See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/366720268

Critical heat flux prediction for safety analysis of nuclear reactors using machine
learning

Conference Paper · August 2022

DOI: 10.1109/IBCAST54850.2022.9990190

CITATIONS READS
9 91

4 authors:

Rehan Zubair Khalid Atta Ullah

Pakistan Institute of Engineering and Applied Sciences Pakistan Institute of Engineering and Applied Sciences
8 PUBLICATIONS 29 CITATIONS 88 PUBLICATIONS 523 CITATIONS

SEE PROFILE SEE PROFILE

Asifullah Khan Mansoor Inayat

Pakistan Institute of Engineering and Applied Sciences National Centre for Physics
290 PUBLICATIONS 9,992 CITATIONS 33 PUBLICATIONS 560 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Rehan Zubair Khalid on 18 April 2024.

The user has requested enhancement of the downloaded file.

 Critical heat flux prediction for safety analysis of
nuclear reactors using machine learning
2022 19th International Bhurban Conference on Applied Sciences and Technology (IBCAST) | 978-1-6654-6051-4/22/$31.00 ©2022 IEEE | DOI: 10.1109/IBCAST54850.2022.9990190

Rehan Zubair Atta Ullah Asifullah Khan Mansoor H.Inayat

Department of Chemical Department of Chemical Department of Computer and Department of Chemical
Engineering Engineering Information Sciences Engineering
Pakistan Institute of Engineering Pakistan Institute of Engineering Pakistan Institute of Engineering Pakistan Institute of Engineering
and Applied Sciences (PIEAS) and Applied Sciences (PIEAS) and Applied Sciences (PIEAS) and Applied Sciences (PIEAS)
Islamabad, Pakistan Islamabad, Pakistan Islamabad, Pakistan Islamabad, Pakistan
[email protected] [email protected] [email protected] [email protected]

MCHFR value for a commercial PWR has to be at least 1.3 at

full power transient as a margin of safety [3]. As a result,
Abstract— Critical Heat Flux (CHF) prediction is vital for predicting CHF under operating circumstances is critical for
pressurized water reactors because it is a crucial metric for nuclear reactor safety and efficiency. The relevance of CHF
nuclear power plants' economic effectiveness and reliability. It is has prompted a lot of experimental and theoretical research.
also essential for the safety analysis of nuclear reactors as it is a However, because there is no consensus on the mechanism that
limiting factor for the design and operation of nuclear power causes CHF, the route to predicting exact CHF remains elusive
plants. Due to absence of no deterministic theory exists on its [4]. This lack of consensus stems from a lack of knowledge of
predictions, the method of predicting accurate CHF is not the highly complicated physics of multi-phase flow heat
straight forward. This paper develops an intelligent tool for CHF transfer phenomena, which was because of a lack of
prediction that covers a wide range of pressure, mass flowrate, experimental capacity to take measurements at the spatial and
and thermal flux. Two machine learning (ML) techniques, temporal scales characteristic of CHF phenomena [5]. Despite
namely random forest and artificial neural networks, are used to this, researchers have developed a wide range of CHF
predict CHF. A new dataset is developed from literature, prediction tools, including data-driven methods, i.e., look-up
covering a wide range of two-phase flow operating conditions to tables (LUTs) and empirical correlations to physics-based
train and test the model. Then both ML techniques are compared analytical models. Form one side, data-driven best-fitting
with the widely used look-up table method for predicting CHF. strategies, including LUTs and empirical correlations are
The results show the superior predicting capabilities of both ML
simple to utilize. These methods show good agreement for a
techniques, especially deep neural networks (DNN), over the
specific dataset range but not beyond their scope of validity of
conventional method. Parametric trends are also compared
among these techniques. Furthermore, it can also be used as an
operating ranges. On the other hand, analytical models depend
online monitoring tool for reactor cores of nuclear power plants on assumptions relying on a decent but limited grasp of the
to predict extreme conditions. underlying mechanisms, which are then reinforced with
primarily empirical constitutive connections to satisfy the
Keywords— Critical heat flux, Flow boiling, Prediction, Look- conservation equations.
up table, Artificial Neural Network, Random Forest
Due to advancements in computing capacity and optimization
I. INTRODUCTION methodologies, predictive approaches based on artificial
intelligence have emerged as a possible alternative to
CHF is used in the thermal-hydraulic analysis of various conventional data-based tools. The machine learning (ML)
process equipment, such as heat exchangers, reactor cores, and method is particularly beneficial in engineering domains where
fuel cells. It is also critical for nuclear reactor safety design. complicated physical processes are hard to describe. Because
CHF, also known as a departure from nucleate boiling, is of its improved performance when interacting with severely
described as a rapid increase in heating surface temperature non-linear relations, an artificial neural network (ANN) is
caused by the creation of a gas film between the heating among the utmost preferred options in this area [6, 7]. ANN
surface and the bulk fluid, resulting in a loss in heat transfer consists of multiple layers capable of non-linear mapping
and eventual mechanical failure of the heating device [1]. Due through transfer function. It offers a variety of benefits over
to its complexity, CHF is a significant for the thermal- traditional methods, including non-linear mapping capability,
hydraulics variable analysis and one of the least understood accurate prediction, easiness of training and rapid updating,
processes [2]. The CHF imposes a regulatory restriction for and a good generalization ability [8]. Tree-based ensemble
industrial PWRs during licensing worldwide. The CHF ratio method, namely random forest (RF), is another prominent
(CHFR) is the smallest ratio that allows reactors to run safely supervised ML method for regression. It is highly efficient ML
between the anticipated CHF and the actual operating heat algorithm that requires minimum hyperparameter tunning.
flux. The CHFR varies across the fuel length and depends upon
reactor design. The minimum value of CHFR (termed
MCHFR) should be more than one for every point in the core.

