work and thus generated appropriate data from these properties for the temperature values from
Table 13. The results are shown in the following
Table 14.
Using the data from column two and applying the methodology with the set of Equations (
2)–(
12), the geometric parameters of the couples of the TEG engine were obtained, see
Table 15.
Subsequently, with the data from the third and fourth columns, the averages of the quantities
,
,
, and
in the temperature range (
= 293.15 K = 20 °C and
= 423.15 K = 150 °C) were obtained. The results are shown in
Table 16.
The data obtained for the legs of the
engine (
Table 15 and
Table 16), as well as the data of the metallic bridge (
Table 10), were used in combination with the following equation for the power produced by the
engine.
where
is the power produced by the TEG engine;
n is the number thermocouples, which in this case is
; and
I is the electric current (which is the independent variable and is measured in amperes). The other quantities that appear in Equation (
20) have already been indicated above. The graph of the electrical power produced by the three-couple
engine is shown in
Figure 17.
In
Figure 17, it can be seen that the graph obtained adjusted well to the original Crane curve at
= 150 °C (curve in yellow); see
Figure 14. It can be seen that the same maximum value was reached below
W around 8 A in a range of 0–15 A. Therefore, it is evident that the methodology (which combined the reduced variables method with supervised machine learning) proposed in this work managed to reproduce the three-couple
engine model and the experiment of Crane et al. [
17].
7. Conclusions
This work has shown that the fusion of supervised machine learning with the reduced variables technique can be a useful tool for designing and adjusting them to the conditions imposed by the operating environment of the system, specifically when facing the challenge of knowing a reduced set of measured values. Thanks to its ability to calculate geometric and prediction parameters, it is possible to
(a) Approximate values of thermoelectric properties for any temperature value;
(b) Generate new data from few experimental values, even when it is not possible to perform a new experiment;
(c) Design
for any range of temperature; if the value of a thermoelectric property for a specific temperature is not known experimentally, it is possible to predict it with
(
Figure 2 and
Figure 5) and use it for design calculations;
(d) Analyze power simultaneously with respect to various parameters for any temperature value and determine optimization conditions.
The previous characteristics were verified through the design and analysis of the power of the conventional
(
Section 3.1), segmented
(
Section 4.1), and
engine (
Section 6) systems.
When conducting the power analysis, one of the outstanding findings was the result shown in
Figure 11, indicating that the curve that corresponded to the highest temperature intersected with the curves of the other temperatures. Each intersection point specifically corresponds to the maximum performance of a segment. These intersection points occur very close to the value
in which the power of the segmented
is maximized.
As shown throughout this work, the proposed methodology was translated into a code in mathematics; its usability is now evaluated in terms of the following ten techno-economic aspects:
(1) The calculation scheme will be updated soon, to introduce a procedure based on heat transfer and to consider the physical aspects of the heat source;
(2) From an economic point of view, although Wolfram is a licensed software, it currently allows the user a free basic plan account, in which the code notebook could be published and shared with those interested in designing thermoelectric systems with this methodology;
(3) The code automatically selects the appropriate prediction method according to the training data. For the study of conventional and segmented systems, the methods that the code selected to predict the values were linear regression, decision tree, and first neighbors;
(4) The methodology could be transferred in code to another type of software that is freely available; for example, it could be implemented in python. In that case, there would be the advantage of being able to modify the prediction method and make more robust codes that can be better adjusted to the training data;
(5) This methodology may allow for the design of , taking advantage of the results of experimental measurements reported in various papers. This aspect is very useful for researchers who want to design and who do not have a laboratory, the equipment to develop experiments, or specialized software for design and simulation;
(6) So far, the code has been tested with the thermoelectric materials , , and . Currently, there are new materials, for example, organic materials or new alloys; thus, tests must be carried out using the properties of these new materials, to adapt the code to the needs of new applications.
(7) In addition to what is mentioned in point , the designed code does not require high-end computing equipment compared to specialized design software. In the case of the calculations developed in this work, an AMD Ryzen 3 processor with Radeon Vega Mobile Gfx 2.60 GHz, with 8 GB RAM, was used;
(8) The methodology has only been applied for constant cross-sectional areas with a quadrangular geometry. The code should be extended to include the design of with cross-sectional areas with geometries other than quadrangular, for example, circular. Variable cross-sectional areas regarding leg length could also be included. This could be achieved by reformulating Equations (16) and (17) in terms of and ;
(9) The code does not send any warning in the event that the user makes some type of error when capturing the information for the design; this still depends on the skills and knowledge of the user regarding the specifications of the system to be designed, but it is intended that soon some kind of table with reference values will be added to act as a design guide;
(10) It would be very useful to link this code with an interface that collects data in real time from experiments carried out in various laboratories around the world. This would allow various thermoelectric material research groups to make a quick evaluation of the possibility of developing new devices.
Artificial Intelligence Applied in Thermoelectrics
Supervised machine learning (which is an area of AI) has been used to predict values of thermoelectric properties and the electrical power produced; the input values are the operating temperature and the space between the legs. The application of this powerful computational tool is novel in the field of thermoelectric devices, as seen in [
18]. Artificial neural networks were applied to model a thermoelectric generator’s maximum energy generation and efficiency. The authors concluded that neural networks demonstrated an extremely high prediction accuracy, greater than 98%, and they can operate under a constant temperature difference and heat flux. The physical model considered contact resistance, electrical, surface heat transfer, and other thermoelectric effects.
Furthermore, Chika Maduabuchi’s paper [
19] presented the first AI-enabled optimization of a TEG performed using deep neural networks (DNN). The effects of strategic parameters on TEG power output, efficiency, and thermal stress performance were investigated. The parameters were hot and cold junction temperatures, heat transfer coefficients, incident heat flux, external load resistance, span height TE, area, and area ratio.