Nuclear Engineering and Design 395 (2022) 111862

Contents lists available at ScienceDirect

Nuclear Engineering and Design

journal homepage: www.elsevier.com/locate/nucengdes

Artificial intelligence-driven thermal design for additively manufactured

reactor cores☆
Emilian Popov a, *, Richard Archibald a, Briana Hiscox a, Vladimir Sobes b
a
Oak Ridge National Laboratory, 1 Bethel Valley rd., Oak Ridge, TN 3780, United States
b
Department of Nuclear Engineering, University of Tennessee, 1502 Cumberland Ave., Knoxville, TN 37996, United States

A R T I C L E I N F O A B S T R A C T

Keywords: The paper describes application of machine learning (ML), specifically deep learning and Gaussian processes, to
Artificial intelligence (AI) optimize the design of complex nuclear engineering systems for which predictive, but computationally intensive,
Machine learning (ML) multiphysics solutions are available. The approach combines reduced-order modeling, simulation, and ML for
Computational fluid dynamics (CFD)
computational design. High-fidelity reactor thermal hydraulics (TH) simulations are utilized effectively, mini­
Thermal hydraulics (TH)
Reactor design
mizing the computational and physical costs of large-scale simulations and expensive prototyping, by using
Optimization reduced-order models to explore the design space. Models are bias corrected by ML error correction using in­
formation from the high-fidelity simulations. This method draws on the uncertainty quantification intrinsic to
Gaussian processes to provide the best predictions of optimal designs with error bounds. Based on the multi-
objective loss function, the method proposes a new set of simulations that maximizes reduction of uncertainty
in the regions identified as the most promising to minimize loss.
This work explores the automatic construction of physics-informed ML methods using emulators validated by
very sparse sampling of coupled thermal diffusion and viscous, turbulent flow solutions. The emulators become
more accurate as new data are generated by steady-state reduced-order modeling, in which an individual design
can be executed on a single graphics processing unit (GPU) in a few minutes. The parallel nature of probing the
design space results in almost half a million different reactor designs that can be evaluated in an hour on Summit,
an Oak Ridge Leadership Computing Facility (OLCF) supercomputer.

1. Introduction the application of AM in the nuclear industry (Betzler, et al., 2019),

leveraging advanced materials, data science, and rapid testing and
Continued developments in advanced manufacturing technologies deployment to drive down costs and development times, ultimately
are fundamentally altering the way components are designed and improving future commercial viability. This approach is being demon­
fabricated (Thompson et al., 2016). The potential application space for strated under the US Department of Energy Office of Nuclear Energy
these technologies within the nuclear industry is very broad because of (DOE-NE) Transformational Challenge Reactor (TCR) program (Oak
the rigorous requirements and inherent multidisciplinary nature­ Ridge National Laboratory, 2019).
—encompassing civil, mechanical, electrical, and nuclear engineer­ Previous efforts in nuclear system optimization formulated the
ing—of large nuclear power plants (Tobin et al., 2018). Beyond problem as a combination with fixed geometry instead of an optimiza­
applications for existing technologies, these new manufacturing tion over continuously variable geometry parameterization (Pereira and
methods, advanced materials, and dimensional constraints can be Lapa, 2003; Pereira et al., 1999; Jayalal et al., 2015; Pazirandeh and
applied to the nuclear core design problem (Sobes, et al., 2020). A Tayefi, 2012; Gomez-Fernandez et al., 2020; Liu et al., 2019; Betzler
manufacturing-informed design approach yields the most benefit from et al., 2019). Genetic algorithms have been the predominant choice of AI

