Page 1
SLAC-PUB-8889
September 2001
physics/0108063
Room Temperature Accelerator Structures for Linear Colliders
Work supported by Department of Energy contract DE–AC03–76SF00515.
Presented at the IEEE Particle Accelerator Conference (PAC2001),
6/18/2001—6/22/2001, Chicago, IL, USA
R. H. Miller et al.
Stanford Linear Accelerator Center, Stanford University, Stanford, CA 94309
Page 2
ROOM TEMPERATURE ACCELERATOR STRUCTURES FOR LINEAR
COLLIDERS*
R.H. Miller, R.M. Jones, C. Adolphsen, G. Bowden, V. Dolgashev, N. Kroll Z. Li, R. Loewen, C.
Ng, C. Pearson, T. Raubenheimer R. Ruth, S. Tantawi, J.W. Wang, SLAC, Menlo Park, CA, USA
Abstract
Early tests of short low group velocity and standing
wave structures indicated the viability of operating X-
band linacs with accelerating gradients in excess of 100
MeV/m. Conventional scaling of traveling wave traveling
wave linacs with frequency scales the cell dimensions
with λ. Because Q scales as λ1/2, the length of the
structures scale not linearly but as λ3/2 in order to preserve
the attenuation through each structure. For NLC we
chose not to follow this scaling from the SLAC S-band
linac to its fourth harmonic at X-band. We wanted to
increase the length of the structures to reduce the number
of couplers and waveguide drives which can be a
significant part of the cost of a microwave linac.
Furthermore, scaling the iris size of the disk-loaded
structures gave unacceptably high short range dipole
wakefields. Consequently, we chose to go up a factor of
about 5 in average group velocity and length of the
structures, which increases the power fed to each structure
by the same factor and decreases the short range dipole
wakes by a similar factor. Unfortunately, these longer
(1.8 m) structures have not performed nearly as well in
high gradient tests as the short structures. We believe we
have at least a partial understanding of the reason and will
discuss it below. We are now studying two types of short
structures with large apertures with moderately good
efficiency including: 1) traveling wave structures with the
group velocity lowered by going to large phase advance
per period with bulges on the iris, 2) π mode standing
wave structures.
1 HIGH GRADIENT RF BREAKDOWN
The high gradient RF breakdown testing is reported in
detail by Adolphsen [1]. There are several interesting and
some surprising results that affect the choice of structure
design, which we will discuss here. The first is that as
suggested by Adolphsen the viable operating gradient
appears to vary almost linearly with the inverse of the of
the group velocity. A related observation is that the
structures process very rapidly with a relatively small
number of arcs up to a gradient that also varies slightly
less than linearly with the inverse of the group velocity.
Above that gradient the arcing rate increases dramatically
and progress to higher gradients is very, very slow.
The second and perhaps the most surprising result
occurred during the simultaneous testing of a 105cm long
structure and a 20cm long structure driven by the same
klystron through a 3 dB coupler so that the drive levels
and history would be identical. Both structures were
designed to be approximately constant gradient, but
precisely constant peak surface field on all disks. Both
had an initial group velocity of 5% of the velocity of light.
The short structure was identical to the first 20 cm of the
long structure. One might have expected the 105cm
structure to have roughly 5 times as many arcs as the
20cm structure. Instead, the two structures had equal
numbers of arcs at all power levels within the statistical
variation, except during the very early processing. This is
less surprising in view of the fact that the vast majority of
the arcs in the long structure occur in the first 20cm.
The third interesting fact emerging from the high
gradient testing is observed when a structure has been
processed up to some level with a short pulse and the
pulse length is increased significantly. The rate of
breakdowns increases dramatically with the arcs
distributed uniformly in time within the pulse, not
concentrated in the added portion of the pulse.
The fourth interesting result occurred in an experiment
studying high electric field gradients in rectangular
waveguide, Dolgashev [2]. The large dimension of the
waveguide had been reduced to lower the group velocity
to about 0.18c in order to raise the field strengths that
could be reached with available power. The striking
observation was that when the pulse length was less than
400ns and the peak surface gradient was 80 MV/m the
arcs never degraded the high gradient performance of the
waveguide. When the pulse length was more than 500 ns
the arcs frequently degraded the high gradient
performance. The degradation observed was a higher rate
of arcing, or the inability to reach 80MV/m and higher
xray levels on pulses where no arcing occurs.
