See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/344614751

Hardware-In-The-Loop Tests of Complex Control Software for Rocket

Propulsion Systems

Conference Paper · October 2020

CITATIONS

4READS

164

4 authors:

Some of the authors of this publication are also working on these related projects:

NetSat View project

University Wuerzburg Experimental Satellite 4 (UWE-4) View project

Günther Waxenegger-Wilfing

German Aerospace Center (DLR)

28 PUBLICATIONS   83 CITATIONS   

SEE PROFILE

Kai Dresia

German Aerospace Center (DLR)

12 PUBLICATIONS   43 CITATIONS   

SEE PROFILE

M. Oschwald

German Aerospace Center (DLR)

410 PUBLICATIONS   2,853 CITATIONS   

SEE PROFILE

Klaus Schilling

University of Wuerzburg

370 PUBLICATIONS   2,368 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Günther Waxenegger-Wilfing on 12 October 2020.

The user has requested enhancement of the downloaded file.

71st International Astronautical Congress, Virtual, 12-14 October 2020. Copyright c© 2020 by International AstronauticalFederation (IAF). All rights reserved.

IAC–2020–C4.1.15

Hardware-In-The-Loop Tests of Complex Control Software for RocketPropulsion Systems

Gunther Waxenegger-WilfingDLR (German Aerospace Center), Germany, guenther.waxenegger@dlr.de

Kai DresiaDLR (German Aerospace Center), Germany, kai.dresia@dlr.de

Michael OschwaldDLR (German Aerospace Center), RWTH Aachen University, Germany, michael.oschwald@dlr.de

Klaus SchillingUniversity Wuerzburg, Germany, schi@informatik.uni-wuerzburg.de

Software of rocket propulsion systems has a very high complexity but must also guarantee an outstandinglevel of reliability. Therefore, extensive tests are inevitable during the development process to meet therequirements. However, ground tests of the entire propulsion system are extremely expensive and can notaddress all nominal and off-nominal behavior that could possibly occur during the mission. Hardware-in-the-loop simulations allow realistic testing of embedded systems (e.g., the engine controller or an actuator)by combining a software simulation with emulated electrical sensor and actuator signals. This contributiondiscusses hardware-in-the-loop methods for anomaly detection and control systems of rocket propulsionsystems. We develop a roadmap surveying a step-by-step approach to demonstrate the reliability of advancedcontrol algorithms and sketch how to deploy such algorithms on space-graded embedded systems. In thefuture, adaptive and reconfigurable control for the propulsion system will become increasingly important.Adaptive control can tune the control actions to uncertain or changing system characteristics. Also, couplingthe control algorithms to suitable anomaly detection systems allows the safe reaction to unforeseen events,which further improves the reliability of the propulsion system. The DLR Institute of Space Propulsionoperates extensive test facilities for developing propulsion system technologies to operational maturity andensuring their quality. Specific attitude pointing aspects of the propulsion system software can be simulatedby high precision turntables at the company Zentrum fur Telematik (ZfT). Within the cooperation betweenthe DLR Institute and ZfT, different aspects of complex control software for propulsion systems are analyzedand tested using the available facilities.keywords: hardware-in-the-loop simulation, rocket engine control, control system development, rocketpropulsion test facilities

1. Introduction

The demands placed on the software are constantlyincreasing in the area of space transportation. Thereasons for this are that mission profiles are becomingmore complex, e.g. landing of first stages for subse-quent reuse, and that control aspects are gaining inimportance.4

Liquid rocket engines like the Space Shuttle mainengine use closed-loop control at most near steadyoperating conditions. The control of the transientphases is traditionally performed in open-loop due tohighly nonlinear system dynamics. The situation isunsatisfactory, in particular for reusable engines. Theopen-loop control system cannot react to external dis-

turbances. It is therefore intended to extend the useof closed-loop control to the transient phases. Onlyoptimal control can guarantee a long life expectancyof the engine without damaging pressure and tem-perature spikes. Although the importance of a suit-able closed-loop control system has been evident formany years, the majority of rocket engines still em-ploy valves which are operated with pneumatic ac-tuators, too inefficient for sophisticated closed-loopcontrol. The development of an all-electric controlsystem3 started in the late 90s in Europe. The fu-ture European Prometheus engine will have such asystem.16 Other countries are also well advanced inthe research and development of electrically operated

IAC–2020–C4.1.15 Page 1 of 8

71st International Astronautical Congress, Virtual, 12-14 October 2020. Copyright c© 2020 by International AstronauticalFederation (IAF). All rights reserved.

flow control valves.1 Due to the electrification of theactuators and the grown demands, the interest in op-timal engine control increased recently and will fur-ther rise in the future.