☆
This manuscript has been authored by UT-Battelle, LLC, under contract DE-AC05-00OR22725 with the US Department of Energy (DOE). The US government
retains and the publisher, by accepting the article for publication, acknowledges that the US government retains a nonexclusive, paid-up, irrevocable, worldwide
license to publish or reproduce the published form of this manuscript, or allow others to do so, for US government purposes. DOE will provide public access to these
results of federally sponsored research in accordance with the DOE Public Access Plan (https://energy.gov/downloads/doe-public-access-plan).
* Corresponding author.
E-mail addresses: [email protected] (E. Popov), [email protected] (R. Archibald), [email protected] (B. Hiscox), [email protected] (V. Sobes).

https://doi.org/10.1016/j.nucengdes.2022.111862
Received 30 November 2021; Received in revised form 25 April 2022; Accepted 16 June 2022
Available online 27 June 2022
0029-5493/© 2022 Elsevier B.V. All rights reserved.
E. Popov et al. Nuclear Engineering and Design 395 (2022) 111862

Fig. 1. Optimization problem geometry, including 1/12 of the real reactor core (left). Colors indicate different components: yellow represents the moderators, blue
among hexagons is the coolant, and cyan is the fuel. The red hexagons show the overlapped original (baseline) core geometry. Right, an individual computational cell
showing the optimization parameters. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 2. TH simulator results showing the temperature in the core components at three axial elevations (middle) for a randomly selected optimization set. The insert
on the left shows the input power as mapped to the grid, and the insert on right shows the surface grid.

2020). However, arbitrary geometry could only be considered through a

Table 1
voxel representation. Lastly, the authors published a study on the
Baseline design compliance comparison with data from (Jain et al., 2022).
framework for an arbitrary geometry optimization of nuclear systems
Parameter Relative difference, % and gave some demonstrations on simplified challenge problems (Sobes,
Core outlet temperature 0.4 et al., 2020). This work is the application of that framework to a nuclear
Maximum fuel temperature 2.5 reactor full core.
Minimum fuel temperature 0.3
This work focuses on solutions of the Navier-Stokes equations, which
Maximum channel velocity 2.0
Average channel velocity 0.5
are tightly coupled to the heat conduction equation. The solution to this
Maximum moderator temperature 1.8 set of coupled equations in nuclear engineering is tied up in several
Minimum moderator temperature 0.7 complex computer codes and is computationally intensive. Therefore, an
efficient multiparameter optimization search using high-fidelity physics
models is prohibitively expensive. In this work, the authors explore the
for the combinatorial optimization problem. When more continuously
automatic construction of physics-informed ML methods by using em­
variable geometries were studied, more structured optimization ap­
ulators with validation from very sparse sampling of the predictive high-
proaches were used (Pevey et al., 2020). A recent work considered the
fidelity physics simulations. The physics-informed emulators are based
nuclear systems design by using genetic algorithms created from scratch
on a steady-state reduced-order model through Gaussian kernel convo­
rather than optimizing an existing configuration (Archibald et al.,
lution that allows for a fast evaluation on a single GPU (Archibald et al.,

2
E. Popov et al. Nuclear Engineering and Design 395 (2022) 111862

Fig. 3. Displays of (a) initial design based upon 4 parameter optimizations with loss 14.7 ◦ C, (b) best design after first round of optimization with loss 13.9 ◦ C, (c)
best design after second round of optimization with loss 13.4 ◦ C, (d) final round of optimization with loss 13.3 ◦ C.

Fig. 4. Initial (left) and final (right) core component configuration showing increased amount of fuel and cooling, and decreased amount of moderator. Fuel
indicated with yellow, and moderator with cyan. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of
this article.)

