Article https://doi.org/10.

1038/s41467-023-43318-8

Discovery of ultrafast spontaneous spin

switching in an antiferromagnet by
femtosecond noise correlation spectroscopy
Received: 23 January 2023 M. A. Weiss 1, A. Herbst 1, J. Schlegel1, T. Dannegger 1
, M. Evers1, A. Donges1,
M. Nakajima2, A. Leitenstorfer 1, S. T. B. Goennenwein 1
, U. Nowak 1 &
Accepted: 6 November 2023
T. Kurihara 1,3

Check for updates Owing to their high magnon frequencies, antiferromagnets are key materials
1234567890():,;
1234567890():,;

for future high-speed spintronics. Picosecond switching of antiferromagnetic

spin systems has been viewed a milestone for decades and pursued only by
using ultrafast external perturbations. Here, we show that picosecond spin
switching occurs spontaneously due to thermal fluctuations in the anti-
ferromagnetic orthoferrite Sm0.7Er0.3FeO3. By analysing the correlation
between the pulse-to-pulse polarisation fluctuations of two femtosecond
optical probes, we extract the autocorrelation of incoherent magnon fluc-
tuations. We observe a strong enhancement of the magnon fluctuation
amplitude and the coherence time around the critical temperature of the spin
reorientation transition. The spectrum shows two distinct features, one cor-
responding to the quasi-ferromagnetic mode and another one which has not
been previously reported in pump-probe experiments. Comparison to a sto-
chastic spin dynamics simulation reveals this new mode as smoking gun of
ultrafast spontaneous spin switching within the double-well anisotropy
potential.

Some of the most intriguing effects in physics rest on fluctuations. magnons. Antiferromagnets (AFM) possess especially high-frequency
Incoherent thermal fluctuations of spins critically determine the magnons reaching into the THz range16, and are thereby attracting
magnetic properties of correlated systems. Thermally excited mag- enormous attention from the viewpoint of accelerating the conven-
nons are a driving force of a rich variety of fundamental physical tional ferromagnet-based spintronics devices17,18. However, due to
phenomena such as phase transitions and spin caloritronic effects1–3, their ultrafast dynamics that go beyond state-of-the-art electronics,
whereas their non-deterministic properties promise unique applica- conventional SNS cannot detect antiferromagnetic spin fluctuations.
tions such as probabilistic computing4. Incoherent spin fluctuations The rich spin physics in AFMs arising from their complicated spin
have traditionally been studied either indirectly through the tem- textures have only been resolved with ultrafast pump-probe spectro-
perature dependence of macroscopic properties such as heat capacity, scopy, where time resolutions down to femto- or even attoseconds19
conductivity, and magnetic susceptibility5,6, or in the frequency are available. Still, such pump-probe measurements are of perturba-
domain through optical probes relying on, e.g., the Raman effect7,8 or tive nature and therefore, they cannot detect the incoherent dynamics
diffraction9. For measuring relatively slow spin fluctuation dynamics in that spontaneously exist due to thermal or quantum mechanisms.
paramagnets in the range of MHz to GHz frequencies, spin noise In this work, we experimentally demonstrate the spontaneous
spectroscopy (SNS)10–15 has been employed. In contrast to para- incoherent sub-THz magnon fluctuation dynamics in the AFM. This is
magnets, correlated spin systems exhibit collective excitations, i.e., achieved by a unique experimental principle, inspired by the emerging

1
Department of Physics, University of Konstanz, D-78457 Konstanz, Germany. 2Institute of Laser Engineering, Osaka University, 565-0871 Osaka, Japan. 3The
Institute for Solid State Physics, The University of Tokyo, 277-8581 Kashiwa, Japan. e-mail: [email protected]