Since rocket engines operate at the limits of whatis technically feasible, they are inherently suscepti-ble to anomalies. The immense costs associated withthe loss of the launch vehicle clearly show the im-portance of a suitable condition monitoring system.Machine condition monitoring systems must providea proper diagnosis in real-time from existing sensordata to detect abnormal behavior and, for example,to trigger an emergency shutdown. Condition mon-itoring algorithms must be able to detect all criticalfaults that may occur during the operation of a rocketengine. Furthermore, it is not enough to detect andcategorize a fault. One must also be able to react ac-cordingly to it. Fault-tolerant control systems aim atan optimal control reconfiguration and satisfy the in-dustrial demand for enhanced availability and safety,in contrast to traditional reactions to faults, whichbring about sudden shutdowns and loss of availabil-ity.

Modern control methods offer significant advan-tages compared to classical techniques. However, theuse in a safety-critical area such as space transporta-tion places great demands on robustness and reliabil-ity. For this reason, the testing of appropriate soft-ware is of utmost importance.

2. Rocket Propulsion Systems Control

Rocket propulsion systems produce thrust byejecting a gas that has been accelerated to high speedthrough a propelling nozzle. Efficient engines usethe combustion of reactive chemicals, consisting offuel and oxidizer components, within a combustionchamber to supply the necessary energy. Accordingto Newton’s third law, the gas expansion pushes theengine in the opposite direction. The net thrust of arocket engine depends on the mass flow of the exhaustgas and its effective exhaust velocity. Changing thedirection of the thrust can be achieved by gimbalingthe whole engine, which is the standard procedurefor big liquid rocket engines used on modern launchvehicles. For a given propellant combination, nozzlegeometry, and ambient conditions, the magnitude ofthe thrust is a function of the combustion chamberpressure and mixture ratio, i.e. oxidizer to fuel ratio.Thus, those factors are usually the focus of the maincontrol loops and are controlled via adjustable flowcontrol valves.12 Linear actuators are used to changethe gimbal angle of the engine and therefore the direc-

𝜙 |𝐹|

Vehicle

Fig. 1: Gimbaled Engine

tion of the thrust, compare Fig. 1. Further actuatorsare the ignition systems. Sequences like the start-up of a rocket engine must be carried out in a closelycoordinated manner. Pressure peaks may cause back-flows in the feed lines, which can destroy the engine.Other constraints are given by temperature and ro-tational speed limits. Especially the reuse of enginesnecessitates the use of condition monitoring methods.This is the only way to monitor changes in systemhealth. Condition monitoring and anomaly detectionare also a prerequisite for health management tech-niques and emergency shutdowns. As already men-tioned, it is common to handle the demanding tran-sients in open-loop. In steady-state operation, classiccontrol methods, e.g. of the PI type, are used if at all.Concerning the condition monitoring simple red linemethods are state-of-the-art. Although classical con-trol algorithms are well tested and can be made veryrobust, they have fundamental limitations and maynot be sufficient for future reusable engines in termsof performance. To make matters worse, the perfor-mance of a reusable engine might degrade over timedue to soot depositions, increased leakage mass flows,which are caused by seal aging, or turbine blade ero-sions. Therefore, it must be possible to easily adaptthe control system to guarantee optimal performance.The last years have witnessed an enormous interestin the use of artificial intelligence methods for auto-matic control, especially machine learning algorithmsincluding neural networks. These methods seem to be

IAC–2020–C4.1.15 Page 2 of 8

71st International Astronautical Congress, Virtual, 12-14 October 2020. Copyright c© 2020 by International AstronauticalFederation (IAF). All rights reserved.

Trained Network Rocket Engine ControllerSupervised / Reinforcement

Learning

Training (CPU/GPU) Format Conversion Deployment - Inference

Fig. 2: Development Flow. Figure adapted from Fig. 1 by Nadeski11

.

particularly promising for adaptive control, fast opti-mal control, and condition monitoring, but challengesremain.

