Real-time Control of a Tokamak Plasma

Using Neural Networks

Christopher M. Bishop‡ , Paul S. Haynes† , Mike E. U. Smith† ,

Tom N. Todd† , and David L. Trotman†
‡ Current address: Microsoft Research,
7 J J Thomson Avenue, Cambridge, CB23 0FB, U.K.
[email protected]
http://research.microsoft.com/∼cmbishop

† AEA Technology, Culham Laboratory,

Oxfordshire OX14 3DB
(Euratom/UKAEA Fusion Association)

Published in Neural Computation 7, No. 1 (1995) 206–217.

Abstract
In this paper we present results from the first use of neural networks for real-time control
of the high temperature plasma in a tokamak fusion experiment. The tokamak is currently
the principal experimental device for research into the magnetic confinement approach to
controlled fusion. In an effort to improve the energy confinement properties of the high
temperature plasma inside tokamaks, recent experiments have focussed on the use of non-
circular cross-sectional plasma shapes. However, the accurate generation of such plasmas
represents a demanding problem involving simultaneous control of several parameters on
a timescale as short as a few tens of microseconds. Application of neural networks to
this problem requires fast hardware, for which we have developed a fully parallel custom
implementation of a multilayer perceptron, based on a hybrid of digital and analogue
techniques.

1
1 Introduction
Fusion of the nuclei of hydrogen provides the energy source which powers the sun. It also
offers the possibility of a practically limitless terrestrial source of energy. However, the
harnessing of this power has proved to be a highly challenging problem. One of the most
promising approaches is based on magnetic confinement of a high temperature (107 – 108
Kelvin) plasma in a device called a tokamak (from the Russian for ‘toroidal magnetic
chamber’) as illustrated schematically in Figure 1. At these temperatures the highly

Figure 1: Schematic cross-section of a tokamak experiment showing the toroidal vacuum vessel
(outer D-shaped curve) and plasma (shown shaded). Also shown are the radial (R) and vertical
(Z) coordinates. To a good approximation, the tokamak can be regarded as axisymmetric about
the Z-axis, and so the plasma boundary can be described by its cross-sectional shape at one
particular toroidal location.

ionized plasma is an excellent electrical conductor, and can be confined and shaped by
strong magnetic fields. Early tokamaks had plasmas with circular cross-sections, for which
feedback control of the plasma position and shape is relatively straightforward. However,
recent tokamaks, such as the COMPASS experiment at Culham Laboratory, as well as
most next-generation tokamaks, are designed to produce plasmas whose cross-sections are
strongly non-circular. Figure 2 illustrates some of the plasma shapes which COMPASS
is designed to explore. These novel cross-sections provide substantially improved energy
confinement properties and thereby significantly enhance the performance of the tokamak.
Unlike circular cross-section plasmas, highly non-circular shapes are more difficult to
produce and to control accurately, since currents through several control coils must be
adjusted simultaneously. Furthermore, during a typical plasma pulse, the shape must
evolve, usually from some initial near-circular shape. Due to uncertainties in the current
and pressure distributions within the plasma, the desired accuracy for plasma control
can only be achieved by making real-time measurements of the position and shape of the
boundary, and using error feedback to adjust the currents in the control coils.
The physics of the plasma equilibrium is determined by force balance between the
thermal pressure of the plasma and the pressure of the magnetic field, and is relatively
well understood. Particular plasma configurations are described in terms of solutions of
the Grad-Shafranov equation (Shafranov, 1958), given by

∂2Ψ
!
∂ 1 ∂Ψ
R + = −RJ(Ψ, R) (1)
∂R R ∂R ∂Z 2

1
Figure 2: Cross-sections of the COMPASS vacuum vessel showing some examples of potential
plasma shapes. The solid curve is the boundary of the vacuum vessel, and the plasma is shown
by the shaded regions. Again, R and Z are the radial and vertical coordinates respectively in
units of meters.