Nature Communications | (2023)14:7651 1

Article https://doi.org/10.1038/s41467-023-43318-8

field of sub-cycle quantum optics20–22. Here, we analyse magnon fluc- a-axis at the lower critical temperature T L to the c-axis at the upper
tuation dynamics via their temporal autocorrelation function, by critical temperature T U . The critical temperatures have been pre-
measuring the statistical correlations of polarisation noise imprinted viously measured to be T L, ref = 310 K and T U, ref = 330 K24. The SRT
on two subsequent femtosecond probe pulses (see Fig. 1a). The two is expressed by a change in the free energy potential (see Fig. 1b). T L
linearly polarised, spectrally separate pulses with a variable time delay marks the temperature where the anisotropy difference between a-
Δt are focused on the sample. Upon transmission of the pulses at times and c-axis disappears, causing an enhanced magnetic susceptibility in
t and t + Δt, transient magnetisation fluctuations δM c ðtÞ and the a-c-plane and strong softening of the qF mode resonance
δM c ðt + ΔtÞ parallel to the propagation direction introduce polarisa- frequency24,30. When such a system is thermally populated, one
tion changes δαðtÞ and δαðt + ΔtÞ to each probe, respectively, via the expects a strong enhancement of the angular distribution width at T L .
Faraday effect. The polarisation states of the transmitted probe pulses
are individually analysed with independent polarimetric detectors. The Volume scaling of the magnon noise
pulse-to-pulse fluctuations of the detector output are extracted by First, we investigate the magnon noise dynamics as a function of the
sub-harmonic lock-in amplification20, multiplied in real time and time delay Δt and its scaling with the probed volume Ω at a tempera-
averaged over ~108 pulses at each delay position. By this method, the ture of 294.15 K (Fig. 2a). The confocal position of the sample is iden-
time correlation

trace of the out-of-plane sub-THz magnetisation tified by the maximum signal amplitude (z = 0 µm, green graph). A
dynamics δM c ðt ÞδM c ðt + Δt Þ is precisely unravelled22,23 (see “Meth- systematic decrease of the amplitude is observed with extending z up
ods” section for further details). to ±20 µm (blue, yellow, magenta, and red graphs in Fig. 2a). The sig-
nals are symmetric around Δt = 0, consistent with the fact that we
Results probe an autocorrelation of the temporal dynamics of the system. All
Experimental sample waveforms exhibit a distinct peak at Δt = 0 followed by a gradual
Our sample is a single crystal of the orthoferrite Sm0.7Er0.3FeO324. In decrease and slow oscillations around the zero level that lasts for
this material, the electron spins of the Fe3+ ions are anti- several tens of picoseconds. Figure 2b shows the noise amplitude at
ferromagnetically coupled. A Dzyaloshinskii–Moriya interaction25,26 Δt = 0 as a function of the longitudinal sample position with respect to
(DMI) results in a slight spin canting and a weak net ferromagnetic the optical focus (blue circles). As illustrated by the black graph, the
moment (net magnetisation M). Orthoferrites have two exchange correlated Faraday noise amplitude is fitted well by a function which is
modes with resonance at multi-THz frequencies and two magnon inversely proportional to the volume Ωð z Þ probed by the fundamental
modes in the sub-THz regime27,28. The latter include a quasi-
ferromagnetic mode (qF mode) and a quasi-antiferromagnetic mode
(qAF mode). The qF mode is characterised by a precession of the weak a
0.20 -20 μm
ferromagnetic moment around its equilibrium axis, whereas the qAF
-10 μm
mode results in its longitudinal oscillation. Sm0.7Er0.3FeO3 shows a
correlated noise (mV )

0 μm
2

temperature-induced second-order spin reorientation transition (SRT) 0.15 +7.5 μm

close to room temperature29, in which the magnetic anisotropy (but +20 μm
0.10
not the antiferromagnetic spin order) change. In the SRT region
ðT L < T < T U Þ, M continuously rotates from along the crystallographic
0.05

0.00
a 't
GD t
BPD -0.05
a L2 DM
Amp
BP -40 -20 0 20 40
L1 GD t+'t HWP time delay Δt (ps)
c Subh
WP
Sm0.7Er0.3FeO3 b
Data analysis h
Sub Experiment
BPD Fit
0.20
correlated noise (mV )

