nature photonics

Article https://doi.org/10.1038/s41566-024-01419-w

Experimental demonstration of attosecond

pump–probe spectroscopy with an X-ray
free-electron laser

Received: 9 May 2023 Zhaoheng Guo 1,2,18, Taran Driver 1,3,18, Sandra Beauvarlet 3,4, David Cesar1,
Joseph Duris 1, Paris L. Franz1,2, Oliver Alexander 5, Dorian Bohler1,
Accepted: 8 March 2024
Christoph Bostedt6,7, Vitali Averbukh 5, Xinxin Cheng 1,8, Louis F. DiMauro9,
Published online: xx xx xxxx Gilles Doumy 10, Ruaridh Forbes1,3,8, Oliver Gessner 11, James M. Glownia1,8,
Erik Isele1,2,3, Andrei Kamalov1,8, Kirk A. Larsen 1,3, Siqi Li 1, Xiang Li 1,8,
Check for updates
Ming-Fu Lin 1,8, Gregory A. McCracken9, Razib Obaid1,8, Jordan T. O’Neal1,3,12,
River R. Robles 1,2, Daniel Rolles 13, Marco Ruberti 5, Artem Rudenko 13,
Daniel S. Slaughter 11, Nicholas S. Sudar1, Emily Thierstein1,2,3, Daniel Tuthill 9,
Kiyoshi Ueda 14, Enliang Wang 13, Anna L. Wang1,2,3, Jun Wang 1,2,3,
Thorsten Weber 11, Thomas J. A. Wolf 1,3, Linda Young 10,15, Zhen Zhang 1,
Philip H. Bucksbaum1,2,3,12, Jon P. Marangos 5, Matthias F. Kling 1,2,16,
Zhirong Huang 1,2,16, Peter Walter 1,8, Ludger Inhester 17, Nora Berrah 4,
James P. Cryan 1,3 & Agostino Marinelli 1,3,16

Pump–probe experiments with subfemtosecond resolution are the

key to understanding electronic dynamics in quantum systems. Here
we demonstrate the generation and control of subfemtosecond pulse
pairs from a two-colour X-ray free-electron laser. By measuring the
delay between the two pulses with an angular streaking diagnostic, we
characterize the group velocity of the X-ray free-electron laser and show
control of the pulse delay down to 270 as. We confirm the application
of this technique to a pump–probe measurement in core-ionized
para-aminophenol. These results reveal the ability to perform pump–probe
experiments with subfemtosecond resolution and atomic site specificity.

Pump–probe spectroscopy is the primary workhorse for ultrafast exper- The ultrafast dynamics of electrons are integral to many chemical
iments1. This form of stroboscopic measurement requires an initial light and physical processes. For example, electron–light interactions, and the
pulse, which excites, or ‘pumps’, the system into a non-equilibrium state. resulting electronic motion, are the fundamental mechanisms by which
A subsequent light pulse is then used to probe the time evolution of the light is absorbed in matter, a basic building block of solar energy technol-
system. When either the pump or probe pulse is tuned to X-ray wave- ogy and photosynthesis7. The timescale for this electronic motion is set
lengths, the interaction of the pulse is strongest with highly localized by the electronvolt-scale binding energy of the most chemically relevant
core-level electrons, resulting in an atomic-site-specific interrogation valence electrons, and is in the range of a few femtoseconds to subfem-
of electronic densities2. This technique is routinely applied to follow tosecond. Therefore, pump–probe experiments that are capable of
photo-induced chemical and material transitions on femtosecond accessing coherent electronic motion require the use of subfemtosecond
timescales3–6. In extending the pump–probe technique to subfemto- pulse pairs with sufficient intensity to enable multi-photon interactions.
second timescales we can probe the motion of valence electrons in Isolated light pulses with subfemtosecond pulse duration were
quantum systems. initially demonstrated as early as 20018, by exploiting the process

A full list of affiliations appears at the end of the paper. e-mail: [email protected]; [email protected]

Nature Photonics
Article https://doi.org/10.1038/s41566-024-01419-w

a I attosecond pump–probe spectroscopy requires a detailed experimen-

c-VM
tor tal understanding of time-dependent FEL physics. The relative slippage
ndula
nd u
Seco ω pulse between the X-rays and the electrons during the amplification process
2
modifies the pump–probe delay at the few-femtosecond level27; there-
lator ser
undu e red la
fore, the group velocity of the FEL must be characterized to control the
Energy

First p u ls Infra
ω pump–probe delay with subfemtosecond accuracy.
e
e ican
Tim y ch In this work, we demonstrate the generation, diagnosis and con-
Dela
trol of two-colour X-ray pulse pairs with attosecond-scale delays and
beam duration at the Linac Coherent Light Source30. The delay between
tron
Elec the two pulses can be controlled with an accuracy of hundreds of
b c 1
attoseconds, and their relative timing jitter is estimated to be less
than 270 as. We also report the application of this novel technique

Electron yield (a.u.)