2020), which is suitable for ideal scalability in the search over the vast 2. Problem statement
design space. The surrogate modeling method that was developed in this
work falls under the category of multi-fidelity surrogate modeling. The The problem represents 1/12 symmetry of the entire reactor core
information generated by the dense sampling of reduced-order modeling (Fig. 1). The real core configuration was idealized (modified and
is merged with sparse full physics simulation via Gaussian processes simplified) to allow for efficient parametrization. The baseline design
(GPs) (Williams and Rasmussen, 2006). GPs have recently been used in can be parametrized with four parameters (two for the moderator radii,
various studies for computational simulation and analysis of nuclear P1 and P2, and two for the fuel radii, P3 and P4). In the space between
reactors (West et al., 2021; Jiang et al., 2020; Figueroa and Göttsche, fuel and moderator, the coolant channels form as those four parameters
2021; Majdi, 2020). This work uniquely developed reduced-order vary. The fuel element baseline design is indicated in Fig. 1 by the red
models of the neutronics and thermofluidic physics and were designed hexagons, which provides a conceptual idea of the modeling approach.
for a fast evaluation on a single GPU. Thus, it is possible to evaluate Further details are restricted by export control and proprietary
millions of designs by using GPU-based high-performance computing regulations.
(HPC) systems, such as Summit (Oak Ridge Leadership Computing Fa­ This paper presents results from the initial four parameter optimi­
cility, 2022). zation problem. The parameters are allowed to vary as to always pro­
duce criticality within a desired band, with constraints for minimum

3
E. Popov et al. Nuclear Engineering and Design 395 (2022) 111862

Fig. 5. Local loss function prediction, in the initial (left) and final (right) distributions as applied over the core fuel elements. The position of the critical element is
clearly identifiable in the lower right corner, and the local loss function reduces from 21.3 down to 19.9 ◦ C.

parameters with the baseline design. Because of this paper’s limited

Table 2
scope, this verification is not discussed in detail.
Correlation coefficients between input and output quantities. The highest co­
The optimization objective is to minimize the thermal loads on core
efficients are highlighted.
components. Three criteria were initially selected and computed for
P1 P2 P3 P4
each individual element: temperature, temperature difference, and
Keff − 0.7 0.7 0.4 0.0 temperature standard deviation. These criteria were computed on the
Max T − 0.3 0.3 0.3 − 0.7 baseline core configuration, as shown by the red hexagons. As part of
DT − 0.3 0.3 0.3 − 0.6
DP 0.3 − 0.3 − 0.6 0.7
another structural mechanic’s study, the same criteria were used as
Power − 0.3 0.3 0.4 0.4 forcing functions to determine which criterion leads to the largest
Std T − 0.5 0.3 0.4 − 0.2 component loading. This work has not been finalized yet and the ana­
Std T cog − 0.3 0.2 0.2 − 0.4 lyses have been carried out for the three objective functions above.

cooling channel clearance. The four-parameter problem does not allow 3. Simulator and emulator model overview
for radial or axial grading of core components, and it is thus limited. This
initial step is intended to confirm the original design (nominal param­ The neutronics code Monte Carlo N-Particle (MCNP) was chosen for
eter values) and explore this limited optimization space. The adequacy this analysis due to its ability to calculate power in non-fuel materials
of model performance was confirmed by comparing key design (MCNP, 2005). Four mesh tallies were used to calculate the total energy
deposition at each spatial location. The geometry was broken down into

Fig. 6. Plots of Keff versus moderator volume (left), and pressure drop versus coolant channel size (right) in relative units, for the full set of core averaged data points.

