Overview of advances in ASDEX Upgrade plasma control to support critical physics research for ITER and beyond

The successful operation of fusion reactors requires plasma scenarios with good core confinement and acceptable first wall heat loads that are stable and robust to external perturbations. This poses both physical and technological challenges. One of the technologies that addresses these challenges is a complex feedback control system that supports advances in physical understanding and helps to ensure stable operating conditions. The operation of marginally stable plasmas often leads to off-normal events (such as disruptions) and feedback control can prevent these to some extent. This contribution gives an overview of the main results of the development and operation of the feedback control algorithms on ASDEX Upgrade (AUG). Fueling actuators, using a combination of gas valves and pellet injection, can simultaneously control neutral density of the divertor and the density of the plasma core above the Greenwald limit. Impurity injection is employed to control the position of the X-point radiator, allowing the creation of an ELM-suppressed H-mode with high radiation fraction. Heating actuators are used to control the plasma energy content, which supports advanced tokamak experiments and enables stable I-mode operation, and the electron temperature control, which supports turbulence studies. In control technology, AUG has pioneered the use of virtual actuators, which allow effective use of the limited number of heating actuators, adaptive control policies, and exception handling. Such technologies will also be used in ITER. Advanced nonlinear state observers (RAPTOR, RAPDENS) and codes to evaluate the power deposition properties (RABBIT, TORBEAM) are available for routine use in the AUG feedback controllers. Extensive use of the AUG discharge control system further enhances the research capabilities of this machine.


Introduction
The ultimate goal of tokamak research is a reliable and economically competitive fusion power plant with low environmental impact and inherently safe operation.Fusion power generation requires good confinement of the plasma core with temperatures and densities sufficiently high for effective fusion power production.This must be compatible with an acceptable heat exhaust solution that limits the steady state heat loads on the divertor targets to approximately 10 MW/m 2 , otherwise the unmitigated divertor heat loads in DEMO would reach 300 MW/m 2 which is not compatible with any of the existing materials.In order to reduce the divertor heat fluxes to acceptable levels, more than 95% of the exhaust power must be dissipated, thus reducing the heat flux reaching the divertor targets [1].This is possible in detached regimes, which are characterised by significantly reduced heat and particle fluxes to the divertor plates and are typically achieved by injecting a mixture of impurities and fuel gas.The simultaneous achievement of good core confinement and divertor detachment often drives the plasma into only marginally stable regimes, prone to instabilities that can degrade confinement or cause disruptions with a sudden loss of plasma kinetic and magnetic energy, accompanied by effects that can lead to damage to the device [2,3].For example, keeping the plasma detached can drive the discharge to a density limit [4].Violation of the density limit can lead to confinement degradation and disruption.
For a successful operation of a fusion reactor, the plasma state must be monitored in real-time and if deviations from the desired state occur, action must be taken to eliminate them.This is the role of tokamak control systems, which use feedback control to monitor the plasma state and take appropriate action if the state deviates from the reference.In addition, tokamak control systems face the challenge of selecting the optimal actuators available to achieve a given control objective, which is handled by a control system component called the actuator manager.Finally, feedback control of plasma parameters is an important tool to support the physics studies that establish and stabilize the required plasma conditions.For example, in a diverted plasma such as that planned for ITER or DEMO, the plasma tends to be vertically unstable and no physics studies with this type of device are possible without feedback control of the plasma vertical position or vertical velocity of the plasma [5].
ASDEX Upgrade (AUG) is a medium sized tokamak (R = 1.65 m, a = 0.5 m, B t ⩽ 3.2 T, I p ⩽ 1.4 MA) equipped with a full tungsten wall.It is used to study ITER and DEMO relevant issues [6].Plasma feedback control research is fully integrated into its experimental program, and AUG faces similar issues and challenges in this area.These challenges are even more pronounced for larger devices.Both the ITER plasma control system (PCS) and the AUG discharge control systems (DCS) have to deal with highly non-linear systems, sometimes without any existing quantitative model.Both control systems face similar problems in terms of diagnostics or actuator failures, that need to be properly handled.Disruption avoidance is critical for ITER, and many of the methods can already be tested on AUG.The ITER PCS [7], like the AUG DCS, will be highly modular and configurable.
The DCS places great emphasis on reproducibility and predictability.Due to the high configurability of the DCS, past discharge programmes can be run with minimal effort to adapt the control system.Control algorithms based on a simple and transparent set of rules are preferred, so that the behaviour of the system can be predicted and understood.The AUG control system facilitates studies in several key areas of tokamak research and this paper presents the most significant results that have been achieved.The first of these studies is access to the reactor-relevant regimes at the very high density required for fusion power production (see section 3).For this purpose, a density observer RAPDENS (see section 2.1) as well as methods for handling the discrete nature of the pellet injection (see section 3) have been developed.The control system is an important tool for exploring novel exhaust solutions that critically depend on feedback control of effects without an existing quantitative physics model (see section 4).It also helps to perform physics studies focusing on fast particle transport by stabilizing the electron temperature, or on the LH transition by changing the ion heating (see section 5).The tools to compute those quantities, RAPTOR and RABBIT (see section 2.1) have been developed and used in the real-time control system.Feedback control also allows exploration of the operating space around the density limit (see section 6).The operation of AUG and the applications mentioned above have benefited from the introduction of virtual actuators (see section 2.3), which have improved robustness to heat source failure and increased the actuation amplitude of the heating controllers.