2 ω
py (a.u.) 2ω to a time-dependent pump–probe measurement of core-ionization
dynamics in para-aminophenol molecules. Our results extend
1 the attosecond-pump/attosecond-probe capability to the X-ray
regime, enabling the study of electronic dynamics with the relevant
subfemtosecond time resolution and with atomic site specificity. The
0 0
0 1 2 use of two different colours enables us to separate the spectroscopic
px (a.u.) signatures of the pump and probe pulses in our data. More generally,
two-colour pulse pairs are important for enabling a broad suite of
Fig. 1 | Experimental configuration for the two-colour attosecond XFEL.
nonlinear spectroscopies31,32, where dynamics are initiated and inter-
a, A modulated electron bunch with a high-current spike lases in the first set of
magnetic undulators (red stripes) with a resonant photon energy of 370 eV.
rogated using X-ray interactions with different electronic orbitals and
A magnetic chicane can be used to delay the electron bunch with respect to this atomic sites.
first pulse before lasing in the second undulator section (blue stripes), with a
resonant photon energy that is double that of the initial undulator, 740 eV. Demonstration of two-colour attosecond pulse
b, X-ray pulses ionize the target sample (CF4) and the photoelectrons are pairs
collected using a c-VMI spectrometer (see text). Atom colours: C, black; F, green. Figure 1a shows a schematic representation of the experiment. An
c, Measured two-dimensional (2D) projection of the photoelectron momentum enhanced self-amplified spontaneous emission (ESASE)33 method
distribution recorded via c-VMI in the absence of the circularly polarized is used to produce a short current spike by shaping the photocath-
streaking field, where px is chosen to lie along the X-ray polarization direction. ode laser pulse17. This high-current spike is capable of generating sub-
The black line shows the electron momentum distribution retrieved from an femtosecond X-ray pulses. The soft X-ray undulator beamline of the
inverse Abel transform of the data. LCLS-II (the upgraded LCLS XFEL) is composed of two sections, each
tuned to a different resonant frequency (ω and 2ω). Both undulator
sections are composed of individually tunable modules (see Meth-
of strong-field-driven high harmonic generation. The majority of ods). The shaped electron bunch emits a single-spike soft X-ray pulse
time-resolved experiments reaching few- or subfemtosecond resolu- from each half of the undulator, with a typical duration of between 200
tion have relied on a strong infrared field as either the pump or probe9,10. and 700 as depending on the exact experimental conditions16. The
This highly non-perturbative interaction is challenging to model and relative delay is controlled using a magnetic chicane. The minimum
does not lend itself to quantitative comparison with theory. High har- delay achievable with a split undulator is usually limited to a few fem-
monic generation sources have generated attosecond-scale pulse toseconds due to the relative slippage of the first pulse with respect
pairs in the ultraviolet domain11–13, which can be used for time-resolved to the electrons27. To access subfemtosecond delays, the second pulse
measurements14,15. Extending attosecond-pump/attosecond-probe must be tuned to a harmonic of the first. In this case, the first pulse
capabilities to the soft X-ray domain will enable ångström-scale prob- introduces strong harmonic microbunching in the electron beam,
ing of the electronic density, due to the atomic site specificity of X-ray which seeds the emission of the second pulse. Therefore, the second
spectroscopies. pulse can saturate in a short undulator length, minimizing the relative
X-ray free-electron lasers (XFELs), which can produce X-ray pulses delay between the two pulses. In our demonstration, we upconvert to
with sufficient intensity to drive nonlinear interactions, can also the second harmonic, but, in principle, this scheme applies to higher
generate isolated attosecond pulses16,17 and pulse trains18,19. Whereas harmonics as well.
these pulses often have properties that can vary from shot to shot, The delay between the two pulses is measured using attosecond
this has been shown to have little impact on the electronic dynam- angular streaking16,34,35, as illustrated in Fig. 1b. The co-propagating
ics induced by the FEL pulses20 or on the resulting spectroscopic ω/2ω pulse pairs are incident on a molecular target (tetrafluorometh-
observables21. ane, CF4) and drive X-ray ionization in the presence of a circularly
Pump–probe techniques that combine an external laser pulse with polarized, 1.3 μm laser field. We record the transverse momentum
attosecond XFELs present serious challenges in the attosecond regime. distribution (with respect to the propagation axis of the X-rays) of
The XFEL arrival-time jitter is dominated by the radiofrequency phase the ionized electrons using a co-axial velocity map imaging (c-VMI)
jitter, and is of the order of tens to hundreds of femtoseconds22,23. To spectrometer35,36. The effect of the laser field is to shift (or streak) the
mitigate this effect, the arrival time of the X-rays with respect to an measured electron momentum distribution in the direction opposite
external laser can be reliably measured on a shot-to-shot basis with the laser vector potential at the time of X-ray ionization34,37:
femtosecond accuracy24; subfemtosecond time sorting has been
reported25,26 but with relatively long laser and X-ray pulses (tens of p(t → ∞) = p0 + e A(t0 ), (1)
femtoseconds and few femtoseconds, respectively). The timing stabil-
ity can be greatly improved using multi-pulse FEL techniques. In this where p(t → ∞) is the momentum of the electron measured at the
case, both pump and probe pulses are emitted by the same electron detector, p0 is the momentum of the electron in the absence of the
t0
beam, and the pump–probe delay is independent of the arrival-time laser field, A(t0 ) = − ∫−∞ ℰ L (t′ )dt′ is the vector potential of the laser
jitter of the electrons27–29. The application of two-colour techniques to field ℰ L at the time of ionization t0, and e is the electron charge, taken

Nature Photonics
Article https://doi.org/10.1038/s41566-024-01419-w

a ∆τ = 0.40 fs ∆τ = 0.50 fs ∆τ = 0.65 fs ∆τ = 1.12 fs

0 0.5 1.0 t (fs) 0 0.5 1.0 t (fs) 0 0.5 1.0 t (fs) 0 0.5 1.0 t (fs)

b 3 3 3 3 1

∆ electron yield (a.u.)

2 2 2 2

1 1 1 1
py (a.u.)

py (a.u.)

py (a.u.)

py (a.u.)
∆ф ∆ф ∆ф ∆ф
0 0 0 0 0
–1 –1 –1 –1

–2 –2 –2 –2

–3 –3 –3 –3 –1
–3 –2 –1 0 1 2 3 –3 –2 –1 0 1 2 3 –3 –2 –1 0 1 2 3 –3 –2 –1 0 1 2 3
px (a.u.) px (a.u.) px (a.u.) px (a.u.)

c 0.1

Correlation coefficient
300 300 300 300
θ2ω (deg)

θ2ω (deg)

θ2ω (deg)

θ2ω (deg)
200 200 200 200
0

100 100 100 100

∆ф ∆ф ∆ф ∆ф
0 0 0 0 –0.1
0 100 200 300 0 100 200 300 0 100 200 300 0 100 200 300
θω (deg) θω (deg) θω (deg) θω (deg)

Fig. 2 | Attosecond pump–probe delay determination using angular shift of the photoelectrons ionized by the ω (2ω) pulse. c, 2D maps of the
streaking. a, The time delay Δτ between the ω (red) and 2ω (blue) pulses. b, 2D correlation coefficients between the streaked photoelectron features (described
projections of the photoelectron momentum distribution for different values of in the text) as a function of the detector angles θω and θ2ω for different values of
Δτ. We have subtracted a background from each panel collected in the absence of Δτ. The difference in streaking angles is related to the delay via Δϕ = 2π × Δτ/TL.
the streaking laser. Red (blue) arrows represent the direction of the momentum

as e = −1 in atomic units. The streaking interaction therefore establishes This approach offers control of the delay up to roughly 0.7 fs. The delay
a mapping from angle to time: the direction of the transverse momen- can be increased further using a magnetic chicane to increase the path
tum shift determines the time of arrival of the X-ray pulses within a length of the electron bunch, and thus increase the delay between the
laser period. pulses, as shown in the measured data point labelled ‘Exp: chicane on’.
We measure the delay by recording the difference in the angular
shift of the photoelectrons produced by each pulse. The angle, Δϕ, Discussion and interpretation of delay
between the momentum shift of the two photoelectron features relates measurements
to the delay Δτ between the two pulses according to the relationship: The dependence of the temporal separation of the pulses on the num-
ber of undulator modules is due to a mismatch in the group velocity
Δϕ
Δτ = × TL , (2) between the resonant 2ω pulse and the non-resonant ω pulse. Figure 3d
2π
shows the results of a start-to-end simulation of the experiment. The
where TL is the period of the streaking laser (TL = 4.3 fs in our experi- radiation power of the probe is plotted as a function of the undulator
ment). The two photoelectron features are shown in Fig. 1c: 68 eV elec- length and the bunch coordinate. The mismatch in the group veloci-
trons from ionization from the carbon K shell by the ω (370 eV) pulse ties between the ω and 2ω pulses can be seen clearly in the divergence
and 45 eV electrons from ionization from the fluorine K shell by the between the centre of mass of the ω pulse (dashed white line) and that
2ω (740 eV) pulse. of the 2ω pulse (dashed black line).
Figure 2 shows the differential transverse momentum dis- To interpret our experimental results, we parameterize the FEL
tribution (with and without the infrared laser) for different ω/2ω group velocity as follows:
delays. To extract the average delay between ω/2ω pulses we use a
λr
correlation-based method that exploits the large shot-to-shot variation vg = vb (1 + α ), (3)
λu
in the laser/X-ray arrival time in our experiment (~500 fs (ref. 22)). This
method is benchmarked in Supplementary Section 3.2 by comparison where vg and vb are the group velocity and the beam velocity (which are
with a single-shot analysis for a subset of delays. both close to the speed of light c), λr is the radiation wavelength, λu is
Figure 3a shows the measured average delays between the ω/2ω the period of the magnetic undulator and α is a positive number, typi-
pulse pairs for different undulator and chicane configurations. From cally between 0 and 1, which accounts for the slippage effects during
the correlation analysis it is also possible to infer an upper limit for the the lasing process (a value of α = 1 corresponds to propagation at the
arrival-time jitter between the two pulses, which we estimate to be 270 as speed of light, or vg = c). In the one-dimensional (1D) limit, an FEL pulse
(see Supplementary Section 3.1.2). To control the delay in the subfem- in the exponential gain regime has a group velocity given by the electron
tosecond regime we exploit the relative slippage of the two pulses and beam velocity plus one-third of the slippage rate38,39, corresponding to
vary the number of undulator modules used in the second stage. The α = 1/3. For a pulse generated by a high-current ESASE spike, the radia-
delay between ω/2ω pulses increases roughly linearly in the second stage tion remains close to the current spike because the gain in the high-
at a rate of 130 ± 19 as per undulator module (87 periods per module). current region is larger than in the rest of the bunch, an effective