978-1-6654-6051-4/22/$31.00 ©2022 IEEE

Proceedings of 2022 19th International Bhurban Conference on Applied Sciences & Technology (IBCAST) 314
Authorized licensed use limited to: PAKISTAN INST OF ENGINEERING AND APPLIED SCIENCES. Downloaded on September 21,2023 at 04:53:59 UTC from IEEE Xplore. Restrictions apply.
Islamabad, Pakistan, 16th – 20th August, 2022
 TABLE I. CHF DATASET

Author Mass flow rate Pressure Exit quality Heated Length Heated Diameter CHF # of
[Kg/m2s] [MPa] [-] [mm] [mm] [MW/m2] samples
Williams [9] 325-4683 2.75-15.2 -0.03 to 0.92 1800 9.5 0.38-4.1 129
Kim [10] 20-277 0.11-0.95 0.32 to 1.2 300-1700 6-12 0.12-1.6 512
Stein [11] 24-304 1.1-7.1 -0.002 to 1 120-450 9 0.24-4.7 383
Becker [12] 100-5450 0.21-9.9 0 to 0.99 400-3750 3.9-24.9 0.27-7.5 3473
Lowdermilk [13] 59.6-596.7 3.4 0.71 to 0.94 152.4 3 0.47-3.3 21
Clark [14] 28-102 3.4-13.8 0.66 to 0.99 238 4.58 0.23-9.97 67
Reynold [15] 1166-2888 3.5-10.6 0 to 0.47 228 4.6 3.6-9 67
Inasaka [16] 4300-12300 0.29-0.99 -0.17 to -0.05 100 3 7.3-10.1 6
Total 20-12300 0.11-15.2 -0.17 to 1.25 100-3750 3-24.9 0.12-10.1 4658