Cross-correlation
2
p

BP GD1(t) GD2(t+'t)
Am

b 0.15
F F F F
T T
T T -
-
0.10

T < T T T
T<<TL T≈TL TL<T<TU T≥TU
0.05
Fig. 1 | Schematic illustration of the experimental setup and spin system. a Due
to the Faraday effect, two spectrally separated fs pulses (orange and red) of variable 0.00
time delay Δt experience a polarisation rotation proportional to out-of-plane spin
-40 -20 0 20 40
fluctuations δM c . Corresponding rotation angles δα 1,2 are measured in separate
sample position (μm)
elipsometers. After extraction of the pulse-to-pulse fluctuations from each branch,


the cross-correlation function δα 1 δα 2 is calculated in real time as a function of Fig. 2 | Dependence of magnon noise dynamics on probing volume. a Spin noise
the delay time Δt. L1,L2: lenses; DM: dichroic mirror; HWP: half-wave plate; WP: waveforms recorded at a constant temperature of 294.15 K for multiple long-
Wollaston prism; BPD: balanced photodetector; BP: electronic bandpass filter; itudinal sample positions relative to the laser focus. The confocal position z = 0 was
Amp: transimpedence amplifier; Subh: sub-harmonic demodulation scheme. determined by maximising the amplitude at Δt = 0 (green graph). Amplitudes
b Temperature evolution of the free energy F and its influence on the spin noise decrease monotonically with increasing distance from focus (blue, yellow, magenta
dynamics (red cones) close to the spin reorientation in Sm0.7Er0.3FeO3. The weak and red graphs). b Correlated noise amplitude at Δt = 0 as a function of lateral
ferromagnetic moment M (black arrow) gradually rotates from θ = 0° at T ≤ T L to sample position relative to the focus (blue open circles). The longitudinal position
θ = ± 90° at T ≥ T U . Here, θ is the angle between M and the a-axis of the sample. dependence was fitted with the function given in the Methods section.

Nature Communications | (2023)14:7651 2

Article https://doi.org/10.1038/s41467-023-43318-8

noise amplitude (mV )

a c b 1.2
Experiment Simulation
294.15 K 301.37 K
3.0 0.8
0.6
303.15 K 301.72 K 0.4

correlated noise (arb. units)

correlated noise (mV )
2

2
2.0
0.4 307.15 K 301.89 K
300 310 320
temperature (K)

noise amplitude (arb. units)

(0.5x) 313.15 K (0.5x) 302.41 K d 1.2
1.0
0.2
319.15 K 303.45 K 0.8

0.4
323.65 K 311.00 K
0.0 0
0
-80 -40 0 40 80 -80 -40 0 40 80 300 305 310 315
time delay Δt (ps) time delay Δt (ps) temperature (K)

Fig. 3 | Ultrafast magnon noise dynamics near spin reorientation in the uncertainty associated with the background subtraction procedure (see
Sm0.7Er0.3FeO3. a Correlated noise as a function of time delay Δt between probing “Methods” section). c Magnon noise simulation based on atomistic spin models
pulses for multiple temperatures near the spin reorientation in Sm0.7Er0.3FeO3. and the stochastic Landau-Lifshitz-Gilbert equation. d Temperature evolution of
b Experimentally determined time-zero amplitude as a function of temperature. In the simulated time-zero amplitude.
the region of noise enhancement around 312 K, error bars are added considering