Nature Photonics
Article https://doi.org/10.1038/s41566-024-01419-w

a c
1,600 Linear fitting 1
Sim: chicane off 0
Sim: 1 fs chicane delay
1,200 Sim: 0.6 fs chicane delay
Exp: chicane off

Delay (as)
Exp: chicane on
800

400

Spectral intensity (a.u.)

0
1 2 3 4
Number of undulator modules

b
25
Chicane off
Chicane on
20
2ω pulse energy (µJ)

15

10

1
0
0
1 2 3 4
365 370 375 735 740 745
Number of undulator modules
Photon energy (eV)

d Distance in the undulator (m)

0 3 6 9 12
1.0
ω pulse centre of mass

Power normalised to peak

2.0
2ω pulse centre of mass
Bunch coordinate (fs)

0.8

1.5
0.6
1.0
0.4
0.5
0.2
0
0
0 1 2 3 4
Number of undulator modules

Fig. 3 | Pulse parameter in the two-colour pulse-pair configuration. shows the average two-colour spectrum without smoothing. The bottom 15
a, Measured (Exp) and calculated (Sim) delays between ω/2ω pulse pairs at spectra show 15 single-shot spectra smoothed with a Gaussian kernel σ of five
different undulator beamline configurations. The centres of the data points and spectral pixels (0.35 eV for ω spectra, 0.85 eV for 2ω spectra). d, Start-to-end
the error bars represent, respectively, the mean and three times the standard simulation of the average ω/2ω pulse pair (averaged over 100 shots) in the second
deviation calculated from 200 bootstrapping iterations. The number of shots undulator section with the chicane off. The power of the 2ω pulse at each location
used to analyse each dataset is listed in Supplementary Table 1. b, Measured in z is normalized by the maximum at that location. The white (black) dashed line
average 2ω pulse energy at different beamline configurations. The error bars shows the centre of mass of the ω (2ω) pulse as a function of the electron beam’s
show the standard deviation in the calibration of the average 2ω pulse energy travel distance in the undulator. The forward propagation of XFEL pulses in
calculated from 29,190 shots. c, Measured spectra of ω/2ω pulse pairs with no quadrupole magnets and free-space drifts with respect to the bunch coordinate
chicane delay and three undulator modules for 2ω pulses. The top spectrum has been removed.

temporal gating effect. This limits the group velocity to the average saturation point. By comparing the experimental data with the simula-
velocity of the electrons40, thus α = 0. In the nonlinear regime, the group tion, we find that the 2ω pulse is quickly amplified to above gigawatt
velocity is typically higher than in the exponential gain regime and can level and propagates ahead of the ESASE current spike, exhibiting
exhibit exotic behaviours such as superluminal propagation (α > 1)41,42. behaviour that is consistent with the early onset of superradiant propa-
Our experiment enables the group velocity of an XFEL pulse to gation41,43,44. Our observations are consistent with the start-to-end
be measured. By comparing the change in delay between ω/2ω pulse simulations plotted in Fig. 3a.
pairs with the slippage per undulator module, we determine the group To further increase the delay, one can use the magnetic chicane
velocity of the 2ω pulse as up to a delay value of tens of femtoseconds. We measure the time delay
for a nominal magnetic chicane value of 1 fs (‘Exp: chicane on’ data
λr point). The discrepancy between the simulated chicane delay and the
vg,2ω = vb (1 + (0.73 ± 0.04) ). (4)
λu nominal experimental chicane delay is probably due to the magnetic
hysteresis of the chicane dipoles operating far from saturation. In
We measure the group velocity of the 2ω pulse to be faster than in the future experiments, one could minimize this effect using one or more
exponential gain regime, because the pulse is generated close to the off-resonance magnetic undulators to introduce such short delays.

Nature Photonics
 Article https://doi.org/10.1038/s41566-024-01419-w

a Demonstration of a time-resolved experiment

To demonstrate the performance of these two-colour pulse pairs,
we carry out a simple pump–probe experiment to measure the
dynamics of core ionization in para-aminophenol molecules. The
para-aminophenol molecule is a prototypical molecular system that
contains the four most abundant atomic components in living organ-
isms (carbon, nitrogen, oxygen and hydrogen) and represents an ideal
testbed for soft X-ray experiments with elemental specificity. We tune
Ekin (eV) the lower photon energy pulse to ℏω = 295.5 eV, to prepare the system in
250 300 350 400 450 500 a non-equilibrium state or to ‘pump’ the system. At this photon energy,
it is most probable that the pump pulse will remove electrons from
299 C A–M
ROI b c
pump the carbon K shell of the para-aminophenol molecule, thus creating
Yield (a.u.)

single core-hole states. On the basis of the measured fluence of the

O A–M
298
C 1s
probe N A–M
probe pump pulse we estimate that nearly every molecule in the focal volume
probe is ionized. The binding energy (BE) of the carbon K-shell electrons in
para-aminophenol is between 289.71 and 291.41 eV, depending on the
BE (eV)