where the coordinates R and Z are defined in Figure 1, the function Ψ is called the
poloidal flux function, and the plasma boundary corresponds to a surface of constant Ψ.
The function J(Ψ, R) specifies the plasma current density, and for the work reported here
we have chosen the following representation
( )
(1 − β)
J(Ψ, R) = b βR + (1 − Ψα1 )α2 (2)
R
which is motivated by plasma physics considerations. Here b is a constant, β controls the
ratio of plasma pressure to magnetic field energy density, and the parameters α1 and α2 are
numbers ≥ 1 which can be varied to generate a variety of current profiles. Fortunately, the
plasma configurations obtained by solution of the Grad-Shafranov equation are relatively
insensitive to the precise choice of representation for the function J(Ψ, R).
Due to the non-linear nature of the Grad-Shafranov equation, a general analytic so-
lution is not possible. However, for a given current density function J(Ψ, R), the Grad-

2
Shafranov equation can be solved by iterative numerical methods, with boundary con-
ditions determined by currents flowing in the external control coils which surround the
vacuum vessel. On the tokamak itself it is changes in these currents which are used to
alter the position and cross-sectional shape of the plasma. Numerical solution of the
Grad-Shafranov equation represents the standard technique for post-shot analysis of the
plasma, and is also the method used to generate the training dataset for the neural net-
work, as described in the next section. However, this approach is computationally very
intensive and is therefore unsuitable for feedback control purposes.
For real-time control it is necessary to have a fast (typically ≤ 50µsec.) determination
of the plasma boundary shape. This information can be extracted directly from a variety
of diagnostic systems, the most important being local magnetic measurements taken at a
number of points around the perimeter of the vacuum vessel. Most tokamaks have several
tens or hundreds of small pick up coils located at carefully optimized points around the
torus for this purpose. We shall represent these magnetic signals collectively as a vector
m.
The position and shape of the plasma boundary can be described in terms of a set of
geometrical parameters such as vertical position and elongation, which we collectively de-
note by yk . These parameters are illustrated in Figure fig:geometry, and will be discussed
in more detail in the next section. The basic problem which has to be addressed, there-

Z ka
q
Z0
a

R0 R

plasma
boundary

Figure 3: Schematic illustration of a cross-section of the toroidal vacuum vessel showing the
definitions of various coordinates and parameters. The elliptical curve denotes the plasma
boundary, whose center is at R = R0 , Z = Z0 . The parameter κ describes the elongation of
the plasma, and θ is called the poloidal angle. The triangularity δ (not shown) describes the
departure of the plasma boundary from a simple ellipse. (Values of κ = 1 and δ = 0 correspond
to a circular plasma boundary).

fore, is to find a representation for the (non-linear) mapping from the magnetic signals
m to the values of the geometrical parameters yk , which can be implemented in suitable
hardware for real-time control.
The conventional approach presently in use on many tokamaks involves approximating
the mapping between measured the magnetic signals and the geometrical parameters by a
single linear transformation. However, the intrinsic non-linearity of the mappings suggests

3
that a representation in terms of feedforward neural networks should give significantly
improved results (Lister and Schnurrenberger, 1991; Bishop et al., 1992; Lagin et al.,
1993). Figure 4 shows a block diagram of the control loop for the neural network approach
to tokamak equilibrium control.

Neural magnetic
Network signals m

plasma position Tokamak

and shape yk

desired position error Control

and shape ykd signals Amplifiers

Figure 4: Block diagram of the control loop used for real-time feedback control of plasma position
and shape. The neural network provides a fast non-linear mapping from the measured magnetic
signals onto the values of a set of geometrical parameters yk (illustrated in Figure 3) which
describe the position and shape of the plasma boundary. These parameters are compared with
their desired values ykd , and the resulting error signals used to correct the currents in a set of
feedback control coils using standard linear PD (proportional-differential) controllers.