Gaussian mode hδα ðt Þ2 i / Ωð1z Þ : Note that in paramagnets, the ampli- mode, which is expected to reduce as δM c / cosðθÞ towards higher
tude δα of the statistical fluctuations of N independent spins within the temperature. Therefore, the noise amplitude is expected to decrease
probing volume Ω is known to follow the scaling law δα / p1ffiffiffi Ω
/ p1ffiffiffi
N
11,15
. continuously. These temperature dependencies indicate that the
Here, we find the same volume scaling of Faraday noise as in conven- amplitude of the observed magnon noise can be naively understood as
tional SNS. Note that this dependence is not trivially understood for the angle distribution of spins due to thermal occupation of the ani-
correlated spin systems where individual spins are coupled to form sotropy potential well by magnons, consistent with the model descri-
collective magnons. Our observation indicates the existence of bed in Fig. 1b.
mutually incoherent oscillators whose spatial extent is smaller than the These findings are analysed exploiting simulations of the spin
probe spot size. While the origin of such oscillators cannot be fully noise around the SRT in a generic orthoferrite with parameters fitting
identified at this stage, we speculate that this scaling may be indicative the equilibrium properties observed experimentally (see “Methods”
of a multi-domain magnetic state that effectively limits the coherence section). The theory is based on an atomistic spin model and the sto-
length of magnons. Nevertheless, in the following, we fix our sample chastic Landau-Lifshitz-Gilbert equation31,32. Figure 3c depicts the
position to z = 0 µm to maximise the signal based on this feature. simulated c-axis magnon noise autocorrelation trace in a
192 x 192 x 192 orthoferrite spin lattice. The simulation reproduces the
Temperature dependence of the waveforms around the SRT temporal shape of the noise waveforms in Fig. 3a, including the sym-
Figure 3a depicts the striking variation of spin noise autocorrelation metry around Δt = 0 and the temperature evolution of both the time-
waveforms found around the SRT. The amplitudes, oscillation periods zero peak amplitude and the subsequent slow oscillation. The peak
and lifetimes depend strongly on temperature. Since in the SRT only amplitude of the calculated waveforms is shown as a function of
the orientation of the spin system changes, this strong variation of the temperature in Fig. 3d. In the simulations, the SRT manifests as a
signal indicates that the observed autocorrelation signal truly comes strong noise enhancement around 302 K followed by a decrease at
from magnetisation noise dynamics. The finite width of the auto- higher temperatures. The nearly quantitative agreement between the
correlation around the peak (Δt = 0) indicates that the system exhibits theoretical calculation and the experimental data allows us to inves-
non-zero coherence on the ultrafast timescale. The oscillations with a tigate the stochastic nature of the spin noise dynamics from a micro-
period of tens of picoseconds, clearly visible in some of the auto- scopic viewpoint in the following discussions.
correlation waveforms (e.g., 294.15 K and 303.15 K) on both sides of the
peak, reflects the spin precession mode. The temporal dynamics is Spectral analysis of magnetisation fluctuations
discussed later in more detail. We now analyse the dynamics in the frequency domain. The Fourier
Now, we first focus on the temperature evolution of the signal spectra of the detected noise autocorrelation waveforms are shown in
amplitude. The amplitude at Δt = 0 is depicted in Fig. 3b. A sharp Fig. 4a. Interestingly, two distinct spectral peaks (purple and green
amplitude increase is observed in the region around 312.15 K. This arrows) are observed for most temperatures. The frequencies and
point is slightly higher but close to the estimated lower threshold amplitudes of both peaks are strongly dependent on temperature.
temperature T L ~ 305 K of the SRT in our sample, around which tem- While the two peaks are clearly distinguishable for T ≤ 304 K, they
perature the anisotropy softening results in an enhanced magnetic exhibit similar frequencies when approaching the noise enhancement
susceptibility27. Beyond this temperature, the noise amplitude region around T L and eventually become indistinguishable due to the
decreases continuously, almost disappearing around the upper strong broadening. At temperatures T ≫ T L , the spectral amplitude is
threshold T U = 320 K. The sharp decrease observed towards the significantly reduced because of the SRT. Figure 4b shows the fre-
higher temperature side is explained by the equilibrium rotation of the quencies of each peak as a function of temperature. Both peaks
spin system within the SRT. In this temperature region, the net mag- experience a strong frequency reduction around T L ≈ 305 K, which
netisation M continuously rotates from M // a (in-plane) to M // c (out- closely resembles the softening behaviour of the qF mode around the
of-plane) (see Supplementary Information SI2). Our Faraday probe is SRT. The temperature dependence of the high-frequency (HF) peak
sensitive only to the c-axis magnetisation fluctuation δM c of the qF (green full circles in Fig. 4b) is shown to quantitatively match with

Nature Communications | (2023)14:7651 3

Article https://doi.org/10.1038/s41467-023-43318-8

5.0
2.0 a c

frequency (GHz)
Experiment Simulation b 50
TL
40
294.15 K 301.37 K
4.0 30
spectral amplitude (arb. units)

spectral amplitude (arb. units)

HF mode
1.5 303.15 K 301.72 K LF mode 20
PP data
10
3.0
295 300 305 310 315
1.0 temperature (K)
307.15 K 301.89 K 2.0

frequency (GHz)
d 150
0.5 TL,sim
(0.2x) 313.15 K (0.2x) 302.41 K 1.0 100
319.15 K 303.46 K
50
0 323.65 K 311.00 K
0
0
20 40 60 80 0 20 40 60 80 298 300 302 304 306
frequency (GHz) frequency (GHz) temperature (K)