297
relative position of the core-level vacancy to the nitrogen or oxygen
atomic sites47. Ionization by the pump pulse produces slow photoelec-
296 trons with kinetic energy of ~5 eV. We probe the dynamics induced by
PCI (10 eV) the pump pulse with subsequent ionization by the probe pulse tuned
PCI + A–M decay (10 fs) to 2ℏω = 591 eV. From the spectral measurements performed at these
295 Data
photon energies, the estimated pulse duration was 530 as and 290 as
for ω and 2ω, respectively.
0 1 2 3 4 5 6 7
Figure 4a shows a schematic view of the experimental setup.
Time delay (fs)
Photoelectrons ionized by the pump and probe pulses are collected
Fig. 4 | Experimental demonstration of the attosecond pump–probe with a magnetic-bottle time-of-flight electron spectrometer (MBES).
capability. a, Schematic of the experimental setup. Electrons ionized by the Figure 4b shows the electron kinetic energy (Ekin) spectrum recorded
pump and probe pulses are collected using an MBES. Atom colours:
with a delay of 2.1 fs between the pump and probe pulses. We analyse the
C, black; N, blue; O, red; H, white. b, Raw photoelectron spectrum showing
photoelectron feature corresponding to the ionization of the carbon
the Auger–Meitner (A–M) features for carbon, nitrogen and oxygen sites in
K shell (C 1s) by the probe pulse in the kinetic energy window between
para-aminophenol molecules. The pump and probe photon energies are
295.5 and 591 eV, respectively. The shaded region of interest (ROI) shows the
285 and 300 eV. We extract the binding energy of this feature, defined
carbon K-shell (C 1s) ionization feature produced by the probe. c, Black data as BE = 2ℏω − Ekin, and plot the result as a function of the delay between
points showing the pump–probe-delay-dependent binding energy (BE) values the two X-ray pulses (in Fig. 4c). We find that the measured binding
obtained using partial covariance analysis. The centres of the data points and energy of this photoelectron feature increases by ~1.3 eV in the first
the vertical error bars for the binding energy values represent, respectively, few femtoseconds.
the mean and standard deviation of 200 bootstrapping iterations. The number The fast photoelectrons (Ekin ≃ 300 eV) produced by the probe
of shots in each dataset is listed in Supplementary Table 2. The centres of the pulse will rapidly overtake the slower photoelectrons produced by the
data points and the horizontal error bars of the time delay values represent, pump pulse (Ekin ≃ 5 eV). As a consequence of this interaction, the fast
respectively, the mean and standard deviation of 50 shots of simulations with electrons gain energy because they experience additional screening
different random seeds. The grey dashed line shows the static binding energy from the molecular ion. This kinetic energy shift decreases as the spatial
of a K-shell electron in the core-ionized cation (referred to as BEDCH in the
position where the fast electron overtakes the slow electron moves
Supplementary Information). The orange curve shows a model that includes
further away from the ion, and thus the shift vanishes for long pump–
a classical simulation for the PCI effect59 considering a slow pump electron
probe delays48. Such shifts have been referred to as a post-collision
with a kinetic energy of 10 eV and a fast probe electron with a kinetic energy of
300 eV. The green curve adds the A–M decay of a single core-hole state with a
interaction (PCI) effect, and can be modelled with a classical propa-
10 fs timescale on top of the PCI effect. gation in a Coulomb potential. In Fig. 4c we compare our data with
this model in the orange curve. The green curve shows the result of a
calculation that includes both this PCI model and the effect of a finite
For this small value of the chicane delay, the harmonic microbunching core-hole lifetime (10 fs), which is consistent with the short-time behav-
is not entirely suppressed and the delay remains consistent with a 2ω iour of the data (see Supplementary Sections 5.3–5.4 for more details
group velocity larger than the beam velocity. on the models). For the longer time delay (7 fs), other effects should
Figure 3b also shows the average 2ω pulse energy as a function of be taken into account, for instance, the dissociation of a proton (H+)
the undulator length. The 2ω pulse energy increases with the number could occur on this timescale. This would lead to a shift of the bind-
of undulator modules and decreases as the chicane delay is turned on, ing energy to lower values, closer to the core-ionized cation binding
due to the suppression of microbunching induced by chicane disper- energy (the grey dashed line in Fig. 4c). A detailed understanding of the
sion39. This delay-dependent pulse energy can be easily measured dynamics in core-ionized para-aminophenol is beyond the scope of this
on a shot-to-shot basis using, for example, a grating spectrometer. work; however, the measured time-dependent shift demonstrates the
Figure 3c shows an example of such a single-shot characterization of ability to perform pump–probe measurements with subfemtosecond
these pulses, showing simultaneous ω/2ω spectra measured using resolution using an XFEL.
a variable-line-spacing grating spectrometer 45. The full-width at
half-maximum bandwidths are 2.6 ± 0.8 eV and 3.4 ± 0.6 eV for the Conclusion
ω (370 eV) and 2ω (740 eV) pulses, respectively. On the basis of the We have demonstrated the generation of gigawatt-level, two-colour
typical time–bandwidth product observed in both simulation and (ω/2ω), attosecond soft X-ray pulse pairs with controllable, synchro-
previous experiments16,46, we estimate the corresponding full-width nized delays. This is achieved using the split-undulator method in a
at half-maximum pulse durations to be roughly 700 as and 530 as for harmonic configuration, where the harmonic microbunching induced
370 eV and 740 eV pulses, respectively. by the first pulse can seed the second pulse. We directly measure the