2 Software Simulation Results

The dataset for training and testing the network was generated by numerical solution
of equation 1 using a free-boundary equilibrium code. This code contains a detailed
description of the COMPASS hardware configuration, and allows the boundary conditions
to be expressed directly in terms of currents in the control coils. The data base currently
consists of over 2,000 equilibria spanning the wide range of plasma positions and shapes
available in COMPASS. Each configuration takes several minutes to generate on a fast
workstation.
For a large class of equilibria, the plasma boundary can be reasonably well represented
in terms of a simple parameterization, governed by an angle-like variable θ, given by

R(θ) = R0 + a cos(θ + δ sin θ)

Z(θ) = Z0 + aκ sin θ (3)

where we have defined the following parameters

4
 R0 radial distance of the plasma center from the major axis of the torus,
Z0 vertical distance of the plasma center from the torus midplane,
a minor radius measured in the plane Z = Z0 ,
κ elongation,
δ triangularity.
Thus, for instance, if the triangularity parameter δ is zero, the boundary is described by an
ellipse with elongation κ. These parameters (except for the triangularity) are illustrated in
Figure 3. Each of the entries in the database has been fitted using the form in equation 3,
so that the equilibria are labelled with the appropriate values of the shape parameters.
On the COMPASS experiment, there are some 120 magnetic signals which could be
used to provide inputs to the network. Since each input could either be included or ex-
cluded, there are potentially 2120 ' 1036 possible sets of inputs which might be considered.
In order to find a computationally tractable procedure for selection of a suitable subset
of inputs, we have used forward sequential selection (Fukunaga, 1990), based on a simple
linear mapping (discussed shortly) to provide a selection criterion. Simulations aimed
at finding a network suitable for use in real-time control have so far concentrated on 16
inputs, since this is the number available from the initial hardware configuration.
It is important to note that the transformation from magnetic signals to flux surface
parameters involves an exact linear invariance. This follows from the fact that, if all of
the currents are scaled by a constant factor, then the magnetic fields will be scaled by
this factor, and the geometry of the plasma boundary will be unchanged. It is important
to take advantage of this prior knowledge and to build it into the network structure,
rather than force the network to learn it by example. We therefore normalize the vector
m of input signals to the network by dividing by a quantity proportional to the total
plasma current. A scaling of the magnetic signals by a common factor then leaves the
network inputs (and hence the network outputs) unchanged. Compared with learning
by example, this explicit use of prior knowledge brings three distinct advantages: (i) the
network exhibits exact, rather than approximate, invariance to rescaling of the currents;
(ii) the relative output accuracy can be maintained over a wide range of plasma current
(which typically varies from a few kA to a few 100kA during the plasma pulse); (iii)
the network training can be performed with a smaller dataset than would otherwise be
possible, which can be generated for just one value of total plasma current. Note that the
normalization has to be incorporated into the hardware implementation of the network,
as will be discussed in Section 3.
The results presented in this paper are based on a multilayer perceptron architecture
having a single layer of hidden units with ‘tanh’ activation functions, and linear output
units. Networks are trained by minimization of a sum-of-squares error using a standard
conjugate gradients optimization algorithm, and the number of hidden units is optimized
by measuring performance with respect to an independent test set. Results from the
neural network mapping are compared with those from the optimal linear mapping, that
is the single linear transformation which minimizes the same sum-of-squares error as is
used in the neural network training algorithm, as this represents the method currently
used on a number of present day tokamaks. This minimization can be expressed in terms
of a set of linear equations whose solution can be found efficiently and robustly using
the technique of singular value decomposition (Press et al., 1992). Note that the same
normalization of the inputs was used here as in the neural network case.
Initial results were obtained on networks having 3 output units, corresponding to the
values of vertical position Z0 , major radius R0 , and elongation κ; these being parameters

5
 which are of interest for real-time feedback control. The smallest normalized test set error
of 11.7 is obtained from the network having 16 hidden units. By comparison, the optimal
linear mapping gave a normalized test set error of 18.3. This represents a reduction in
error of about 30% in going from the linear mapping to the neural network. Such an
improvement, in the context of this application, is very significant.
For the experiments on real-time feedback control described in Section 4 the currently
available hardware only permitted networks having 4 hidden units, and so we consider the
results from this network in more detail. Figure 5 shows plots of the network predictions
for various parameters versus the corresponding values from the test set portion of the
database. Analogous plots for the optimal linear map predictions versus the database val-