II. DATASET DESCRIPTION B. Artificial neural network

ANN is a machine learning technique biologically inspired
In this work, a wide range of nonproprietary CHF
by the human brain, consisting of many discrete processing
experimental data available in the literature is evaluated, and
components known as neurons. Among the different varieties
then a data set is created. The dataset consists of 4658 samples,
of ANNs accessible today, in engineering, the feed-forward
covering a wide range of operating conditions. The dataset
neural network has become the most popular. [19].
includes both geometrical and hydraulic parameters for vertical
up-flow water in a tube that has been uniformly heated. The The input layer, one or more hidden layers, and the output
input variables from raw data include mass flux (G), pressure layer are the three layers that make up a neural network in
(P), local equilibrium quality (x), heated tube length (L), and general. A layer that receives information from input
tube equivalent diameter (D). The output/desired target parameters and transmits it into a hidden layer is called the
variable is CHF, and the heat flux distribution profile input layer. The output layer is a layer that receives
throughout the axis is uniform. Table I shows the experimental information from the hidden layer and produces the final
ranges of the CHF dataset. To increase the prediction accuracy, output. A deep neural network consisting of many hidden
standardization (i.e., zero arithmetic mean and unity standard layers, usually with a non-linear activation function, can learn
deviation) is applied to input variables. both linear and non-linear relationships between input and
output [20]. These weights and biases are changed during the
III. METHODOLOGY training phase to enhance the network accuracy, also known as
A. Look-up table learning. Because the weights and biases are set at random,
they must be changed iteratively to minimize the difference
CHF look-up tables (LUT) are frequently used in the between the expected and desired outputs, which is why the
nuclear reactor thermal-hydraulics field to predict CHF. It is a backpropagation learning technique is so popular. Architecture
more accurate and widely used data-driven approach for of ANN is shown in Fig. 1.
forecasting CHF for nuclear reactor safety assessments. The
CHF LUT is a normalized data bank for an 8 mm water-cooled C. Random forest
vertical round tube. It contains more than 30,000 data points Random forest (RF) is a supervised machine learning
covering a wide range of mass fluxes (0 to 8000 kg/m2s), technique used to solve both classification and regression
pressures (0 to 21 MPa), and exit/local qualities (-0.50 to 0.90). problems. It employs ensemble learning, a method that
The CHF LUT approach has many advantages, such as a full combines many classifiers to solve complex problems. A tree-
range of practical applications, ease of use, and iteration not based ensemble learning approach offers reliable results with
being required for predicting CHF [17]. little hyperparameter adjustment [21]. RF uses a number of
To tackle diameter correction of tube, Groeneveld et al decision trees. Their results are then aggregated using the
(1996) [18] has proposed following general equation: bagging technique (prediction averaging strategy) to increase
accurate prediction, avoid overfitting, and reduce variance), as
shown in Fig. 2.
𝐶𝐻𝐹𝑑 8 0.5
= ( )
𝐶𝐻𝐹𝑑=8 𝑑

Proceedings of 2022 19th International Bhurban Conference on Applied Sciences & Technology (IBCAST) 315
Authorized licensed use limited to: PAKISTAN INST OF ENGINEERING AND APPLIED SCIENCES. Downloaded on September 21,2023 at 04:53:59 UTC from IEEE Xplore. Restrictions apply.
Islamabad, Pakistan, 16th – 20th August, 2022
 Fig.1. ANN Architecture Fig. 2. Typical RF Architecture

approximator since it can estimate any function using a

IV. TRAINING AND TESTING combination of ReLU. To further validate the results of the
Both RF and ANN are used to train and test the data. For proposed model, 10-fold cross-validation is also performed.
this purpose, the deep learning library, i.e., Keras application
interface programming backend by tensor flow, is used with Similarly, RF hyperparameters are explored to attain the
the sci-kit-learn library in python 3.9. During this process, 80% best performance. The best RF model has 100 n-
of data is used for training, and 20% is used to evaluate the estimators/number of trees.
performance. The data are randomly distributed to check how V. RESULTS AND DISCUSSION
well performed on unseen data. The structure of an ANN
specifies the appropriate number of hidden layers, neurons in In this paper, three different approaches are used to predict the
the hidden layers, the activation function employed in the CHF for nuclear reactor safety analysis. One is the most
layers, and the connection of layers. The architecture of a deep popular approach (LUT approach) among the reactor thermal-
neural network has a major impact on the network's hydraulic community for the prediction of CHF during the
generalization and how accurately predicate the output for a design and operation of the reactors. The other two are
given input; therefore, it should be optimized. Physical artificial intelligence-based techniques, including ANN and
parameters dictate the neurons in the input and output layers. RF. The performance of three different models (as depicted in
Therefore, hidden layers, neurons in each layer, and activation Table III) is compared through relative root-mean-square error
function in each layer are important for an ANN's architecture. (rRMSE), which is defined as
The following architecture and hyperparameters, as
depicted in Table II, are used to achieve the best accuracy 𝑚
during the training process. 𝑦𝑖 − 𝑦𝑝𝑖 2
𝑟𝑅𝑀𝑆𝐸 = √1/𝑚 ∑ ( )
TABLE II. DETAILS OF BEST ANN MODEL 𝑦𝑖
𝑖=1