A major challenge in rocket engine applications isthat the control systems are safety-critical. A rocketengine can destroy or degrade itself and the environ-ment around it if improperly controlled. Reliabilityand safety have top priority in aerospace applications.Considering operation constraints is fundamentallynecessary in such cases. Constraints are not onlyimportant during system operation, but also duringadaption phases as well. Although the stability ofneural network controllers is in general not guaran-teed, there has been promising recent work on certi-fying the stability of certain control policies.8 Theverification and validation of neural networks andadaptive systems have intensively been examined aswell.7 Another challenge is given by the limited com-putational resources associated with typical embed-ded systems used in space applications.

3. Deployment on Embedded Systems

As precisely stated by Roth et al.,13 the computingpower during the operating phase of machine learningalgorithms is typically limited. This is particularlytrue for space applications, where robust and fail-safe computing hardware is used in the harsh envi-ronment of space or during a rocket launch. Further-more, high-performance computing hardware withhigh weight is not feasible for space launch vehiclesas weight is one of the most critical design param-eters limiting the launcher’s performance. In recentyears, different techniques were developed to reduce

the computational demands during neural networktraining and inference.13 These techniques can re-duce memory usage, and increase inference speed andenergy efficiency.

However, these improved methods are only neces-sary for very deep and complex neural network ar-chitectures used for computer vision (e. g., object de-tection or video tracking), natural language process-ing, or when training the neural network directly onthe embedded system. For most applications, it isenough to deploy only the trained neural network onthe embedded system. One would train the neuralnetwork using suitable data sets (supervised learning)or simulation environments (reinforcement learning)on a dedicated workstation in Python with commonlyused machine frameworks such as TensorFlow or Py-Torch. Then, one would convert the neural networkto C/C++, and copy the network to the embeddedsystem on the actual space system, compare Fig. 2.Thus, the embedded system just needs to handle theinference, which has much lower computational de-mands than the training of the network.

Furthermore, the necessary network architecturesfor engine control and condition monitoring are muchsmaller than those used for computer vision. Forexample, a closed-loop controller of the combustionchamber pressure and mixture ratio was realized withonly two hidden layers of 400 and 300 neurons.19

A single inference on this network needs only ap-proximately 250 000 floating-point operations (flops).A space-graded microcontroller used in a variety ofspacecraft is the RAD750 with a computing perfor-mance of 80 million flops per second.10 Clearly, thismicrocontroller can handle the inference in real-time.

IAC–2020–C4.1.15 Page 3 of 8

71st International Astronautical Congress, Virtual, 12-14 October 2020. Copyright c© 2020 by International AstronauticalFederation (IAF). All rights reserved.

DevelopmentPC

EmbeddedController

HIL Simulator

Sim

ula

ted

Senso

r D

ata

Contro

l Inputs

Set-P

oin

t

Fig. 3: HIL Configuration

For computer vision or similar applications, whichneed bigger neural networks and more computing per-formance, rugged computers based on industrial AIboards could be used after strengthening them forspace conditions.

There is a loss in accuracy if a neural networkis trained only in a simulated environment. Espe-cially when aerodynamics and fluid dynamics haveto be considered. There are reinforcement learningapproaches that explicitly consider modeling errors.Domain randomization can produce controllers thatgeneralize well to a wide range of environments.17

Nevertheless, fine tuning of the control law on thereal propulsion system is preferable. Since the testbench measurement command and control systemslacks corresponding restrictions on computing power,such a procedure would be executed using test benchinfrastructure. For control law adaption one couldalso use recorded flight data.

4. Hardware-in-the-Loop Testing

Hardware-in-the-loop (HIL) testing2 refers to theinclusion of physical devices (hardware) into the sim-ulation process. A typical example would be connect-ing a real physical controller, i.e., the same one thatwould perform this duty in the actual plant, with asoftware simulation of the plant. This allows validat-ing a control system without putting the plant (inour case the propulsion system) at risk. It is not

DevelopmentPC

EmbeddedController

HIL Simulator

Sim

ulat

edSe

nsor

Dat

a

Control Inputs

Set-

Poi

nt

Actuators

Position

Command

HydraulicTorque

Position

Fig. 4: Extended HIL Configuration

only more economical to develop and test while con-nected to a HIL simulator than the real plant, butalso increases the scope of the testing. HIL tests en-able testing at or beyond the range of the nominalengine parameters and verification of the system atfailure conditions. A suitable mathematical systemrepresentation is required, which must be real-timecapable.