These four observations suggest that it may be
important to consider the energy deposited at the site of
an arc when an arc occurs for two reasons. First, it may
alter the microwave parameters of the structure by causing
a tiny, deposited-energy-dependent change in the resonant
frequency of the cell in which the arc occurs and thus
change the phase advance and the match of the structure.
Secondly, there probably is a deposited energy threshold
above which the high gradient performance of the
____________________________________________
*Work supported by the U.S. DOE, Contract DE-AC03-76SF00515.
Page 3
structure is degraded. Above this threshold an arc is
likely to cause surface damage which causes successive
arcs to occur at or in the vicinity of the original site. For
many years people have observed that RF processing is
not always monotonic; that sometimes an arc causes a
setback, lowering the RF power which the device being
processed will accept. It has also been realized that it was
advantageous to high power process with a short pulse
until the desired gradients are reached, and then slowly
increase the pulse length. The energy deposited in an arc
in a travelling wave structure should vary linearly with the
incident power, the pulse length, and the group velocity.
It may also depend on other parameters such as the
frequency, phase advance per cell and geometric factors
such as the gap across which the breakdown occurs. The
damage threshold almost surely depends on the physical
properties such as melting point of the material from
which the structure is made. Despite our ignorance of
these issues, it still may be useful to define a damage
parameter, D.
D = PTvg/c (1)
P is the input power, T is the pulse length, vg is the group
velocity and c is the velocity of light. To get an idea of
what may reasonable we can look at SLAC (ignoring any
frequency and geometry dependence). The highest power
operation of the standard SLAC S-band sections with a
rectangular pulse (unSLEDded) occurred in the injector
where several sections ran routinely with 25 MW 2.5 µs
long pulses. The initial group velocity of these constant
gradient sections was .02c. With these parameters the
damage parameter is 1.2 Joules. It is important to note
that only a minuscule fraction of this can be dissipated at
the site of the arc. In the high gradient testing the X-band
1.8 meter Damped Detuned Structures (DDS) processed
very easily and essentially monotonically up to about 35
MeV/m with a 44 MW, 240ns pulse. The initial group
velocity is .12c, giving a damage parameter of 1.3 J,
suggesting that this may be a useful parameter. The four
observations from the high gradient testing reported above
all suggest that for gradients above some damage
threshold most of the arcs result from damage caused by
previous arcs rather than from the initial existing defects
such as inclusions and microscopic points in our copper
structures. We have also observed both with the SLAC S-
band structures and with the NLC X-band structures that
it is possible to have arc damage from which it is
impossible to recover. That is that a reasonable amount of
high gradient processing cannot recover the gradient at
which the structure had been operating. The conservative
course of action may be to design to run below the arc
damage threshold for which our damage parameter D may
(or may not) be a useful indicator. If D were a precise
measure of the limit for monotonic processing then the
gradient at which monotonic processing ends would scale
linearly with the 1/vg. We observe a less than linear
scaling, but the number of samples is very small. Perhaps
vg/c in the damage parameter needs an experimentally
determined exponent which is less than unity.
2 TRAVELING WAVE STRUCTURES
The present design for the unloaded gradient for NLC is
72 MeV/m. To achieve this with a damage parameter of
the order of 1 Joule will require an initial group velocity
in our approximately constant gradient structures of
between .03c and .05c. We are presently testing
structures at each of these values. The tested structures
would not be satisfactory for NLC because the apertures
are too small causing excessive dipole wakefields. Z. Li
[3] has designed structures for each of these group
velocities using 150o phase advance per cell and thicker
disks to achieve these lower group velocities with the
same average iris diameter as in the vg = 0.12c structure.
The 0.05c structure is 90cm long, while the 0.03c
structure is 60 cm long. R.M. Jones [4] is studying
detuning these and damping them using either manifold
damping or local damping to reduce the long-range dipole
wakes by a factor greater than 100. The initial results
look quite promising. The wakefield for the 0.05c
structure with 10% Gaussian detuning and manifold
damping is shown in Fig. 1. It is difficult to go below an
initial group velocity of .03c without reducing the average
iris diameter, which we don't want to do because of the
short-range dipole wakes. We are uncomfortable about
100
Wake ?V?pC?mm?m?
2468
????
s ??????? ?
m ?
0.001
0.01
0.1
1
10
Figure 1: Wakefield for two 90cm long manifold-damped
10% detuned structures with .05c initial vg.
going to phase advances larger than 150o per period
because of the reduced bandwidth, and thicker disks hurt
the shunt impedance.