4
E. Popov et al. Nuclear Engineering and Design 395 (2022) 111862

small cells and the energy deposition due to neutrons and photons was 3.2. ML with uncertainty
calculated for each cell. Analysis was done to ensure the mesh was fine
enough to provide good resolution while having a reasonable compu­ Kernel-based ML provides an efficient method for ML applicable to
tational cost. This multi-physics problem has a weak dependence on physics-oriented problems in engineering sciences. Specifically, given a
thermal feedback due to the gas used as coolant and solid moderators. It set.
is, therefore, sufficient to pass the spatial power deposition calculated by {( ) }
pi , Pi (x), Ti (x), Vi,c (x), Vi,f (x), Vi,m (x), Li , σ i : i = 1, ..., N (1)
MCNP directly to the TH model. The high fidelity neutronic calculation
is performed at the appropriate thermal conditions that provide accurate
of training data the loss can be determined for any parameter set. Here, x
properties. Additional iterations between the neutronics solver and the
is the position vector, and pi is the parameter vector for the ith design.
TH model are unnecessary.
The functions are the power Pi , component temperature Ti , fractional
The optimization was performed using the information obtained
coolant indicator Vi,c , fractional fuel indicator Vi,f , and fractional
from a combination of TH reactor core simulations and low-order TH
moderator indicator Vi,m for the N simulated training sets. The loss Li for
physics-based emulators. The low-order information was used to
any design is defined to be the standard deviation of temperature for
augment the data for training the Gaussian process ML for optimized
every domain with positive fuel indicator. The last item in this collection
reactor design. A brief description of both simulator and emulator is
is the error estimation of the loss σi . The loss of any design p is predicted
given in this section. A more detailed description can be obtained from
by using the kernel-based ML method defined as:
Sobes et al. (2021) together with the computer code used to perform the
calculations. ∑N
L(p) = i=1
ci k(p, pi ) (2)

3.1. TH reactor core simulator 1 2

where kernel function, k(p, pi ) = e− 2‖p− pi ‖ was used. The coefficients of
The TH simulator is based on a solution of the Reynolds-averaged the kernel-based ML are found by solving.
Navier-Stokes (RANS) equations for the coolant flow in the reactor ∑N ( )
Lj = ci k pj , pi ≡ Kc (3)
core channels, combined with conjugated thermal conduction equation i=1

for the solid structures (fuel and moderators). The simulator solutions
for j = 1, ..., N training sets, where Lj losses are known. The matrix ele­
serve to train and validate the emulator models. The implementation ( )
uses the STAR-CCM + commercial multiphysics platform (Siemens, ments are given as Ki,j = k pi , pj for 1 ≤ i, j ≤ N and the coefficient
2020). vector c = (c1 , ..., cN ). One of the strengths of the Gaussian process is the
The RANS method used a two-equation approximation (realizable k- ability to deal with uncertainty in the data and to estimate uncertainty in
epsilon model) to solve for the turbulent kinetic energy and its dissi­ predictions. Specifically, the uncertainty in the loss of any design p is
pation rate. The spatial discretization was selected to minimize given by.
computational cost, which reduces this model to a high wye plus (Y + )
version. θ(p) = k(p, p) − kp T K − 1 kp (4)
The constantly varying geometry required regeneration of compu­
where the vector is defined as kp = (k(p, p1 ), ..., k(p, pN ) ).
tational grids at every optimization iteration. Grid factors were chosen
after a grid study (not discussed here). This was implemented in an
automated algorithm, which allowed each computing set of samples to 3.3. Data augmentation and optimization
run without user intervention. The prototype geometry was built and
meshed so that it would permit modifications according to the variation Simulation-based computational design can quickly become an
of the optimization parameters. intractable problem depending upon both the size of the computational
The boundary conditions included power from concurrent neu­ design space and the computational complexity of design simulation.
tronics calculations and a constant total coolant flow rate with a pre­ However, the kernel-based ML method is perfect for learning the opti­
defined inlet temperature. The thermal power was supplied as a mapped mization loss function for any given computational design if it has
to the computational grid function, and consistency checks were per­ enough design simulations to train. Unfortunately, the full-physics
formed to ensure that the total power is preserved as the grid varies. simulation is expensive, thus limiting the number of simulations that
Volumetric power densities were calculated for both fuel and moderator. can be run to produce a data training set. It is possible to augment the
The flow problem was simulated by assuming the flow is distributed training data with low-resolution simulations. The kernel-based method
among cooling channels based on the total pressure gradient across the above can make predictions using uncertain data, and it is therefore
core. Two plena, inlet and outlet were added to the model. A constant possible to use both physics simulation data and low-resolution data to
flow rate was applied to the inlet plenum, and an outlet pressure train the same machine. In this situation, the physics simulation data
boundary condition was applied to the outlet plenum. The pressure drop have no uncertainty, and the low-resolution simulations have uncer­
is a resultant quantity, which was calculated for each set of parameters. tainty proportional to the error between the low resolution and full fi­
Temperature fields were extracted from the solution and transferred to delity simulations. The simulations used in this study are relatively
the ML interface for training. inexpensive to run on a GPU. Thus, on a GPU-based HPC resource, many
A plot of temperature at three axial elevations in the entire domain low-resolution simulations can be generated to augment training data.
(fuel, moderator, and coolant) is shown in Fig. 2. The case was randomly
selected to illustrate the results of the TH simulator. The plot shows the 3.4. Emulator model
temperature distribution in core components (middle), the power input
(left), and a typical grid (right). The number of cases used for training The data from the full-fidelity physic model are augmented with a set
varied, and for the four-parameter problem it was around one hundred. of M emulated models:
This low number is an indication of the efficiency of the ML training {( ) }
algorithm and the proper selection of the emulator background physics ̃ i (x), Vi,c (x), Vi,f (x), Vi,m (x), Li , σ i : i = N + 1, ..., N + M
pi , Pi (x), T
models.
(5)
The ML method assumes that the full fidelity physical models are
exact, or σi = 0 for i = 1,...,N. In the case of i = N + 1,...,N + M, a low-