Fig. 4 | Magnon noise spectra near SRT in Sm0.7Er0.3FeO3. a Fourier spectra of mode) frequency data obtained by THz pump-near IR probe24 are shown as pink
the measured magnon noise waveforms from Fig. 3a. Two maxima are resolved for crosses. The pump-probe data is shifted by −7.5 K to compensate for the different
T < 307.15 K (purple and green arrows), while only the low-frequency (purple amount of stationary heating in our experiment. c Fourier spectra of the simulated
arrows) peak prevails for higher temperatures. b Temperature evolution of the waveforms (Fig. 3c). Green and purple arrows indicate the centre frequencies of
high-frequency peak (dark green full circles) and the low-frequency peak (purple the qF mode and the low-frequency feature, respectively. d Temperature evolution
full circles). The values were obtained either by fitting the spectra in a with a of the simulated c-axis projection of the qF mode frequency. The values were
double Lorentzian function or evaluation of the 2nd derivative zero crossing points obtained by fitting the spectra in c with a triple Lorentzian function (see “Methods”
(see “Methods” section). For comparison, quasi-ferromagnetic magnon mode (qF section).

pump-probe data24 (pink crosses), clearly identifying this HF peak as Information SI5). With increasing temperature, the number of
the qF mode in Sm0.7Er0.3FeO3. On the other hand, the low-frequency observed switches gradually decreases, while at the same time, the
peak (LF peak) has no correspondence in pump-probe measurements. distance Δmc between the up-and down states increases. For suffi-
It suggests that this LF peak is linked to the randomness of the spon- ciently high temperatures T ≫ T L, sim , no more switching events are
taneous magnetic dynamics, which has now become accessible by the recorded. Here, mc =mS always fluctuates around a preferred state.
stochastic nature of our technique. When comparing the temperatures at which the LF peak is observed in
The Fourier spectra of the stochastic LLG simulations are depicted the simulated spectra (Fig. 4c) with the temperatures at which RTN is
in Fig. 4c. The appearance of the two peaks and their softening around recorded in the magnetisation time traces (Fig. 5a), it becomes clear
T L, sim ≈ 301.5 K (Fig. 4d) matches the experimental results in Fig. 4b. that the emergence of the LF feature is inherently linked to the RTN.
This agreement between the temperature dependence of the simu- Note that the autocorrelation of a pure two-level random tele-
2
2jΔt j
lated qF mode fluctuation and the HF mode seen in the experiment graph noise follows an exponential decay of the form ðΔm 2 Þ e
c τ for

further supports our assignment to the original qF mode. Conversely, finite time Δt , with τ being the mean dwell time. Since both Δmc and
36

the LF feature appears in the experimental data from the spectrum Δt increase with the progress of SRT, the RTN dynamics should con-
recorded at a temperature of T = 294.15 K to temperatures well tribute to an exponentially decaying component with an increasing
beyond T L ≈ 305 K, whereas it is observed in a narrower temperature amplitude and decay time. This behaviour is indeed observed in both
region T ⪆ T L, sim in the simulation (Fig. 4c). This quantitative dis- the simulated and measured autocorrelation traces (Fig. 3a, c), which
crepancy stems from the practical limitations of our simulation in unambiguously proves that the measured noise signals reflect the RTN
finding a set of parameters to fit the experimental results (see “Meth- dynamics. In the frequency domain, it manifests itself as the LF peak. It
ods” section). Still, both the experimental and simulated temperature should be mentioned that the Fourier transform of a purely expo-
dependence of the LF spectral amplitude (Suppl. Figs. 1 and 4) follow a nentially decaying autocorrelation function results in a Lorentzian
similar trend as the time-zero amplitude as a function of temperature spectrum centred around zero36,37. In contrast, in our observations the
shown in Fig. 3b. This finding suggests the underlying LF dynamics to LF feature exhibits a peak at finite frequencies. We attribute this dif-
contribute significantly to the total noise amplitude (Fig. 3b, d). ference to the limited time window over which our traces are analysed.
To investigate this RTN behaviour in further detail, we plot the
Comparison with simulated temporal evolution of trajectory of the LLG-simulated magnetisation for different tempera-
magnetisation tures around the SRT (Fig. 5b–f and Supplementary Video 9). At
To gain insights into the physical origin of the LF feature, we now T = 298:24 K < TL, sim (Fig. 5b), the equilibrium magnetisation points
investigate the simulation data in more detail. Figure 5a shows results along the a-axis (θ = 0) and qF mode noise can be observed in the b-
for the temporal evolution of the c-axis projection of the normalised and c-projections (see Suppl. Fig. 5).
magnetisation mc =mS (mS is the magnetisation at saturation). For At T = T L, sim ≈ 301:54 K (Fig. 5c), an enhancement of the qF mode
T < T L, sim , the equilibrium axis of the normalised magnetisation is noise is observed. The spin system then rotates towards finite angle θ
parallel to the a-axis. Consequently, fluctuations of mc =mS centred for T > T L, sim . The ±θ states are energetically degenerate and
around the origin are observed. When approaching T L, sim , the fluc- switching between them occurs for temperatures slightly above
tuations increase in amplitude and oscillation period in agreement T L, sim , i.e., for T = 302:88 K ⪆ TL, sim (Fig. 5d). This results in an
with our previous discussion. For T ⪆ T L, sim , mc =mS switching additional magnetisation noise contribution (see Suppl. Fig. 5c) and
between two discrete states with similar amplitude but different sign accounts for the strongly enhanced noise amplitude in the
(up- and down-state) become prominent, resembling random tele- T = 302:41 K ⪆ TL, sim waveform in Fig. 3c. The configuration of the
graph noise (RTN33–35) on a picosecond timescale (see Supplementary sublattice magnetisation vectors and the net magnetisation for the