AUG real-time controllable actuators
AUG operates a large number of controllable actuators.Firstly, there are 8 neutral beam injectors (NBI) [8], which can provide up to 20 MW of total power.The electron cyclotron resonance heating (ECRH) [9] is provided by 8 gyrotrons, with a total power of up to 6 MW.Four of these gyrotrons are connected to fast-moving mirrors, that can be controlled in real-time to deposit the ECRH power at the locations required by the control system.The ion cyclotron resonance heating (ICRH) can be controlled on 2 pairs of antennas and provides up to 5 MW of power [10].
There are several independent gas circuits that can be controlled in real-time.Three of these circuits are typically used to supply the plasma fuel (deuterium, hydrogen, or helium).The other two circuits are used to supply impurities: either heavier elements such as nitrogen or noble gases for increased divertor radiation, or gases such as methane that are used for impurity transport studies [11].To fuel the core, there is a real-time controllable pellet system that injects pellets from the highfield side of the vacuum vessel at a frequency of up to 80 Hz. [12][13][14].
For the magnetic actuators, 8 independent poloidal field coils and 3 central solenoid coils are controlled by 10 independent power supplies providing feedback control of plasma current, position, and shape.In addition, the currents of the resonant magnetic perturbation coils currents are also controllable, potentially allowing the error field control [15].

AUG real-time diagnostics
Many of the AUG diagnostics provide real-time data to the control system.There are 58 magnetic probes and 18 flux loops [16] which provide local measurements of the magnetic field or magnetic flux.A set of Mirnov coils is used to provide information on magnetohydrodynamic mode number and amplitude.The line-integrated electron density is measured by 5 interferometers with different lines of sight [17].There are two detectors that measure the total bremsstrahlung radiation and 128 bolometers that measure the line-integrated radiation along different lines of sight [18].In the lower divertor, the line integrated radiation is measured by an AXUV camera along 16 lines of sight [19].The electron temperature (T e ) is measured by 60 electron cyclotron emission (ECE) channels at different radial positions [20].The mapping of the ECE measurement to the flux surface coordinates is performed by the real-time equilibrium reconstruction [16].

The AUG discharge control system
The commands to the real-time controllable actuators are generated by the Discharge Control System, which processes the diagnostic signals into high-level variables to be controlled and employs a comprehensive set of feedback control strategies.For example, magnetic signals are used to calculate the magnetic axis coordinates and plasma shape parameters.The standard calculations are performed in a synchronized data-driven chain of evaluations, but algorithms that take longer time to complete their calculations can be run asynchronously at a rate slower than the cycle time of the control system (typically 1.5 ms).The standard real-time processes, which run every cycle, use the latest result of the asynchronous algorithms.Technical details of the DCS architecture have been reported elsewhere [21][22][23].