5
E. Popov et al. Nuclear Engineering and Design 395 (2022) 111862

resolution approximation of the physics is used to estimate the tem­ in the table, parameters P1 and P2, which define the moderation,
perature function T
̃ by solving: strongly affect the multiplication factor (Keff), highlighted in the table,
whereas parameter P4 (coolant channel) influences the temperature by
( )
− αc Vi,c (x) + αf Vi,f (x) + αm Vi,m (x) ΔT̃ i (x) = υi (x)
∂T̃ i (x)
+ Pi (x) (6) controlling the cooling. These dependencies are illustrated in Fig. 6 in
∂x which the data were core averaged for the final set of the optimization
cycle.
̃
where υi (x) is the flow field, ∂T∂ix(x) is the temperature gradient along the
flow direction, and αc , αf , and αm are constants. For any design p, the 5. Conclusions
flow field was calculated based on the coolant flow rate. The flow field is
zero in the solid material of the reactor. The flow field and constants αc , This paper discusses the TH aspects of an optimization problem that
∑
αf , and αm are calculated such that N ‖Ti (x) − T ̃ i (x)‖ is minimized, effectively uses ML and AI for reactor core design. The work is an
i=1 2
where Ti(x) is the training set temperatures. example to demonstrate the approach, and the algorithms will continue
When the data are augmented, the Gaussian process is calculated by: to be permanently evaluated and modified to improve the result. The
entire methodology targets AM core components for which many de­
∑N+M
L(p) = ci k(p, pi ) (7) grees of geometry freedom exist, and for which the traditional human-
oriented design is inefficient. However, the presented case is an initial
i=1

1 2 step in the progression suite of problems, in which the least number of

where the kernel function, k(p, pi ) = e− 2‖p− pi ‖ is used. The coefficients of
optimization parameters were considered. The outcome of this four-
the kernel-based ML are found by solving.
parameter problem demonstrates that the baseline design is almost
∑N+M ( ( ) )
Lj = ci k pj , pi + σ 2i δi,j ≡ Kc (8) optimal and confirms the accuracy of the developed modeling tools.
i=1

for j = 1, ..., N +M and σ i = 0 for i ≤ N, where the matrix elements are Declaration of Competing Interest
( )
given as Ki,j = k pi , pj for 1 ≤ i, j ≤ N + M, the coefficient vector c =
The authors declare that they have no known competing financial
(c1 , ..., cN+M ), and σi is an estimate for the error in the emulation. The interests or personal relationships that could have appeared to influence
uncertainty in the loss of any design p is given by. the work reported in this paper.
( )∑N+M 2
θ(p) = k(p, p) − kp T K − 1 kp σ k(p, pi ) (9)
i=1 Acknowledgment