Nature Communications | (2023)14:7651 4

Article https://doi.org/10.1038/s41467-023-43318-8

1 a T g
298.24 K

0 0 T < TL

-1
1 301.54 K T h

T = TL
mc/mS (x10 )

0 0
-3

-1
1 302.88 K T i

0 0 T ⪆ TL

-1
1 303.46 K T j

0 0 T >> TL

-1
E
200 400 600 800 1000 1200 Etherm
time (ps)
b T=2 c T=3 d T=3 e T=3 f
98.2 01.5 02.8 11.00
4K 4K 8K K

Fig. 5 | Picosecond random switching in Sm0.7Er0.3FeO3. a Simulated time traces (red) for the energetically degenerate ±θ states. The canting angle of the sublattice
of the c-axis component mc of the magnetisation normalised to the saturation magnetisation vectors and the thus resulting net magnetisation is highly exag-
magnetisation mS for multiple temperatures near spin reorientation in gerated for visibility. g–j Orthoferrite potential landscape for different tempera-
Sm0.7Er0.3FeO3. The first 50 ps are not shown, because here the system still equi- tures across the SRT. The red-dotted line indicates the thermal energy Etherm of the
librates from the initial conditions. b–e Simulated trajectories (purple lines) of the system. At temperatures T < T L (g), the potential exhibits a parabolic shape and the
normalised magnetisation (red arrow) for multiple temperatures near spin reor- fluctuations are restricted to a single minimum around θ = 0. Slightly above T L
ientation in Sm0.7Er0.3FeO3. At temperature T ⪆ T L,sim (d) switching events are (h, i), a double-well potential develops and the particle randomly switches between
recorded in the c-direction. The precession cones of the qF mode magnon are the minima located at ±θ. For T U > T ≫ T L (j), the energy barrier between the
indicated in yellow. f Illustration of the sublattice magnetisation vectors in minima is larger than the thermal energy of the system and no more switching
Sm0.7Er0.3FeO3 magnetisation (blue, mint, green, yellow) and the net magnetisation events occur.