Nature Photonics
Article https://doi.org/10.1038/s41566-024-01419-w

average delay of the ω/2ω pulse pair using an angular streaking tech- References
nique for each beamline configuration. The angular streaking measure- 1. Zewail, A. H. Laser femtochemistry. Science 242, 1645–1653
ment shows that subfemtosecond delays can be controlled in steps of (1988).
(27 ± 4)% of the slippage rate in the second undulator section (130 ± 19 as 2. Siegbahn, K. Electron spectroscopy for atoms, molecules, and
at 2ℏω = 740 eV) by changing the number of undulator modules in the condensed matter. Rev. Mod. Phys. 54, 709–728 (1982).
second stage. This is a result of the small deviation of the group veloc- 3. Picón, A. et al. Hetero-site-specific X-ray pump–probe
ity of the second pulse from the speed of light. Longer delays can be spectroscopy for femtosecond intramolecular dynamics. Nat.
accessed using a magnetic chicane to delay the electron bunch with Commun. 7, 11652 (2016).
respect to the fundamental pulse. 4. Berrah, N. et al. Femtosecond-resolved observation of the
The FEL group velocity has so far been a point of theoretical inves- fragmentation of buckminsterfullerene following X-ray
tigation41,49, relevant to long-wavelength experiments where the radia- multiphoton ionization. Nat. Phys. 15, 1279–1283 (2019).
tion slippage strongly affects the FEL dynamics42,50. However, the level 5. Barillot, T. et al. Correlation-driven transient hole dynamics
of experimental sophistication required for attosecond X-ray science resolved in space and time in the isopropanol molecule. Phys.
makes it a key element in understanding and executing pump–probe Rev. X 11, 031048 (2021).
measurements at the subfemtosecond scale. This experiment has 6. Schwickert, D. et al. Electronic quantum coherence in glycine
enabled the group velocity of an XFEL to be measured and provides a molecules probed with ultrashort x-ray pulses in real time.
benchmark for time-dependent and nonlinear FEL theory. Our meas- Science Advances 8, eabn6848 (2022).
urements qualitatively confirm the increase in group velocity in the 7. Engel, G. S. et al. Evidence for wavelike energy transfer through
nonlinear regime predicted using 1D theory41, and are quantitatively quantum coherence in photosynthetic systems. Nature 446,
reproduced by 3D time-dependent simulations. 782–786 (2007).
Our two-colour attosecond pump–probe setup can be scaled to 8. Hentschel, M. et al. Attosecond metrology. Nature 414, 509–513
the next generation of XFEL facilities with megahertz repetition rates, (2001).
as it relies on passive beam-shaping methods. This method can be 9. Goulielmakis, E. et al. Real-time observation of valence electron
applied to any wavelengths of ω/2ω by simply adjusting the undulator motion. Nature 466, 739–743 (2010).
parameter K. This enables the probing of different absorption edges, as 10. Calegari, F. et al. Ultrafast electron dynamics in phenylalanine
long as the probe pulse is a harmonic of the pump. For example, ω/2ω initiated by attosecond pulses. Science 346, 336–339 (2014).
pulse pairs can be exploited in non-resonant-pump/resonant-probe 11. Tzallas, P. et al. Generation of intense continuum
experiments of ionic charge motion5, or resonant-pump/non-resonant extreme-ultraviolet radiation by many-cycle laser fields. Nat. Phys.
probing of core-excited state dynamics51. We have demonstrated the 3, 846–850 (2007).
application of this two-colour technique to a pump–probe experiment 12. Fabris, D. et al. Synchronized pulses generated at 20 eV and
with subfemtosecond resolution, and have resolved a time-dependent 90 eV for attosecond pump–probe experiments. Nat. Photonics 9,
shift of the photoelectron binding energy due to a post-collision inter- 383–387 (2015).
action between electrons, which has been identified as an important 13. Takahashi, E. J., Lan, P., Mücke, O. D., Nabekawa, Y. & Midorikawa,
problem in attosecond physics32. The pulse energies achieved with K. Attosecond nonlinear optics using gigawatt-scale isolated
the ω pulse enable nonlinear excitation techniques, such as impulsive attosecond pulses. Nat. Commun. 4, 2691 (2013).
stimulated X-ray Raman scattering52,53, which can be probed resonantly 14. Okino, T. et al. Direct observation of an attosecond electron wave
with the 2ω pulse. packet in a nitrogen molecule. Sci. Adv. 1, e1500356 (2015).
While this paper addresses several challenges in attosecond 15. Tzallas, P., Skantzakis, E., Nikolopoulos, La. A., Tsakiris, G. D. &
and XFEL physics32,54, future developments will allow us to expand Charalambidis, D. Extreme-ultraviolet pump–probe studies of
the range of scientific opportunities enabled by attosecond XFEL one-femtosecond-scale electron dynamics. Nat. Phys. 7, 781–784
pulses. This method can be extended to non-harmonic pump/probe (2011).
photon energies by utilizing a delay line to delay the first X-ray pulse 16. Duris, J. et al. Tunable isolated attosecond X-ray pulses with
and achieve time overlap, or through using fresh-slice methods29. gigawatt peak power from a free-electron laser. Nat. Photonics 14,
Improvements in the pulse duration and temporal resolution can be 30–36 (2020).
achieved with brighter electron bunches, with future plasma-based 17. Zhang, Z. et al. Experimental demonstration of enhanced
injectors promising the ultimate limit of few-cycle pulses and sub- self-amplified spontaneous emission by photocathode temporal
100 as resolution55. The number of delay data points in our pump– shaping and self-compression in a magnetic wiggler. New J. Phys.
probe experiment is largely limited by the relatively small absorption 22, 083030 (2020).
cross-section at X-ray energies. The availability of high-repetition-rate 18. Duris, J. P. et al. Controllable X-ray pulse trains from enhanced
electron bunches from superconducting accelerators will increase self-amplified spontaneous emission. Phys. Rev. Lett. 126, 104802
the available repetition rate from the 120 Hz used in our experiment (2021).
to kilohertz and, ultimately, megahertz levels (for example, up to 19. Maroju, P. K. et al. Attosecond pulse shaping using a seeded
27 kHz for the European XFEL56 and 1 MHz for the LCLS-II57). This will free-electron laser. Nature 578, 386–391 (2020).
dramatically improve the data collection rate and increase the number 20. Grell, G. et al. Effect of the shot-to-shot variation on charge
of delay points measured in future pump–probe experiments. Finally, migration induced by sub-fs x-ray free-electron laser pulses. Phys.
these methods can be extended to hard X-ray energies by utilizing Rev. Res. 5, 023092 (2023).
the ESASE technique discussed here or simpler single-spike shaping 21. Driver, T. et al. Attosecond transient absorption spooktroscopy:
techniques58. a ghost imaging approach to ultrafast absorption spectroscopy.
Phys. Chem. Chem. Phys. 22, 2704–2712 (2020).
Online content 22. Glownia, J. M. et al. Time-resolved pump–probe experiments at
Any methods, additional references, Nature Portfolio reporting sum- the LCLS. Opt. Express 18, 17620–17630 (2010).
maries, source data, extended data, supplementary information, 23. Kang, H.-S. et al. Hard X-ray free-electron laser with
acknowledgements, peer review information; details of author contri- femtosecond-scale timing jitter. Nat. Photonics 11, 708–713 (2017).
butions and competing interests; and statements of data and code avail- 24. Bionta, M. R. et al. Spectral encoding of x-ray/optical relative
ability are available at https://doi.org/10.1038/s41566-024-01419-w. delay. Opt. Express 19, 21855–21865 (2011).