Data evaluation algorithms
Diagnostic signals are processed by several advanced data processing algorithms implemented in the control system.Plasma position, shape, and β are computed in real-time using function parametrization [24,25].This method relies on the statistical analysis of a large database of simulated equilibria to obtain a functional representation (typically quadratic polynomials) of the above mentioned physical parameters in terms of the values of the magnetic probe and flux loop measurements.Full equilibrium reconstruction of the plasma shape and flux surface geometry is provided by the real-time Grad-Shafranov equation solver JANET [16].The existing LabVIEW-based version has been rewritten in C++17 and runs as a DCS satellite application [26,27].The transition from LabVIEW to C++17 is essential for the long-term maintainability of the real-time equilibrium reconstruction code.
An extended Kalman filter (EKF) based state observer, RAPDENS, is used to estimate the plasma density profile.The version implemented on AUG provides the value of plasma density at 12 normalised toroidal flux coordinates [28,29] and runs every DCS cycle (execution time is about 0.6 ms).RAPDENS uses an empirical particle balance model including particle sources, sinks, and several versions of a 1-D particle transport model based on Bohm-Gyrobohm [30].The model estimates are corrected using 5 interferometer channels and 2 bremsstrahlung measurements.This observer is able to detect diagnostic failures such as fringe jumps, that often occur when using ICRH.In some cases this can be corrected by offsetting the error introduced by the fringe jumps, but in most cases the interferometer affected by the fringe jumps is ignored for the rest of the discharge [31].Thanks to its predictive model, the RAPDENS observer even provides a reliable density estimate in the pellet-fuelled discharges.
Another EKF based state observer, RAPTOR, solves the coupled 1-D electron transport equation and 1-D flux diffusion equations [32].This algorithm provides real-time information on the temperature and current profiles on a grid of 11 normalised toroidal flux coordinates and runs every 6 ms.Predictive model estimates are corrected using the ECE measurements mapped to normalised flux coordinates.
The deposition location of the external heating, the fraction of heat that is absorbed by electrons and ions, and the current drive profile are important inputs to the transport and current diffusion equations solved in RAPTOR.These quantities are calculated in real-time for each NBI beam by the RABBIT code [33][34][35] and for each gyrotron by the TORBEAM code [36].The computation time of both codes is about 20 ms and is performed in parallel.
An X-point radiator (XPR) is a strongly radiating plasma zone within the separatrix near the X-point.Detection of the XPR is important for the heat exhaust (see part 4) and disruption avoidance (see part 6) and is performed by applying a peak detection algorithm to the AXUV photodiode lines of sight in the divertor [37].

Feedback control schemes
In the control system, each actuator is allocated to a single feedback controller, which can be either single input single output, or multiple input multiple output (MIMO).The feedback controllers are designed to be sufficiently independent of each other, and an action of one controller is seen as a disturbance by the other controllers.When the action of the actuator interacts strongly with more than one of the controlled quantities, MIMO controllers are used.The most prominent use for MIMO control is the plasma shape control, where a change in the poloidal field coil current affects several controlled quantities (e.g.gaps between the wall and the separatrix).Heating controller (e.g.plasma β or the electron temperature control) uses the ECRH, NBI, and ICRH power, the density controller uses the fuel gas rate, valves and pellet injectors, and the radiation controller uses the impurity gas valves.A plasma current controller uses the central solenoid and the shape and position controller uses the remaining poloidal field coils.If neoclassical tearing mode (NTM) control is required, a dedicated gyrotron and its fast mirror can be used.Actuator management allows multiple NBI sources and gyrotrons to be combined into more powerful virtual actuators (see part 2.3).
The controller input error e can be dynamically scaled where appropriate by a time-varying factor S, which describes the efficiency of a control action on the controlled variable.Let us demonstrate this using a scaled PI control policy, where the controller output would be computed as follows: where P is the proportional gain and I is the integral gain of the controller.For example, the input error of the shape controller is multiplied by the plasma current to account for the fact that it is more difficult to shape the plasma at higher currents.This has the same effect as changing the controller gain based on the plasma state.

Actuator manager: virtual actuators
Some categories of actuators (gyrotrons, NBI, ICRH antennas) are often similar to each other in terms of control action.This makes them interchangeable, and allows the aggregation of these actuators into groups called virtual actuators, which are treated as individual actuators by the controller.The assignment of the actuators to virtual actuators depends on the experimental goals and is done manually by the pulse designer.Virtual actuators provide the controller with their limits and distribute the controller commands to the individual actuators.This increases the range of actuation amplitude and increases the robustness of individual actuators against tripping.A virtual actuator scheme has been introduced for ECRH, NBI, and ICRH [41,42] and is currently being considered for ITER.