Best ANN model Hyperparameters Where m is the number of samples in the dataset.
• network architecture 5/100/100/100/1
• activation function (hidden layers) Rectified Linear Unit (ReLU) TABLE III. PERFORMANCE OF LUT, RF AND ANN
• optimization algorithm Adam (optimal learning rate =
0.001) Technique Test Samples within Samples within
rRMSE ± 10% error ± 20% error
LUT 17.5 60% 70%
It is worth mentioning that there is no universal rule for RF 12.3 90% 95%
determining the ideal topology of a deep neural network, and it ANN 9.5 94% 99%
is often established through the trial and error method [22]. ANN (10-fold 9.7 92% 98%
Optimal network architecture has three hidden layers of 100 cross-validation)
neurons in each. The input layer has five neurons
corresponding to the input variables, and the output layer has Fig. 3. represents the performance of three different
one neuron. The Adam optimization algorithm (optimal techniques in terms of root mean square relative error. By
learning rate = 0.001) is employed to update synaptic weights comparison, it is found that the proposed ANN model
and biases. Adam optimizer is famous for deep learning outperformed as compared to RF and LUT.
problems and requires minimum tuning. It is a stochastic The relative error of CHF of three different techniques are
gradient descent optimization algorithm derived from an also compared with pressure, mass flow rate, tube diameter,
adaptive estimation of 1st order and 2nd order moments. Non- length, and exit quality. The relative CHF error, which is
linearity is added in order to increase predictive capabilities. defined as
The Rectified Linear Unit (ReLU) transfer function is
CHF prediction −CHF experimental
employed in the hidden layers because of its popularity and CHF relative-error =
CHF experimental
performance in the proposed model. It is a universal

Proceedings of 2022 19th International Bhurban Conference on Applied Sciences & Technology (IBCAST) 316
Authorized licensed use limited to: PAKISTAN INST OF ENGINEERING AND APPLIED SCIENCES. Downloaded on September 21,2023 at 04:53:59 UTC from IEEE Xplore. Restrictions apply.
Islamabad, Pakistan, 16th – 20th August, 2022
 Fig. 4. shows ANN model also modifies mostly biased and across a broad spectrum of practical interest. As a result of
scattered parametric trend vs (mass flow-rate, exit quality, these findings, the proposed ANN approach's higher
pressure, heated tube diameter, heated length) with RF, LUT generalization capacity has been demonstrated.

Fig.3. Predicated Vs experimental CHF with LUT, RF and ANN.