Nature Photonics
Article https://doi.org/10.1038/s41566-024-01419-w

25. Hartmann, N. et al. Sub-femtosecond precision measurement of 45. Obaid, R. et al. LCLS in—photon out: fluorescence measurement
relative X-ray arrival time for free-electron lasers. Nat. Photonics 8, of neon using soft x-rays. J. Phys. B At. Mol. Opt. Phys. 51, 034003
706–709 (2014). (2018).
26. Maroju, P. K. et al. Attosecond coherent control of electronic wave 46. Li, S. et al. Attosecond coherent electron motion in Auger–
packets in two-colour photoionization using a novel timing tool Meitner decay. Science 375, 285–290 (2022).
for seeded free-electron laser. Nat. Photonics 17, 200–207 (2023). 47. Zhaunerchyk, V. et al. Disentangling formation of multiple-core
27. Lutman, A. A. et al. Experimental demonstration of femtosecond holes in aminophenol molecules exposed to bright X-FEL
two-color X-ray free-electron lasers. Phys. Rev. Lett. 110, 134801 radiation. J. Phys. B At. Mol. Opt. Phys. 48, 244003 (2015).
(2013). 48. Artemyev, A. N., Streltsov, A. I. & Demekhin, P. V. Controlling
28. Hara, T. et al. Two-colour hard X-ray free-electron laser with wide dynamics of postcollision interaction. Phys. Rev. Lett. 122, 183201
tunability. Nat. Commun. 4, 2919 (2013). (2019).
29. Lutman, A. A. et al. Fresh-slice multicolour Xray free-electron 49. Bartolini, R., Doria, A., Gallerano, G. & Renieri, A. Theoretical and
lasers. Nat. Photonics 10, 745–750 (2016). experimental aspects of a waveguide FEL Nucl. Instrum. Methods
30. Emma, P. et al. First lasing and operation of an ångstrom- Phys. Res. A 304, 417–420 (1991).
wavelength free-electron laser. Nat. Photonics 4, 641–647 (2010). 50. Fisher, A. et al. Single-pass high-efficiency terahertz free-electron
31. Mukamel, S., Healion, D., Zhang, Y. & Biggs, J. D. Multidimensional laser. Nat. Photonics 16, 441–447 (2022).
attosecond resonant X-ray spectroscopy of molecules: lessons 51. Al-Haddad, A. et al. Observation of site-selective chemical bond
from the optical regime. Annu. Revi. Phys. Chem. 64, 101–127 (2013). changes via ultrafast chemical shifts. Nat. Commun. 13, 7170
32. Lindroth, E. et al. Challenges and opportunities in attosecond and (2022).
XFEL science. Nat. Rev. Phys. 1, 107–111 (2019). 52. O’Neal, J. T. et al. Electronic population transfer via impulsive
33. Zholents, A. A. Method of an enhanced self-amplified stimulated X-ray raman scattering with attosecond soft-X-ray
spontaneous emission for x-ray free electron lasers. Phys. Rev. pulses. Phys. Rev. Lett. 125, 073203 (2020).
Spec. Top. Accel. Beams 8, 040701 (2005). 53. Mukamel, S., Healion, D., Zhang, Y. & Biggs, J. D. Multidimensional
34. Hartmann, N. et al. Attosecond time–energy structure of X-ray attosecond resonant X-ray spectroscopy of molecules: lessons
free-electron laser pulses. Nat. Photonics 12, 215–220 (2018). from the optical regime. Annu. Rev. Phys. Chem. 64, 101–127 (2013).
35. Li, S. et al. A co-axial velocity map imaging spectrometer for 54. Lépine, F., Ivanov, M. Y. & Vrakking, M. J. Attosecond molecular
electrons. AIP Adv. 8, 115308 (2018). dynamics: fact or fiction? Nat. Photonics 8, 195–204 (2014).
36. Walter, P. et al. The time-resolved atomic, molecular and 55. Emma, C. et al. Terawatt attosecond x-ray source driven by a
optical science instrument at the Linac Coherent Light Source. plasma accelerator. APL Photonics 6, 076107 (2021).
J. Synchrotron Radiat. 29, 957–968 (2022). 56. Decking, W. et al. A MHz-repetition-rate hard X-ray free-electron
37. Itatani, J. et al. Attosecond streak camera. Phys. Rev. Lett. 88, laser driven by a superconducting linear accelerator. Nat.
173903 (2002). Photonics 14, 391–397 (2020).
38. Huang, Z. & Kim, K.-J. Review of x-ray free-electron laser theory. 57. Abbamonte, P. et al. New Science Opportunities Enabled by
Phys. Rev. Spec. Top. Accel. Beams 10, 034801 (2007). LCLS-II X-Ray Lasers (SLAC National Accelerator Lab., 2015);
39. Pellegrini, C., Marinelli, A. & Reiche, S. The physics of x-ray https://doi.org/10.2172/1630267
free-electron lasers. Rev. Mod. Phys. 88, 015006 (2016). 58. Huang, S. et al. Generating single-spike hard X-ray pulses with
40. Baxevanis, P., Duris, J., Huang, Z. & Marinelli, A. Time-domain nonlinear bunch compression in free-electron lasers. Phys. Rev.
analysis of attosecond pulse generation in an x-ray free-electron Lett. 119, 154801 (2017).
laser. Phys. Rev. Accel. Beams 21, 110702 (2018). 59. Russek, A. & Mehlhorn, W. Post-collision interaction and the Auger
41. Yang, X., Mirian, N. & Giannessi, L. Postsaturation dynamics lineshape. J. Phys. B: At. Mol. Phys. 19, 911 (1986).
and superluminal propagation of a superradiant spike in a
free-electron laser amplifier. Phys. Rev. Accel. Beams 23, 010703 Publisher’s note Springer Nature remains neutral with regard to
(2020). jurisdictional claims in published maps and institutional affiliations.
42. Hajima, R., Nishimori, N., Nagai, R. & Minehara, E. Analyses of
superradiance and spiking-mode lasing observed at JAERI-FEL. Springer Nature or its licensor (e.g. a society or other partner) holds
Nucl. Instrum. Methods Phys. Res. A 475, 270–275 (2001). exclusive rights to this article under a publishing agreement with
43. Bonifacio, R., Souza, L. D. S., Pierini, P. & Piovella, N. The the author(s) or other rightsholder(s); author self-archiving of the
superradiant regime of a FEL: analytical and numerical results. accepted manuscript version of this article is solely governed by the
Nucl. Instrum. Methods Phys. Res. A 296, 358–367 (1990). terms of such publishing agreement and applicable law.
44. Mirian, N. S. et al. Generation and measurement of intense
few-femtosecond superradiant extreme-ultraviolet free-electron © The Author(s), under exclusive licence to Springer Nature Limited
laser pulses. Nat. Photonics 15, 523–529 (2021). 2024

1
SLAC National Accelerator Laboratory, Menlo Park, CA, USA. 2Department of Applied Physics, Stanford University, Stanford, CA, USA. 3Stanford PULSE
Institute, SLAC National Accelerator Laboratory, Menlo Park, CA, USA. 4Physics Department, University of Connecticut, Storrs, CT, USA. 5Blackett
Laboratory, Imperial College London, London, UK. 6Paul-Scherrer Institute, Villigen, Switzerland. 7Laboratory for Ultrafast X-ray Sciences, Institute of
Chemical Sciences and Engineering (LUXS), École Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland. 8Linac Coherent Light Source,
SLAC National Accelerator Laboratory, Menlo Park, CA, USA. 9Department of Physics, The Ohio State University, Columbus, OH, USA. 10Chemical
Sciences and Engineering Division, Argonne National Laboratory, Lemont, IL, USA. 11Chemical Sciences Division, Lawrence Berkeley National Laboratory,
Berkeley, CA, USA. 12Department of Physics, Stanford University, Stanford, CA, USA. 13J.R. Macdonald Laboratory, Department of Physics, Kansas State
University, Manhattan, KS, USA. 14Department of Chemistry, Tohoku University, Sendai, Japan. 15Department of Physics and James Franck Institute, The
University of Chicago, Chicago, IL, USA. 16Department of Photon Science, Stanford University, SLAC National Accelerator Laboratory, Menlo Park, CA,
USA. 17Center for Free-Electron Laser Science CFEL, Deutsches Elektronen-Synchrotron DESY, Hamburg, Germany. 18These authors contributed equally:
Zhaoheng Guo, Taran Driver. e-mail: [email protected]; [email protected]

Nature Photonics
Article https://doi.org/10.1038/s41566-024-01419-w

Methods Here Corr(X, Y) is the Pearson correlation coefficient between X and Y.