Density control
For ITER and DEMO, a pellet-fueled scenario with core density above the Greenwald limit and a divertor pressure that allows detachment, but does not lead to confinement degradation is needed.Such a scenario was developed and feedbackstabilized at AUG by simultaneously controlling the core density and divertor density of neutral particles using pellet and gas injection [14].
There were several challenges in achieving this goal.It was necessary to develop reliable density measurements in pelletfueled discharges.Interferometers are corrupted by pelletinduced fringe jumps that cannot be corrected.The only realtime core density measurement available in these discharges is from bremsstrahlung measurements.This signal is very noisy, making its direct use by the controller difficult.This problem has been solved by the RAPDENS algorithm (see section 2.1), which reconstructs the core density input to the controller using its particle transport model and the bremsstrahlung measurement.
The pellet injector actuator does not act homogeneously on the plasma, but injects frozen pellets intermittently, depending on the desired particle flux.In the control scheme employed at AUG, the density controller modifies the desired particle flux to be provided by the pellet injection.The particle flux request is converted by the DCS into a pellet injector duty cycle, which is commanded to the local pellet injector controller.The pellet injector duty cycle is converted into a sequence of discrete pellet injections by the local sigma-delta [43] controller [44].The auxiliary gas puff is reduced while pellets are injected to avoid an excessive total fuel flux, which could potentially lead to an edge density limit and confinement degradation.The neutral density in the divertor is used as a suitable measure of the edge density, even when pellets are injected [45].
Figure 1 shows the time evolution of a discharge in which pellets drive the core electron density, while reducing of the gas puff rate compensates for the increased particle flux from pellet injection and stabilizes the neutral density level in the divertor.This result inspired the development of dedicated control techniques to improve the handling of discrete pellet injections [46] and to include real-time evaluation of the gas fuelling efficiency, which will help to cope with the coupling of gas puffing and pellet injection not currently considered in the control scheme [47].

Detachment control
The heat load on the target plates of the divertor must be controlled to avoid damage.In AUG, feedback control is typically achieved by measuring the thermocurrent flowing to the target plate (measured indirectly as the voltage across a shunt) and injecting impurity gas to detach the plasma from the target [48,49].
In full detachment the thermocurrent is no longer a valid measurement, but an XPR can be observed, which is a poloidally localized, highly radiative region inside the separatrix close to the X-point [50].An XPR on AUG can be established by injecting impurity gas (either N 2 or Ar ).Establishment of the XPR and control of its location allows access to a new detached and ELM-suppressed regime, which is inherently compatible with safe divertor operation, and is characterised by a good core confinement with H 98 > 0.9.A major advantage of this regime is that it exists for a relatively wide range of the XPR vertical positions within the confined plasma, which can be measured and controlled.When the plasma is close to reattachment, it can be seen that the XPR is close to the Xpoint.If the XPR is too far inside the confined plasma, there is a risk that it will mutate into an unstable MARFE and possibly cause a disruption (see also section 6).However, this regime cannot be reliably achieved without feedback control.
In AUG, the vertical position of the XPR is identified as a radiation peak above the X-point and it is measured by 16 vertically distributed AXUV photodiodes with lines of sight in the vicinity of the X-point.It can be controlled by changing the impurity injection rate, with increased injection causing the XPR to move away from the divertor into the confined plasma, and reduced injection causing it to move towards the divertor.Stable operation of the XPR within the confined plasma region at up to 15 cm above the X-point was demonstrated.The trajectories of the XPR position and actuator response are shown in figure 2. The ELM-suppressed regime is obtained at a position higher than 7 cm.
A reduced model has been derived to explain the physical mechanisms for initiating a stable XPR by analysing the power balance of the parallel heat flux between the upstream position and the XPR volume [51].For feedback control of the XPR, system identification techniques were used to calculate the response of the XPR motion to impurity gas injection by measuring the XPR response to impurity injection sequentially modulated at different frequencies.Based on the frequency response, transfer functions were identified and then used to derive the controller gains for the PI controllers [52] for N 2 which feedback control the amount of the injected gas.