A suitable simulation environment for rocketpropulsion systems is given by EcosimPro, whichis a modeling and simulation tool for 0D and 1Dmultidisciplinary continuous and discrete systems.In EcosimPro, the system description is based ondifferential-algebraic equations and discrete events.Within a graphical user interface, one can combinedifferent components, which are arranged in sev-eral libraries. Of particular interest are the Euro-pean Space Propulsion System Simulation (ESPSS)libraries, which are commissioned by the EuropeanSpace Agency (ESA). These EcosimPro libraries aresuited for the simulation of liquid rocket engines andtest benches. Employees of DLR Lampoldshausenhave many years of experience in the use of Ecosim-Pro/ESPSS and used it for example for the designand optimization of the LUMEN engine.5 Further-more, advanced machine learning based surrogatemodels can be integrated into EcosimPro to improvethe simulation quality.6,18

Using an EcosimPro deck, i.e. is a simulationmodel designed to run as a standalone black box

IAC–2020–C4.1.15 Page 4 of 8

71st International Astronautical Congress, Virtual, 12-14 October 2020. Copyright c© 2020 by International AstronauticalFederation (IAF). All rights reserved.

independent from the main program, and so calledS-functions the model can be easily exported toSimulink. Via the Simulink Coder tool, a package isobtained that is ready for use on different platforms,which can act as HIL simulators like the National In-struments PXI platforms. It must be checked thatreal-time conditions are satisfied in the execution ofthe model selecting an appropriate sample time.

Fig. 3 shows a HIL set up which includes an elec-trical emulation of all sensors and actuators. Theembedded controller receives simulated sensor datafrom the HIL simulator and sends control inputs tothe HIL system.

As a next step, real actuator hardware can be in-cluded. Electrically operated flow control valves arecomposed of three parts: the hydraulic part, the ac-tuator part and the electronics command part. Theactuators, which can be given by one or two electricalmotors, can easily be included in the HIL tests, com-pare Fig. 4. The hydraulic parts have to be replacedby other actuators which provide a realistic torquecalculated by the HIL simulator. Obviously, not allactuators can easily be included in this way, e.g. igni-tion systems are still emulated. A more complex testset up also allows the partial inclusion of representa-tive environmental conditions, e.g., temperature andvibrations.

The HIL simulator can also simulate the occur-rence of anomalies and the gradual deterioration ofsystem health. Therefore this system is perfectlysuited for testing the anomaly detection and condi-tion monitoring functions of the developed real enginecontroller.

5. DLR Rocket Engine Test Facilities

Although HIL tests can make a significant contri-bution to the development of rocket engine controlsystems, they cannot completely replace tests of thecontrol system during the firing of a real engine us-ing a suitable rocket engine test facility. The mainreason is that even the most sophisticated simulationmodels have prediction errors due to not included ef-fects or model miss-specification. This is especiallytrue for complex systems like rocket engines. Thus,real engine tests are needed for system identificationand tuning of the controllers. Fig. 6 shows the devel-opment process for a new engine controller. In thefirst step, the implemented control software is testedwithin a modeling environment to prove the controlsoftware design (software-in-the-loop testing). Next,the previously discussed hardware-in-the-loop testsare performed to test the embedded control archi-

Fig. 5: P8.3 Engine Test Facility

tecture. Finally, the developed control system andthe simulation models are validated by firing tests ofa real engine at a suitable rocket engine test facility.

Sensor data from hot firing tests also provide thetraining data for machine learning based conditionmonitoring systems. Fortunately there are machinelearning algorithms, which only need data for nomi-nal operation. Once trained, neural networks can beused to distinguish nominal and off-nominal behav-ior. By taken the correlations and interactions be-tween various variables into account, the transitionto off-nominal behavior is detected before a classicalred line gets triggered.

The test facilities at the DLR Institute of SpacePropulsion are a fundamental pre-requisite for de-veloping engine technologies to operational maturityand ensuring their quality. The P4 altitude simu-lation test bench played a indispenable role duringthe development of the Vinci upper stage engine,which will be on the upper stage of the Ariane 6launch vehicle. The P5.2 test facility can even beused to qualify not only engines and individual com-ponents, but also complete cryogenic upper stages.Perfectly suitable for researching advanced technolo-gies for future space engines is the P8 test bench,which is operated by CNES, DLR and ArianeGroup.At the interfaces to the test objects in the test cellsP8.1 and P8.2 the high pressure supply system pro-vides fuel at pressures of up to 360 bar. This allowsfor fundamental research and technology testing ofpre-combustion chambers, gas generators and maincombustion chambers for rocket propulsion systems.With the launch of test cell P8.3, , compare Fig. 5, alow-pressure supply will also be available, so that in-vestigations on rocket engines with pump supply arepossible there.