4 1.2
R0 - 55.7 (cm) Z0 (cm) Elongation - 1.0
2
2
0.8
Linear Mapping

Linear Mapping

Linear Mapping
0
0
0.4
-2
-2 0
-4
-4 -2 0 2 4 -2 0 2 0 0.4 0.8 1.2
Database Database Database

4 1.2
R0 - 55.7 (cm) Z0 (cm) Elongation - 1.0
2
2
0.8
Neural Network

Neural Network

Neural Network

0
0
0.4
-2
-2 0
-4
-4 -2 0 2 4 -2 0 2 0 0.4 0.8 1.2
Database Database Database

Figure 5: Plots of the values from the test set versus the values predicted by the linear mapping
for the 3 equilibrium parameters, together with the corresponding plots for neural network with
4 hidden units.

ues are also shown. Comparison of the corresponding figures shows the poorer predictive
capability of the linear approach, even for this sub-optimal network topology.

3 Hardware Implementation
The hardware implementation of the neural network must have a bandwidth of ≥ 20
kHz in order to cope with the fast timescales of the plasma evolution. It must also have

6
an output precision of at least 8 bits in order to ensure that the final accuracy which
is attainable will not be limited by the hardware system. We have chosen to develop
a fully parallel custom implementation of the multilayer perceptron, based on analogue
signal paths with digitally stored synaptic weights (Bishop et al., 1993a). A VME-based
modular construction has been chosen as this allows flexibility in changing the network
architecture, ease of loading network weights, and simplicity of data acquisition. Three
separate types of card have been developed as follows:
• Combined 16-input buffer and signal normalizer
This provides an analogue hardware implementation of the input normalization de-
scribed earlier. For future flexibility this makes use of an EPROM (erasable pro-
grammable read-only memory) to provide independent scaling of groups of 8 inputs
by an arbitrary function of an external reference signal. In the present application
the reference signal is taken to be the plasma current (determined by a magnetic
pick-up loop called a Rogowski coil) and the function is chosen to be a simple inverse
proportionality.
• 16 × 4 matrix multiplier
The synaptic weights are produced using 12 bit frequency-compensated multiplying
DACs (digital to analogue converters) which can be configured to allow 4-quadrant
multiplication of analogue signals by a digitally stored number. The weights are
obtained as a 12-bit 2’s-complement representation from the VME backplane. Note
that the DACs are being used here as digitally controlled attenuators, and not in
their usual role of converting digital signals into analogue signals. Synaptic weights
are downloaded (prior to the plasma pulse) via the VME backplane from a central
control computer, using an addressing technique to label the individual weights.
• 4-channel sigmoid module
There are many ways to produce a sigmoidal non-linearity, and we have opted for a
solution using two transistors configured as a long-tailed-pair, to generate a ‘tanh’
sigmoidal transfer characteristic. The principal drawback of such an approach is
the strong temperature sensitivity due to the appearance of temperature in the
denominator of the exponential transistor transfer characteristic. An elegant solution
to this problem has been found by exploiting a chip containing 5 transistors in
close thermal contact. Two of the transistors form the long-tailed pair, one of the
transistors is used as a heat source, and the remaining two transistors are used to
measure temperature. External circuitry provides active thermal feedback control,
and stability to changes in ambient temperature over the range 0◦ C to 50◦ C is found
to be well within the acceptable range. A separate 12-bit DAC system, identical to
the ones used on the matrix multiplier cards but with a fixed DC input, is used to
provide a bias for each sigmoid.
The complete network is constructed by mounting the appropriate combination of
cards in a VME rack and configuring the network topology using front panel intercon-
nections. The system includes extensive diagnostics, allowing voltages at all key points
within the network to be monitored as a function of time via a series of multiplexed output
channels.