Heating control
Heating actuators are used in feedback control for several applications.AUG allows the plasma β to be controlled using either ECRH or NBI.This allows studies of the I-mode, an ELM-free regime characterised by the presence of a temperature pedestal and the absence of a density pedestal, which exists only within a narrow β window and can only be stabilized by feedback control of the plasma β [53,54].Typically, NBI is used for this purpose.However, in a reversed plasma current configuration, NBI suffers from low efficiency [55] and ECRH is used instead.Another application of the β controller is to assist in the creation of advanced tokamak scenarios [56] by varying the plasma pressure in a controlled manner, thereby helping to create the required current profile.
AUG also has an electron temperature controller, which uses the temperature profile from RAPTOR and controls the ECRH power to keep the electron temperature profile constant during discharges.This has been useful, for example, in developing medium and high power ICRF heating scenarios to investigate and ultimately control the type and level of turbulence by changing the ratio of the electron to ion temperature.Keeping the electron temperature stable while changing the ion temperature and turbulence type makes these studies easier to perform and interpret [57,58].
The final application of the heating actuators presented here is the direct control of the NBI heating properties.In most devices, NBI heating is only controlled by changing system parameters such as injected power or acceleration voltage.AUG has demonstrated feedback control of other NBI heating quantities, such as NBI current drive, fast particle pressure or ion heating.This last type of control is also possible at constant total heating power.Direct control of the ion heat flux (a sum of the ion heating and the equipartition term) has been attempted with promising initial results [59,60].All these quantities are calculated in real-time by RABBIT (see section 2.1).
The following section describes in more detail the feedback control of the ion heating at constant total heating power.Ion heating is the parameter of interest for experiments focusing on the LH transition, which is triggered by the ion heat flux across the separatrix being above a certain threshold rather than by the total separatrix power [61].For low density scenarios, the equipartition coupling between electrons and ions is very small and the ion heat flux is dominated by ion heating.The two actuators that can be used are ECRH, which only heats electrons, and NBI, which heats both ions and electrons.
A MIMO feedback controller calculates the desired NBI power based on its deviation from the ion heating reference.The ECRH power, which is the second output of the same controller, is adjusted to keep the total heating power constant.Figure 3 shows the feedback control of the ion heating at constant total heating power.At the beginning of the discharge, two gyrotrons (1.5 MW) are used to heat the plasma.At 3.0 s, the ion heating reference increase is programmed.Consequently, the controller ramps up the NBI heating power and reduces the ECRH heating power to keep the total heating power constant.During the ramp up of the ion heating power (and therefore increased ion heat flux), the LH transition occurs due to the increased ion heating power, even though the total heating power is constant.The controller is able to follow the ion heating request and keep the total heating power constant except for a time window between 5 and 6 s.This is when the ion heating reference is so high that more NBI power is requested than the total heating power.When such a conflict occurs, the controller will follow the ion heating reference and will not be able to keep the heating power constant as the ECRH power cannot be negative.

Disruption avoidance
Disruptions pose a significant threat to future fusion devices because of the potential damage they can cause.There is an ongoing research into both mitigating the disruption consequences [62] and avoiding them.The development and study of disruption avoidance techniques has been a major topic for AUG in the recent years.A variety of methods to prevent a discharge from disrupting was employed.The main focus has been on disruption avoidance at the density limit, which is characterized by the XPR mutating into an unstable MARFE (see also section 4).The details of this work are discussed in detail in [63].There are only qualitative theoretical models that would describe the transition from XPR to MARFE, and experimental investigation is currently the only method to study this effect.The control system can detect the unstable situation and trigger a recovery action.
Two diagnostic methods have been developed to detect in real-time the unstable state that becomes disruptive.The first is based on direct observation of the XPR/MARFE position above the X-point and activation of disruption avoidance strategies when an empirical threshold is exceeded.The second diagnostic method monitors the position of the global plasma parameters in the space of the critical density fraction and the H 98 factor, where a disruptive region is known from previous experiments [64].When the plasma parameters approach this region, the control system performs disruption avoidance action.The DCS offers the choice of either an abrupt change in the reference trajectories for the actuators to recover the discharge (exception handling), a continuous modification of the actuator actions, or a combination of both.
In response to the evolution of the plasma state, the control system is able to perform scans and evaluate the effectiveness of different types of heating, fueling (pellets, gas, impurities) Figure 4.An example of a non-disruptive scan to study the impact of the N 2 injection rate on the XPR/MARFE stability.The first plot shows the plasma current, the second the heating, third the kinetic energy stored in the plasma, the fourth the injection rate of deuterium and nitrogen in electrons per second, the fifth triangularity, the sixth core and edge density, and the last plot shows the XPR position above the X-point.The red horizontal line in the last plot shows the threshold where the recovery action is triggered.The vertical dotdashed lines show the times when the gas ramps start, and the vertical dotted lines show when the recovery action is triggered.and of shape and position modifications in avoiding disruptions.Once a plasma scenario close to the disruption limits has been established, actuator commands are modified until the plasma becomes disruptive.A recovery action is then taken (typically increased heating) and the plasma is returned to its initial state.This method can then be used to perform a series of scans in a single discharge, allowing the parameter space to be explored in a limited number of discharges without disruptions.
An example of one such scan is shown in figure 4, where the effect of N 2 seeding on MARFE/XPR stability has been studied.This discharge operates at a high density with NBI From these scans, it was found that the MARFE position can be stabilized with additional heating power and that the application of central heating is the most effective.Another important factor is the plasma triangularity.If this is above a certain threshold, the XPR becomes less stable due to improved particle confinement and increased plasma density at the same fueling rate [63].The main destabilizing factor in this case is the gas and impurity injection.Pellet injection has similar effect to gas injection.The effect of the above listed quantities on the XPR/MARFE stability is summarised in table 1.
The methods developed on the AUG for detecting and avoiding density limit disruptions were transferred to TCV and tested.After the necessary adaptations to the different diagnostic setups of the two tokamaks, the density limit disruptions could be avoided in a similar way as on the AUG [65,66].