XFEL setup If the photoelectrons from the carbon and fluorine K shells (ionized by
The temporal profile of the photocathode laser was shaped to produce the ω and 2ω pulses, respectively) are emitted simultaneously, the
a high-current spike in the electron bunch17. Attosecond ω/2ω pulse largest joint variability—that is, the peak in the ρX,̃ Y ̃(θω , θ2ω ) map—will
pairs with controllable subfemtosecond delays were generated by occur along the line θω = θ2ω. Since the time delay Δτ between the two
lasing this high-current spike16 in the split-undulator mode27 with the pulses produces an angular deviation, Δϕ, between the momentum
harmonic configuration. The main accelerator parameters used in shift of the ω/2ω photoelectron features, the largest joint variability
the angular streaking experiment are listed in Extended Data Table 1. will shift to a line θ2ω = θω + Δϕ (see Fig. 2c). The average differential
In the angular streaking experiment, we took two datasets with the angle Δϕ is determined by matching the experimental correlation maps
same undulator and chicane setup but different currents in the second ρX,̃ Y ̃ with strong-field approximation simulations. The error in the time
bunch compressor (BC2). Our angular streaking measurement shows delay measurement was estimated using a bootstrapping approach.
that the change in BC2 current had a negligible effect on the delay For each delay point, the measured data were resampled randomly
between the ω/2ω pulses. with replacement to create 200 unique datasets. For each resampled
The FEL resonance wavelength λr is determined using dataset, an experimental correlation map ρX,̃ Y ̃ was calculated and
matched to the simulated correlation maps to fit a delay. The error bars
λu K2
λr = (1 + ), (5) shown in Fig. 3a represent three times the standard deviation of 200
2γ2 2
delays generated from the bootstrapping method.
where λu = 3.9 cm is the length of one undulator period, γ is the Lorentz We estimated the upper bound of the delay jitter by analysing the
factor of the electron beam and K is the normalized undulator strength amplitudes of the 2D correlation map ρX,̃ Y ̃(θω , θ2ω ) and the amplitudes
parameter. The maximum possible value for the undulator parameter K of two 2D auto-correlation maps ρX,̃ X̃ (θω , θ′ω ) and ρY,̃ Y ̃(θ2ω , θ′2ω ) , where
is ~5.77. The two undulator sections are composed of undulator modules the primed variable corresponds to a different detector angle for the
of 87 periods each. The values of K for each module used to generate same photoemission feature:
the 370 and 740 eV attosecond XFEL pulses in the angular streaking ̃ ω ), X(θ
̃ ′ )),
ρX,̃ X̃ (θω , θ′ω ) = Corr(X(θ ω (7)
experiment are shown in Extended Data Fig. 1. In the split-undulator con-
figuration, the energy spread induced in the electron beam by the first ̃ 2ω ), Y(θ
ρY,̃ Y ̃(θ2ω , θ′2ω ) = Corr(Y(θ ̃ ′ )). (8)
2ω
pulse is detrimental to the generation of the second pulse. Therefore,
to avoid saturation, the parameter K in the first two undulator modules The delay jitter reduces the joint variability between streaking-induced
was slightly detuned to reduce the ω pulse energy. Then, the undulator changes in two photoemission features, which decreases the amplitude
taper from the fifth to the ninth undulator modules matched the energy of the 2D correlation map ρX,̃ Y ̃(θω , θ2ω ). However, the delay jitter does
chirp to generate attosecond 370 eV pulses and to suppress background not influence the amplitudes of two 2D auto-correlation maps
radiation outside the high-current spike. After the magnetic chicane, ρX,̃ X̃ (θω , θ′ω ) and ρY,̃ Y ̃(θ2ω , θ′2ω ). Therefore, by comparing the amplitude
the undulator K in the second section was decreased from ~5.6 to ~3.8 of ρX,̃ Y ̃(θω , θ2ω ) with the amplitudes of ρX,̃ X̃ (θω , θ′ω ) and ρY,̃ Y ̃(θ2ω , θ′2ω ), we
to generate attosecond 740 eV pulses. The number of undulators in the can determine the upper bound for the jitter. We add delay jitter to
second section can be controlled by varying the gap of the individual the strong-field approximation simulations until the amplitude ratio
undulator sections in the second stage, so that only the desired number of the simulation matches the measurement.
of undulators is resonant with the second pulse. A similar setup was used
to generate the 295.5 eV/591 eV attosecond XFEL pulses, with the electron Pump–probe measurements in para-aminophenol
beam energy at 4 GeV and different values of the undulator parameter K. The pump–probe measurements were performed at the time-resolved
Control of the pulse delay is accomplished by varying the number atomic, molecular and optical science (TMO) experimental hutch,
of undulator sections in the second stage or by varying the chicane where the X-ray pulses were focused using a pair of Kirkpatrick–Baez
delay, as discussed above in ‘Demonstration of two-colour attosecond focusing mirrors36. The para-aminophenol was introduced using an
pulse pairs’. Changing the pulse-to-pulse delay involves routine opera- in-vacuum oven heated to ~160 °C. The focused X-ray beam intercepted
tional steps that can typically be performed in less than one minute. the molecular sample in the interaction region of a two-metre-long
MBES. A static 140 V retardation voltage was applied to the flight tube
Analysis of angular streaking data of the MBES to increase the kinetic energy resolution for the electrons
Since the jitter between the X-ray arrival time and the laser pulse is of interest. The spectrum of the probe pulse was measured shot to shot
greater than the laser period, the phase of the streaking laser field when using the variable-line-spacing grating spectrometer.
the X-ray pulse arrives is essentially random. To deal with the random To retrieve the binding energy shift analysed in Fig. 4, we applied a
distribution of laser/X-ray arrival times, we used a correlation analysis partial covariance analysis procedure. Covariance analysis is commonly
method to extract the average delay and the jitter between the ω/2ω used in the analysis of SASE (self-amplified spontaneous emission)
pulses. Each single-shot 2D photoelectron momentum distribution XFEL experiments to gain resolution through the use of fluctuations,
was re-binned to polar coordinates. The electron yield for the nth X-ray since SASE XFEL pulses exhibit shot-to-shot fluctuations of the pulse
pulse was integrated over the the high-energy flank of each photoemis- properties (pulse energy, photon energy, and so on). Analysing the
sion feature to give two 1D traces, Xn(θω) and Yn(θ2ω), for the carbon and partial covariance between the photoelectron spectra and the probe
fluorine photoemission features, respectively, as shown in Extended photon spectra further enables us to isolate the probe contribu-
Data Fig. 2. These two traces contain the information of the tions to the photoelectron spectrum, removing most of the pump
streaking-induced change in the photoelectron momentum distribu- contributions60. This was done by calculating the partial covariance
tion. A 2D map of correlation coefficients, ρX,̃ Y ̃(θω , θ2ω ), was generated matrix,
by calculating the Pearson correlation coefficient between the antisym- −1
pCov(A, B; I ) = Cov(A, B) − Cov(A, I )Cov (I, I )Cov(I, B), (9)
metric parts of two traces:
1 where Cov(X, Y) is the covariance coefficient between X and Y, A is the
Xñ (θω ) = [Xn (θω ) − Xn (θω + π)] ,
2 shot-to-shot spectrum of the probe pulse, B is the measured photo-
Yñ (θ2ω ) =
1
[Yn (θ2ω ) − Yn (θ2ω + π)] , (6) electron spectrum in the vicinity of the features ionized by the probe
2
pulse and I is the fluctuating parameter, which in this instance is the
̃ ω ), Y(θ
ρX,̃ Y ̃(θω , θ2ω ) = Corr(X(θ ̃ 2ω )). intensity of the pump pulse.