7
4 Results from Real-time Feedback Control
Figure 6 shows the first results obtained from real-time control of the plasma in the
COMPASS tokamak using neural networks. The evolution of the plasma elongation,

1.8
shot 9576
elongation
1.4

1.0

0.0 0.1 0.2

time (sec.)
Figure 6: Plot of the plasma elongation κ as a function of time during shot no. 9576 on the
COMPASS tokamak, during which the elongation was being controlled in real-time by the neural
network. The solid curve shows the value of elongation given by one of the network outputs.
The dashed curve shows the post-shot reconstruction of the elongation obtained from a simple
‘filament’ code, which gives relatively rapid post-shot plasma shape reconstruction but with
limited accuracy. The circles denote reconstructions obtained from the full equilibrium code,
which gives closer agreement with the network predictions.

under the control of the neural network, is plotted as a function of time during a plasma
pulse. Here the desired elongation has been preprogrammed to follow a series of steps
as a function of time. The remaining 2 network outputs (radial position R0 and vertical
position Z0 ) were digitized for post-shot diagnosis, but were not used for real-time control.
The graph clearly shows the network responding and generating the required elongation
signal in close agreement with the reconstructed values. The typical residual error is
of order 0.07 on elongation values up to around 1.5. Part of this error is attributable
to residual offset in the integrators used to extract magnetic field information from the
pick-up coils, and this is currently being corrected through modifications to the integrator
design. An additional contribution to the error arises from the restricted number of hidden
units available with the initial hardware configuration. While these results represent the
first obtained using closed loop control, it is clear from earlier software modelling of larger
network architectures (such as 32–16–4) that residual errors of order a few % should be
attainable. The implementation of such larger networks is being pursued, following the
successes with the smaller system.
Neural networks have already been used with great success for fast interpretation
of the data from tokamak plasma diagnostics to determine the spatial and temporal
profiles of quantities such as temperature and density (Bishop et al., 1993b; Bishop et al.,
1993c; Bartlett and Bishop, 1993). There is currently considerable interest in extending

8
these techniques to allow real-time feedback control of the profiles to give more complete
determination of the plasma configuration than is possible by boundary shape control
alone. For such applications, neural networks appear to offer one of the most promising
approaches.

Acknowledgements
We would like to thank Peter Cox, Jo Lister and Colin Roach for many useful discussions
and technical contributions. This work was partially supported by the UK Department
of Trade and Industry.

References
Bishop C M, Cox P, Haynes P S, Roach C M, Smith M E U, Todd T N and Trotman
D L, 1992. A neural network approach to tokamak equilibrium control. In Neural
Network Applications, Ed. J G Taylor, Springer Verlag, 114–128.
Bishop C M, Haynes P S, Roach C M, Smith M E U, Todd T N, and Trotman D L.
1993a. Hardware implementation of a neural network for plasma position control in
COMPASS-D. In Proceedings of the 17th. Symposium on Fusion Technology, Rome,
Italy. 2 997–1001.
Bishop C M, Roach C M and von Hellerman M. 1993b. Automatic analysis of JET charge
exchange spectra using neural networks. Plasma Physics and Controlled Fusion 35
765–773.
Bishop C M, Strachan I G D, O’Rourke J, Maddison G and Thomas P R. 1993c. Recon-
struction of tokamak density profiles using feedforward networks. Neural Computing
and Applications 1 4 – 16.
Bartlett D V, and Bishop C M. 1993. Development of neural network techniques for the
analysis of JET ECE data. In Proceedings of the 8th. International Workshop on
ECE and ECRH (EC8, 1992).
Fukunaga K. 1990. Statistical Pattern Recognition, 2nd. Edition, Academic Press, San
Diego.
Lagin L, Bell R, Davis S, Eck T, Jardin S, Kessel C, Mcenerney J, Okabayashi M, Popyack
J and Sauthoff N. 1993. Application of neural networks for real-time calculations of
plasma equilibrium parameters for PBX-M, In Proceedings of the 17th. Symposium
on Fusion Technology, Rome, Italy. 2 1057–1061.
Lister J B and Schnurrenberger H. 1991. Fast non-linear extraction of plasma parameters
using a neural network mapping. Nuclear Fusion. 31, 1291–1300.
Press W H, Flannery B P, Teukolsky S A and Vetterling W T. 1992. Numerical Recipes
in C: The Art of Scientific Computing 2nd. Edition, Cambridge University Press.
Shafranov V D. 1958. On magnetohydrodynamical equilibrium configurations. Sov.
Phys. JETP, 8 710.