Conclusions and outlook
AUG has been a pioneer in control engineering and physics research.Advanced control tools such as state observers, configurable control objects and actuator management have been extensively used in AUG operation.These tools have been used to simultaneously control the core density above the Greenwald limit and the divertor neutral particle density.Feedback controlled impurity injection has been used to stabilize the XPR and is a potential solution to the DEMO exhaust problem.β control has been used for Imode and advanced tokamak studies, temperature control has been developed for transport studies, and control of NBI current drive, fast particle pressure, and ion heating has been developed.Feedback control has been extensively used to avoid density limit disruptions.
A future goal is to improve the quality of real-time plasma state information by improving existing state observers and adding more real-time diagnostics.On the observer side, there are plans to improve the equilibrium reconstruction by adding pressure constraints provided by the RAPTOR state observer [47,57].In particular, AUG will benefit from realtime Thomson scattering.This will provide the electron density and temperature profiles.Another very useful diagnostic that will be real-time charge exchange recombination spectroscopy, which will allow the measurement of the ion temperature profile and the plasma rotation profile.The feedback control of these profiles will aid transport studies and model validation.It is planned to improve the real-time feedback control of the power crossing the separatrix by adding real-time Gaussian process tomography to provide real-time radiated power densities from the bolometer measurements [67].Machine learning based algorithms are currently under development [68] and will be used to improve the plasma state reconstruction and real-time decision making capabilities of the control system.Using detailed real-time plasma state information, AUG will emulate the alpha heating power as proposed in [69] using an existing set of actuators.The fusion power emulation technique will be used to study the dynamic behaviour and scenario stability to perturbations in the case of events such as NTM appearance, sawtooth crash, W accumulation, or actuator loss.

Figure 1 .
Figure 1.Simultaneous feedback control of electron density in the core and neutral density in the divertor using pellet and gas injection in discharge 38 760.

Figure 2 .
Figure 2. The control of X-point radiator height above X-point (dz) using N 2 injection in AUG discharge 36 655.The second plot shows an electron flux (a standard particle flux unit used in AUG control) which is carried by the injected N 2 particles.

Figure 3 .
Figure 3. Feedback control of the ion heat flux at constant total heating power using ECRH and NBI as actuators.

Table 1 .
The overview of the impact of modifications of various quantities on the MARFE/XPR stability.MW.At t = 2.5 s, both deuterium and nitrogen injections are ramped up until the XPR/MARFE becomes unstable.This is accompanied by H to L transition, which is seen as a kinetic energy drop in figure4.Normally, a disruption would follow shortly afterwards.In this discharge, however, additional ECRH heating is applied after the position of XPR/MARFE above the X-point exceeds a threshold of 5 cm, and the injection of D 2 and N 2 is brought down.This results in stabilisation of the XPR/MARFE.A similar scan is repeated twice more in the same discharge with different rates of N 2 injection, and the discharge is ramped down without disruption.