IAC–2020–C4.1.15 Page 5 of 8

71st International Astronautical Congress, Virtual, 12-14 October 2020. Copyright c© 2020 by International AstronauticalFederation (IAF). All rights reserved.

Software in the Loop

Control Software Design

Hardware in the Loop

System Testing Testbenches

Acceptance Test Requirements

System Modeling

Controller Design

Implementation

Fig. 6: Development Process

6. Zentrum fur Telematik (Zft) Test Facility

In order to characterize the thrust vector, thesatellite orientation plays a crucial role for resultingorbit changes.9 The high precision pointing capabil-ities, which are necessary in a typical dynamic spaceenvironment, can be tested with the turntables of“S4 – Smart Small Satellite Systems GmbH” in theZfT facility. The concept of the Dynamic Bench TestFacility, compare Fig. 7, is based on the combinedoperation of two high-precision three-axis motionsimulators, sensor stimulators and a simulationcomputer system. Arranging motion simulators andsensor stimulators into proper geometrical configu-ration enables HIL simulations, as well as materialstress tests, sensor characterization, and calibration.

This unique precision facility offers testing ofmulti-satellite14 systems with a focus on formationcontrol, relative navigation, inter-satellite communi-cation links and cooperative target observation.15

7. Conclusion

As the demands on the rocket propulsion controlsystems continue to rise, the control software is alsobecoming increasingly complex. The complex soft-ware requires more powerful hardware in turn. Mod-ern embedded system can handle neural networksand diverse model-based approaches. Thus, exten-sive testing becomes even more important and is in-

Axis Outer Middle Inner

Range [◦] ±120 ±120 ∞Torque [N m] 2x4000 2x2500 95Max. Acceleration [◦ s−2] 1000 6000 2000Min. Velocity [◦ s−1] 10−4 10−4 10−4

Max. Velocity [◦ s−1] 150 150 150Pointing Accuracy [◦] 10−4 10−4 10−4

Orthogonality [arc sec] < 5Intersection [mm] 0.2

Table 1: Characteristics of the Turntable

evitable during the development process to guaranteethe reliability of the control system. We sketched adevelopment and test roadmap for rocket propulsionsystems. First, the testing is completely performedin simulation. Second, step by step some virtual el-ements are replaced with real hardware (hardware-in-the-loop tests). Third, real embedded systems aretested with real rocket engines using advanced rocketengine test facilities.

7.1 Acknowledgment

Icons in Fig. 2, Fig. 3, and Fig. 4 made by Freepik,Payungkead, and prettycons from www.flaticon.

com. The authors would like to thank Tobias Kaiserfor valuable discussions concerning space-graded em-bedded systems.

IAC–2020–C4.1.15 Page 6 of 8

71st International Astronautical Congress, Virtual, 12-14 October 2020. Copyright c© 2020 by International AstronauticalFederation (IAF). All rights reserved.

Fig. 7: Dynamic Test Benches at ZfT

References

[1] Hiroya Asakawa, Masahiro Tanaka, TakeshiTakaki, and Kazuhiro Higashi. Component testsof a LOX/methane full-expander cycle rocket en-gine: Electrically actuated valve. In Proceedingsof the 8th European Conference for Aeronauticsand Space Sciences, Madrid, Spain, 2019.

[2] Marko Bacic. On hardware-in-the-loop simula-tion. In Proceedings of the 44th IEEE Conferenceon Decision and Control, Seville, Spain, 2005.IEEE.

[3] Jean-Noel Chopinet, Francois Lassoudiere,G. Roz, O. Faye, S. Le Gonidec, Patrick Alliot,and Guedron Sylvain. Progress of the develop-ment of an all-electric control system of a rocketengine. In 48th AIAA/ASME/SAE/ASEE JointPropulsion Conference, Atlanta, Georgia, 2012.American Institute of Aeronautics and Astro-nautics.

[4] Stephane Colas, Serge Le Gonidec, PhilippeSaunois, Martine Ganet, Antoine Remy, andVincent Leboeuf. A point of view about the con-trol of a reusable engine cluster. In Proceedingsof the 8th European Conference for Aeronauticsand Space Sciences (EUCASS), Madrid, Spain,2019.