Nature Photonics
Article https://doi.org/10.1038/s41566-024-01419-w

Extended Data Figure 3 shows the 2D maps of partial covariance Sciences under awards DE-FG02-04ER15614 and DE-SC0012462.
coefficients for the electron kinetic energy range from 285 to 300 eV, V.A. and M.R. acknowledge support from the UK’s Engineering and
where we observe photoelectron features from carbon K-shell ioniza- Physical Sciences Research Council (EPSRC) through the grant
tion by the probe pulse. These maps are smoothed along the photoelec- ‘Quantum entanglement in attosecond ionization’, grant number
tron kinetic energy axis using a five-bin rolling average. We observe a EP/V009192/1. O.A. and J.P.M. were supported by UK EPSRC grant
dispersive feature with a delay-dependent binding energy. We do not numbers EP/R019509/1, EP/X026094/1 and EP/T006943/1. T.W., D.S.S.
observe a feature corresponding to ground-state un-pumped mol- and O.G. are supported by the US Department of Energy, Office of
ecules, which are expected 5 eV higher in the photoelectron kinetic Basic Energy Sciences, Division of Chemical Sciences, Geosciences,
energy spectrum. This is consistent with the fluences used for which and Biosciences under the contract number DE-AC02-05CH11231.
we expect complete saturation of the K-shell absorption. L.Y. and G.D. were supported by the US Department of Energy, Office
To extract the binding energy from each covariance map, we fit a of Science, Basic Energy Sciences, Division of Chemical Sciences,
dispersive line, 2ℏω = Ekin + BE, to the peak of the covariance map (blue Geosciences, and Biosciences under award DE-AC02-06CH11357. D.R.,
dashed line in Extended Data Fig. 3). The corresponding error bars A.R. and E.W. are supported by the same funding agency under grant
shown in Fig. 4 result from a bootstrapping approach: at each time number DE-FG02-86ER13491. S.B. and N.B. are supported by the US
delay, the data are randomly resampled with replacement to create Department of Energy, Office of Basic Energy Sciences, Division of
200 unique datasets with the same number of shots as in the original Chemical Sciences, Geosciences, and Biosciences under the contract
dataset. For each resampled dataset, a 2D partial covariance map is number DE-SC0012376. A.M. would like to acknowledge L. Giannessi
calculated and a binding energy is fitted to the peak of the covariance. and P. Musumeci for useful discussions and suggestions. L.I. would like
The error bar of the measured binding energy presented in Fig. 4 is the to acknowledge helpful discussion with S.-K. Son and acknowledges
standard deviation of the retrieved binding energy from these 200 support from DESY (Hamburg, Germany), a member of the Helmholtz
datasets. The value of average time delays in Fig. 4c is calculated by Association HGF, and the Cluster of Excellence ‘CUI: Advanced
scaling the measured delays to the photon energy used in the pump– Imaging of Matter’ of the Deutsche Forschungsgemeinschaft (DFG) –
probe experiment with the aid of start-to-end XFEL simulations. The EXC 2056 – project ID 390715994.
error bar of each time delay Fig. 4c represents the standard deviation
of simulated time delays. Author contributions
We adjust the relative calibration between our photon spectrom- Z.G., T.D., J.P.C. and A.M. conceived the two-colour streaking
eter and electron spectrometer using a global offset of 0.36 eV to match experiment. C.B., V.A., M.R., L.F.D., G.D., O.G., D.R., A.R., D.S.S., K.U.,
the previously reported binding energy of a two-site double core hole, T.W., L.Y., P.H.B., J.P.M., M.F.K., P.W., N.B., T.D., J.P.C. and A.M. conceived
that is, 296.53 eV (ref. 47) (Supplementary Table 3). the aminophenol pump–probe experiment. Z.G., D.C., D.B., K.A.L.,
J.D., P.L.F., R.R.R., N.S.S., Z.Z. and A.M. set up the attosecond XFEL
Data availability configuration. Z.G., P.L.F., S.L., T.D., J.W., E.I., K.A.L., J.M.G., X.C., X.L.,
A subset of the raw data used to produce Figs. 1–4 is publicly available via M.-F.L., A.K., R.O., N.S.S., E.T., M.F.K., J.P.C. and A.M. conducted the
Figshare at https://doi.org/10.6084/m9.figshare.23232350.v2 (ref. 61). angular streaking measurements to determine the pulse separation.
This repository also contains a copy of the analysis script used to gen- Z.G., T.D., S.B., D.C., J.D., P.L.F., O.A., C.B., X.C., L.F.D., G.D., R.F., O.G.,
erate the streaking correlation maps and the photoemission spec- J.M.G., E.I., A.K., K.A.L., S.L., X.L., M.-F.L., G.A.M., R.O., J.T.O., R.R.R.,
troscopy data. All other data that support the plots within this paper D.R., A.R., D.S.S., N.S.S., D.T., E.T., K.U., E.W., A.L.W., J.W., T.W., T.J.A.W.,
and other findings of this study are available from the corresponding L.Y., Z.Z., P.H.B., J.P.M., M.F.K., P.W., N.B., J.P.C. and A.M. conducted
authors upon reasonable request. the aminophenol pump–probe experiment. Z.G., S.B., P.L.F., S.L., T.D.,
Z.H., R.R.R., E.I., J.W., L.I., N.B., J.P.C. and A.M. performed the data
References analysis and interpreted the data. Z.G., Z.Z., R.R.R. and D.C. conducted
60. Frasinski, Leszek J. Covariance mapping techniques. J. Phys. B: At. numerical simulations of the FEL. All authors were involved in the
Mol. Opt. Phys. 49, 152004 (2016). writing of the paper.
61. Guo, Z. et al. Raw data for Experimental Demonstration
of Attosecond Pump–Probe Spectroscopy with an X-Ray Competing interests
Free-Electron Laser. FigShare https://doi.org/10.6084/ The authors declare no competing interests.
m9.figshare.23232350.v2 (2023).
Additional information
Acknowledgements Extended data is available for this paper at
Use of the Linac Coherent Light Source (LCLS), SLAC National https://doi.org/10.1038/s41566-024-01419-w.
Accelerator Laboratory, is supported by the US Department of Energy,
Office of Science, Office of Basic Energy Sciences under contract Supplementary information The online version contains supplementary
number DE-AC02-76SF00515. A.M., D.C., P.L.F., R.R.R., Z.H. and Z.G. material available at https://doi.org/10.1038/s41566-024-01419-w.
acknowledge support from the Accelerator and Detector Research
Program of the Department of Energy, Basic Energy Sciences division. Correspondence and requests for materials should be addressed to
Z.G., P.L.F. and R.R.R. also acknowledge support from the Robert James P. Cryan or Agostino Marinelli.
Siemann Fellowship of Stanford University. The effort from T.D., J.W.,
E.I., J.T.O., A.L.W., M.F.K., P.H.B., T.J.A.W. and J.P.C. is supported by the Peer review information Nature Photonics thanks Brian Abbey,
US Department of Energy, Basic Energy Sciences, Division of Chemical Heung-Sik Kang and the other, anonymous, reviewer(s) for their
Sciences, Geosciences and Biosciences (CSGB). C.B. acknowledges contribution to the peer review of this work.
funding from the Swiss National Science Foundation (SNSF) project
grant 200021-197372. L.F.D., D.T. and G.A.M. acknowledge support Reprints and permissions information is available at
from the US Department of Energy, Office of Science, Basic Energy www.nature.com/reprints.

Nature Photonics
Article https://doi.org/10.1038/s41566-024-01419-w

Extended Data Table 1 | Accelerator configuration

Parameter Value

Beam energy 5 GeV

Bunch charge 140 pC
BC1 current 140 A
BC2 current 1700 A/2700 A
XLEAP wiggler period 35 cm
XLEAP wiggler gap 25 mm
XLEAP wiggler Kw 24.7
Machine parameters for generating 370 eV/740 eV XFEL pulses using the split-undulator mode in the angular streaking experiment.

Nature Photonics
Article https://doi.org/10.1038/s41566-024-01419-w

Extended Data Fig. 1 | Undulator configuration in the ω/2ω mode. Values of undulator K were set to be on resonance with ω and 2ω pulses in the first and second
undulator section, respectively. The transparent grey area shows the location of the magnetic delay chicane.

Nature Photonics
Article https://doi.org/10.1038/s41566-024-01419-w

Extended Data Fig. 2 | Measured two-dimensional projection of the electron yield integrated over all detector angles. Two pairs of dashed lines
photoelectron momentum distribution recorded by the c-VMI in the absence label lower and upper bounds of integrating electron yields in 2 photoemission
of the circularly polarised streaking field. The left panel shows the data from features for the delay analysis.
Fig. 1c in polar coordinates. The 1-D trace on the right-hand-side shows the

Nature Photonics
Article https://doi.org/10.1038/s41566-024-01419-w

Extended Data Fig. 3 | Partial covariance between the measured photon averaged photon spectra for each delay. The blue dashed line shows the best fit
energy spectrum and the photoelectron kinetic energy spectrum in the line to the dispersive photoemission feature. The black dotted line is the fit of the
vicinity of the carbon K-shell ionisation features. The partial covariance uses first delay that is used as a reference.
the pump pulse intensity as a fluctuating parameter. The red curve shows the

Nature Photonics