3 STANDING WAVE STRUCTURES
There is some argument for designing the NLC
structure to operate at unloaded gradients as high as 100
MeV/m, to accommodate an energy upgrade above 1 TeV
in the same tunnel. We think this forces us to consider
standing wave structures. Standing wave structures have
several advantages over travelling for high gradient
operation. The first is that for a given gradient the input
power required scales roughly as length, and stranding
wave structures can be made arbitrarily short without
sacrificing efficiency. Secondly, because a standing wave
structure is a high Q resonant cavity the reflection
coefficient goes very, very close to unity almost instantly
when loaded by an arc. In this way a standing wave
Page 4
cavity is more self-protecting than a travelling wave
structure. We are studying 20 and 30cm π-mode
structures with 15 and 23 cells, respectively. We think
this range is a reasonable compromise between tolerances
and the costs associated with a large number of short
structures. Figure 2 is computer cutaway of our first 15-
cell π mode cavity. Two of these are currently in high
power testing. We propose to braze 6 to 10 of these short
structures together to form a single precision aligned
assembly. Each multicell cavity will have an odd number
of cells and be driven through the center cell. Center
driving relaxes the tolerances by a factor of 4.
Figure 2: Computer cutaway of 15-cell π-mode cavity.
The standing wave cavities in each half of the assembly
will be driven by a single oversize rectangular waveguide
with a directional coupler splitting off the appropriate
power for each cavity. When this is done right all the
power reflected from the cavities goes into dummy loads
in the approximation that all the cavities have the same
coupling and the same resonant frequency. Thus, in this
approximation the assembly looks like a matched load to
the RF source. Of course when a cavity arcs it upsets the
match, but since each cavity is getting a small fraction of
the total power the reflected power goes like the small
fraction squared.
3.1 Dipole Wakefield Reduction
The same general methods, namely detuning and
damping, are available for reducing dipole wakefields in
standing wave cavities as were used in the traveling wave
structures. However, because the cavities are so short the
implementation will be different. Tapering the iris
diameters with compensating changes in the cell
diameters still appears to be the best way to tune the first
dipole band while keeping fundamental π mode frequency
constant. By decreasing the thickness of the disks as the
iris diameters decrease we find we can reduce the spread
in cell to cell coupling of the fundamental. We can get an
8% detuning of the first dipole band with the fundamental
mode cell to cell coupling varying from 5.2% at the large
iris cells down to 2.2% at small iris end. We intend to
vary the iris size and therefore the dipole frequencies
monotonically from one end to the other of a 6 or 9 cavity
assembly having a total of about 135 cells. Fig. 3 shows
the amplitude and phase from an equivalent circuit for
each cell of a 15 cell cavity in which the coupling varies
from 3.5% to 2.2% as it might in the last cavity at the high
dipole frequency end of a 6 cavity assembly. The phase
shifts, which are caused by the power flow, can be
compensated for by adjusting the length of the cells to
keep a velocity of light beam on the crest in each cell.
Fig. 4 shows the dispersion curve for each cell (as if it
were in a periodic structure). The cell fundamental mode
resonant frequencies have been tuned to achieve a flat
field, which requires that the π mode frequencies for all
the interior cells coincide at about 11.4 GHz. The cells at
each end, because they are full cells rather than a half-
cells, must be tuned so the π/2 modes are at 11.4 GHz.
Because the cavities are so short, we tried what might
be called end damping. We put a low Q (10) dipole mode
cavity in the drift tube between each pair of 23 cell
cavities and at each end of the full assembly. We have a
preliminary simulation of this and the wakefield is
presented in Fig. 4. It appears to be promising, but this is
certainly neither a full nor optimized design. We have a
concept for the lossy dipole cavities but they have not
been designed.
Figure 3: Phase and amplitude of the fundamental mode
in the 15 cells of a dipole detuned π mode cavity.
Figure 4: Dispersion curves for the fundamental mode of
the 15 cells of a π-mode cavity with dipole mode detuning
Figure 5: Wakefield for six 23 cell π-mode cavities in a
monotonically detuned assembly with a low Q dipole
cavity every 23 cells.
4 REFERENCES
[1] C. Adolphsen, Paper ROAA003 this conference
Page 5
[2] V. Dolgashev, Paper FPAH057 this conference
[3] Z. Li, Paper FPAH061 this conference
[4] R.M. Jones, Papers FPAH058 & MPPH068 this conf.
Download full-text