4. Results from the four parameter problem This study was supported by the Transformational Challenge Pro­
gram by the US Department of Energy, Office of Nuclear Energy.
The four-parameter problem is the first step in the optimization line
of problems in which the baseline design space was explored. The References
problem does not assume radial or axial variation of core components;
instead, it relies on the optimal combination of fixed fuel, moderator, Archibald, A., Sobes, V., Hiscox, B., Popov, E., et al., 2020. Physics Based Machine
Learning for HPC Computational Design. Conference on Data Analysis 2020, Santa
and cooling geometries. The baseline design was human optimized for Fe, New Mexico.
performance and serves as the initial starting point for this optimization Betzler, B.R., Chandler, D., Cook, D.H., Davidson, E.E., Ilas, G., 2019. Design
problem. The four parameters assume their nominal values as per the optimization methods for high-performance research reactor core design. Nucl. Eng.
Des. 352.
baseline design. This initial state was used to evaluate the accuracy of Betzler, B. R., B. J. Ade, et al. 2019. Advanced Manufacturing for Nuclear Core Design.
the simulation tools by comparing selected performance metrics. This Oak Ridge National Laboratory, ORNL/TM-2019/1258 (2019).
model verification step is not discussed in the paper, although, to Figueroa, A., Göttsche, M., 2021. Gaussian processes for surrogate modeling of
discharged fuel nuclide compositions. Ann. Nucl. Energy 156, 108085. https://doi.
illustrate the model compliance with the baseline design data, particu­ org/10.1016/j.anucene.2020.108085.
larly of the thermal–hydraulic core parameters, some key metrics are Gomez-Fernandez, M., Higley, K., Tokuhiro, A., Welter, K., Wong, W.K., Yang, H., 2020.
provided in Table 1. It should be kept in mind that the optimization Status of research and development of learning-based approaches in nuclear science
and engineering: a review. Nucl. Eng. Des., 110479 https://doi.org/10.1016/j.
model is not a full replica of the baseline design model, which includes
nucengdes.2019.110479.
only a fuel cell, and is very detail. Jain, P., Weinmeister, J., Ade, B., Jesse, C., 2022, CFD Modeling for the TCR preliminary
Then, the algorithm was executed, and an optimized set of four pa­ design, NURETH19, paper 35935, Brussels, Belgium, March 6-11.
rameters was obtained. The results from this step are discussed below. Jayalal, M.L., Ramachandran, S., Rathakrishnan, S., Satya Murty, S.A.V., Sai Baba, M.,
2015. Application of genetic algorithm methodologies in fuel bundle burnup
Fig. 3 shows the core component temperatures as they are reduced optimization of pressurized heavy water reactor. Nucl. Eng. Des. 281, 58–71.
during the optimization progression. After four cycles, a reduction of the https://doi.org/10.1016/S0306-4549(98)00036-X.
core wide loss function with about 10% was obtained. Jiang, B.T., Zhou, J., Huang, X.B., Wang, P.F., 2020. Prediction of critical heat flux using
Gaussian process regression and ant colony optimization. Ann. Nucl. Energy 149.
The initial and the final core component configurations are shown in https://doi.org/10.1016/j.anucene.2020.107765.
Fig. 4. The AI optimization suggests a larger amount of fuel, and a lesser Liu, Z., Wang, J., Tan, S., Qiao, S., Ding, H., 2019. Multi-objective optimal design of the
amount of moderation. This results in the cooling being slightly modi­ nuclear reactor pressurizer. Int. J. Adv. Nucl. Reactor Des. Technol. 1, 1–9. https://
doi.org/10.1016/j.jandt.2019.09.001.
fied, leading to a reduction of the core pressure drop. MCNP—AGeneralMonteCarloN-Particle Transport Code, Version5,
For this set of simulations, the loss function is the standard deviation LosAlamosNationalLaboratory, Vol.I,2–71 (2–80) (2005).
of the fuel element temperature. It is visualized in Fig. 5 after being Oak Ridge Leadership Computing Facility, “Summit: America’s Newest and Smartest
Supercomputer,” www.olcf.ornl.gov/Summit/.
overlayed on the real core elements. The loss function reduction can be Oak Ridge National Laboratory, “Transformational Challenge Reactor.” Oak Ridge
clearly seen in the optimized design. As mentioned above, the search for National Laboratory. Accessed Dec. 1. https://tcr.ornl.gov/. (2019).
the optimization function is still ongoing. The result is an outcome from Pazirandeh, A., Tayefi, S., 2012. Optimizing the Fuel Management in a VVER-1000
Reactor Using an Artificial Neural Network. Ann. Nucl. Energy 42, 112–118. https://
a single optimization function problem, but the methodology allows for
doi.org/10.1016/S0306-4549(98)00036-X.
multi-function optimization to be performed if it is found that more than Pereira, C.M.N.A., Lapa, C.M.F., 2003. Coarse-grained parallel genetic algorithm applied
one criterion is important for the core design. to a nuclear reactor core design optimization problem. Ann. Nucl. Energy 30 (5),
Over the course of the analyses, several reactor core operational 555–565.
Pereira, C.M.N.A., Schirru, R., Martinez, A.S., 1999. Basic investigations related to
indicators were calculated. These indicators depend on the optimization genetic algorithms in core designs. Ann. Nucl. Energy 26 (3), 173–193. https://doi.
parameters and the correlation coefficients are given in Table 2. As seen org/10.1016/S0306-4549(98)00036-X.