Proceedings of 2022 19th International Bhurban Conference on Applied Sciences & Technology (IBCAST) 317
Authorized licensed use limited to: PAKISTAN INST OF ENGINEERING AND APPLIED SCIENCES. Downloaded on September 21,2023 at 04:53:59 UTC from IEEE Xplore. Restrictions apply.
Islamabad, Pakistan, 16th – 20th August, 2022
 [3] Todreas, N.E. and M.S. Kazimi, Nuclear systems volume I: Thermal
hydraulic fundamentals. 2021: CRC press.
[4] Celata, G.P., et al., Rationalization of existing mechanistic models for
the prediction of water subcooled flow boiling critical heat flux. 1994.
37: p. 347-360.
[5] Bucci, M., Advanced diagnostics to resolve long-lasting controversies in
boiling heat transfer. 2017.
[6] Leshno, M., et al., Multilayer feed-forward networks with a
nonpolynomial activation function can approximate any function. 1993.
6(6): p. 861-867.
[7] Khan, A., M.H. Shamsi, and T.-S.J.C.M.S. Choi, Correlating dynamical
mechanical properties with temperature and clay composition of
polymer-clay nanocomposites. 2009. 45(2): p. 257-265.
[8] Abiodun, O.I., et al., State-of-the-art in artificial neural network
applications: A survey. 2018. 4(11): p. e00938.
[9] Williams, C. and S. Beus, Critical heat flux experiments in a circular
Fig. 4. CHF relative error vs pressure, mass flow rate, tube diameter, length, tube with heavy water and light water.(AWBA Development Program).
and exit quality 1980, Bettis Atomic Power Lab., West Mifflin, PA (USA).
[10] Kim, H.C., et al., Critical heat flux of water in vertical round tubes at
VI. CONCLUSION low pressure and low flow conditions. 2000. 199(1-2): p. 49-73.
In this work, the predictive capacity of artificial intelligence [11] Stein, M., Systematische Untersuchung der kritischen
(AI) techniques is employed to enhance the safety and Wärmestromdichte beim Strömungssieden von Wasser in lotrechten
Kreisrohren mit und ohne poröser Beschichtung. 2004.
economy of the nuclear reactor system, which mainly depends
[12] Becker, K.M., et al., Burnout data for flow of boiling water in vertical
on the accurate prediction of CHF. A comprehensive round ducts, annuli and rod clusters. 1965, AB Atomenergi.
assessment has shown superior predictive capabilities of [13] Lowdermilk, W.H. and W.F. Weiland, Some measurements of boiling
artificial intelligence approaches as compared to the widely burn-out. 1955.
used conventional data-driven approach, i.e., LUT. [14] Clark, J.A. and W.M. Rohsenow, Technical report no. 4 on local boiling
Furthermore, deep learning approach is outperformed as heat transfer to water at low reynolds numbers and high pressures. 1952,
compared to other machine learning algorithms i.e., RF. One of Massachusetts Inst. of Tech. Div. of Industrial Cooperation.
the most critical characteristics of AI methods is their capacity [15] Reynolds, J.M., Burnout in forced convection nucleate boiling of water.
to expand their applicability domain on-the-fly. 1957, Massachusetts Institute of Technology.
[16] Inasaka, F., H.J.F.e. Nariai, and design, Critical heat flux of subcooled
The finding of this research provides deep insight into the flow boiling for water in uniformly heated straight tubes. 1992. 19(4): p.
future development and enhancement of AI techniques, 329-337.
particularly deep learning models for CHF prediction. More [17] Groeneveld, D., et al., The 2006 CHF look-up table. 2007. 237(15-17):
experimental data is also vital for CHF prediction through AI p. 1909-1922.
approach. Other channel geometries will certainly increase the [18] Tanase, A., et al., Diameter effect on critical heat flux. 2009. 239(2): p.
number of input characteristics. For instance, to estimate CHF, 289-294.
if rod bundle data is given, bundle-specific factors such as non- [19] Rodriguez-Galiano, V., et al., Machine learning predictive models for
mineral prospectivity: An evaluation of neural networks, random forest,
uniform heat flux distribution, grid space layout, and local flow regression trees and support vector machines. 2015. 71: p. 804-818.
conditions may also be taken into account.
[20] Kalogirou, S.A.J.A.e., Applications of artificial neural-networks for
energy systems. 2000. 67(1-2): p. 17-35.
[21] Mohanraj, M., et al., Applications of artificial neural networks for
REFERENCES refrigeration, air-conditioning and heat pump systems—A review. 2012.
16(2): p. 1340-1358.
[22] Gupta, T.K. and K.J.N.P.L. Raza, Optimizing deep feed-forward neural
[1] Ghiaasiaan, S.M., Two-phase flow, boiling, and condensation: in network architecture: A tabu search based approach. 2020. 51(3): p.
conventional and miniature systems. 2007: Cambridge University Press. 2855-2870.
[2] Baglietto, E., E. Demarly, and R.J.A.P.L. Kommajosyula, Boiling crisis
as the stability limit to wall heat partitioning. 2019. 114(10): p. 103701.

Proceedings of 2022 19th International Bhurban Conference on Applied Sciences & Technology (IBCAST) 318
Authorized licensed use limited to: PAKISTAN INST OF ENGINEERING AND APPLIED SCIENCES. Downloaded on September 21,2023 at 04:53:59 UTC from IEEE Xplore. Restrictions apply.
Islamabad, Pakistan, 16th – 20th August, 2022
View publication stats