[5] Jan Deeken and Gunther Waxenegger. LUMEN:Engine cycle analysis of an expander-bleeddemonstrator engine for test bench operation. InDeutscher Luft- Und Raumfahrtkongress, Braun-schweig, Germany, 2016.

[6] Kai Dresia, Gunther Waxenegger-Wilfing, JorgRiccius, Jan Deeken, and Michael Oschwald.Numerically Efficient Fatigue Life Prediction ofRocket Combustion Chambers using ArtificialNeural Networks. In Proceedings of the 8th Euro-pean Conference for Aeronautics and Space Sci-ences (EUCASS), Madrid, Spain, 2019.

[7] J. Hull, D. Ward, and R.R. Zakrzewski. Ver-ification and validation of neural networks forsafety-critical applications. In Proceedings of the2002 American Control Conference (IEEE Cat.No.CH37301), 2002.

[8] Ming Jin and Javad Lavaei. Stability-certifiedreinforcement learning: A control-theoretic per-spective. arXiv:1810.11505 [cs], 2018.

[9] Alexander Kramer, Philipp Bangert, and KlausSchilling. UWE-4: First Electric Propulsion ona 1U CubeSat - In-Orbit Experiments and Char-acterization. Aerospace 2020, 7, 2020.

[10] George Lentaris, Konstantinos Maragos,Ioannis Stratakos, Lazaros Papadopoulos,Odysseas Papanikolaou, Dimitrios Soudris,Manolis Lourakis, Xenophon Zabulis, DavidGonzalez-Arjona, and Gianluca Furano. High-Performance Embedded Computing in Space:Evaluation of Platforms for Vision-BasedNavigation. Journal of Aerospace InformationSystems, 15(4):178–192, 2018.

[11] Mark Nadeski. Bringing machine learning to em-bedded systems. Texas Instruments, 2019.

IAC–2020–C4.1.15 Page 7 of 8

71st International Astronautical Congress, Virtual, 12-14 October 2020. Copyright c© 2020 by International AstronauticalFederation (IAF). All rights reserved.

[12] Sergio Perez-Roca, Julien Marzat, Helene Piet-Lahanier, Nicolas Langlois, Francois Farago,Marco Galeotta, and Serge Le Gonidec. A sur-vey of automatic control methods for liquid-propellant rocket engines. Progress in AerospaceSciences, 107:63–84, 2019.

[13] Wolfgang Roth, Gunther Schindler, MatthiasZohrer, Lukas Pfeifenberger, Robert Peharz,Sebastian Tschiatschek, Holger Froning, FranzPernkopf, and Zoubin Ghahramani. Resource-Efficient Neural Networks for Embedded Sys-tems. arXiv:2001.03048 [cs, stat], 2020.

[14] Oliver Ruf, Stephan Busch, and Klaus Schilling.Challenges and Novel Approaches for TestingLarge Numbers of Small Satellites. In Pro-ceedings of the 68th International AstronauticalCongress, Adelaide, Australia, 2017.

[15] Oliver Ruf and Klaus Schilling. AutomatedAttitude Control Verification of Nano-Satelliteson System-Level. In Proceedings of the 33rdAIAA/USU Conference on Small Satellites, Lo-gan, USA, 2019.

[16] P. Simontacchi, R. Blasi, E. Edeline, S. Sagnier,N. Ravier, A. Espinosa-Ramos, J. Breteau, andPh. Altenhoefer. PROMETHEUS: Precursor ofnew low-cost rocket engine family. In Proceedingsof the 8th European Conference for Aeronauticsand Space Sciences (EUCASS), Madrid, Spain,2019.

[17] Josh Tobin, Rachel Fong, Alex Ray, JonasSchneider, Wojciech Zaremba, and PieterAbbeel. Domain Randomization for Transfer-ring Deep Neural Networks from Simulation tothe Real World. arXiv:1703.06907 [cs], 2017.

[18] Gunther Waxenegger-Wilfing, Kai Dresia,Jan Christian Deeken, and Michael Oschwald.Heat Transfer Prediction for Methane in Regen-erative Cooling Channels with Neural Networks.Journal of Thermophysics and Heat Transfer,34(2):347–357, 2020.

[19] Gunther Waxenegger-Wilfing, Kai Dresia,Jan Christian Deeken, and Michael Oschwald.A Reinforcement Learning Approach forTransient Control of Liquid Rocket Engines.arXiv:2006.11108 [cs, eess, math, stat], 2020.

IAC–2020–C4.1.15 Page 8 of 8

View publication statsView publication stats