6
E. Popov et al. Nuclear Engineering and Design 395 (2022) 111862

Pevey, J., Chvala, O., Davis, S., Sobes, V., et al., 2020. Genetic algorithm design of a Thompson, M.K., Moroni, G., et al., 2016. Design for Additive Manufacturing: Trends,
coupled fast and thermal subcritical assembly. Nucl. Technol. 206 (4). Opportunities, Considerations, and Constraints. CIRP Ann. 65 (2), 737–760.
Radaideh, Majdi I., Kozlowski, Tomasz, 2020. Surrogate modeling of advanced computer Tobin, K. W., J. T. Busby, et al. “Enabling Accelerated Deployment of Nuclear Energy
simulations using deep Gaussian processes. Reliab. Eng. Syst. Saf. 195 https://doi. Systems” Workshop Report, Oak Ridge National Laboratory (United States), ORNL/SPR-
org/10.1016/j.ress.2019.106731. 2018/1025, (2018).
Siemens, “STAR-CCM+ simulation platform”, https://www.plm.automation.siemens.co West, A., Tsitsimpelis, I., Licata, M., et al., 2021. Use of Gaussian process regression for
m/global/en/products/simcenter/STAR-CCM.html, (2020). radiation mapping of a nuclear reactor with a mobile robot. Sci. Rep. 11, 13975.
Sobes, V., Hiscox, B., Popov, E., Archibald, R., et al., 2021. AI-based design of a nuclear https://doi.org/10.1038/s41598-021-93474-4.
reactor core. Nature Sci. Rep. https://doi.org/10.1038/s41598-021-98037-1. Williams, C.K.I., Rasmussen, C.E., 2006. Gaussian Processes for Machine Learning. no. 3.
Sobes, V., B. D. Hiscox, et al. “Artificial Intelligence Design of Nuclear Systems MIT Press, Cambridge, Massachusetts.
Empowered by Advanced Manufacturing.” PHYSOR 2020 – Transition to a Scalable
Nuclear Future, Cambridge, United Kingdom, Mar. 29 - Apr. 2 (2020).