switching ±θ states is shown in Fig. 5f. As the temperature is elevated physical origin of the experimental LF feature. It should be noted that
even further (Fig. 5e), the switching probability becomes lower until no stochastic physical systems exhibiting RTN have lately gathered pro-
more switching occurs. Here, the up-state is always preferred because minence as a key enabler for probabilistic computing. For this pur-
of the initial conditions of the simulation. When approaching the pose, systems showing high-frequency RTN are desired to implement
upper threshold temperature T U, sim , switching is observed in the a- faster computing times and higher precision33–35. While RTN on elec-
projection due to anisotropy softening along this direction. Beyond tronic timescales was intensively studied for decades in systems
T U, sim , the equilibrium magnetisation becomes parallel to the c-axis exhibiting charge carrier traps, e.g., in commercial semiconductor
(not shown) and the qF mode contribution to the noise vanishes (see structures39,40, the fastest RTN device reported so far remained in the
Suppl. Figs. 2 and 4). nanosecond regime exploiting in-plane magnetic tunnel junctions41,42.
The physical picture of the RTN dynamics can be clearly under- To the best of our knowledge, the picosecond RTN reported here
stood by a model considering the stochastic switching between two marks the record-switching speed to date. We attribute the high rate in
energetically degenerate quasi-equilibrium states, which manifest as Sm0.7Er0.3FeO3 to the magnon frequency in the sub-THz region which
±θ rotation states of the equilibrium magnetisation due to the SRT results in shorter intervals between switching attempts as compared to
(see Fig. 5g–j). In Sm0.7Er0.3FeO3, the free energy density exhibits a conventional ferromagnets42,43. This result further highlights the cap-
parabolic shape for T < T L , whereas for T > T L it evolves into a double- ability of ultrafast SNS as a unique tool for probing the stochastic
potential well with minima at ±θ24. The rotation angle θ and the height dynamics near a magnetic phase transition.
of the potential barrier separating the two minima gradually increase In summary, we demonstrate the time-domain observation of sub-
until T = T U 38. Random switching between ±θ states occurs when the THz magnon fluctuations in the antiferromagnetic orthoferrite
height of the potential barrier is low compared to the thermal energy Sm0.7Er0.3FeO3 near the SRT. The drastic increase of amplitude and
of the system (Fig. 5i, T ⪆ T L ). Further increase of the temperature in coherence time within the SRT together with the low-frequency peak
the order of 1 K significantly changes the barrier height, while the observed in the spectrum is direct evidence of ultrafast random spin
thermal energy only changes marginally. As a result, the average life- switching between two equilibrium states of the magnetic free energy.
time τ on each quasi-equilibrium state increases, and the switching The random spin switching speed reported here is the fastest ever
probability declines. Note that this model even reproduces our marked and may provide a key ingredient for ultrafast probabilistic
observation that the temperature at which the time-zero amplitude of computing operating at THz frequencies. Our experimental concept is
the autocorrelation becomes maximal (Fig. 3b, d) is slightly higher not only limited to orthoferrites but also applicable to a wide range of
than T L (Fig. 4b, d). correlated magnetic systems. Furthermore, our results shed new light
The evident connection between RTN in the simulated time traces on THz magnonics in AFMs, where the influence of incoherent spin
and the LF peak in the spectra firmly establishes picosecond RTN as the dynamics has largely been dismissed. In future works, combining the

Nature Communications | (2023)14:7651 5

Article https://doi.org/10.1038/s41467-023-43318-8

proposed concept with coherent excitations is expected to enable the reference. Close to the SRT, the combined effects of magnon softening
seamless observation of spin dynamics from thermal equilibrium to and thermally induced random switching result in the magnetisation
nonequilibrium states. noise not fully decaying to zero within the observation window. In this
case, the abovementioned linear baseline subtraction cannot be used
Methods without distorting the waveform. From 307.15 K upwards, we therefore
Experiment and data post-processing use the average baseline of all waveforms recorded below 307.15 K as
This study exploits a modelocked Er:fibre laser system emitting pulses the reference for our linear baseline subtraction. Furthermore, in the
of a width of 150 fs, a central wavelength of 1.55 µm and a total energy of SRT region above 307.15 K, a strong baseline increase is observed, the
5 nJ at a repetition rate of 40 MHz. This output is frequency doubled in a amplitude of which is strongly temperature dependent and most
periodically poled lithium niobate (PPLN) crystal, reaching a transpar- prominent at around 311 K. The baseline increase is asymmetric around
ency window of the orthoferrites44. Subsequently, the frequency- Δt = 0 and therefore must result from external factors, such as a slight
doubled beam is spectrally split by a dichroic mirror, resulting in two misalignment of the delay stage. In the temperature region between
linearly polarised, spectrally distinct femtosecond pulse trains. The 310.35 K and 311.95 K, no signature of correlated noise was observed
central wavelengths of the two beams are 767 nm and 775 nm with 3 to because of the large asymmetric background. Hence no meaningful
4 nm bandwidths, respectively. One of the probe pulses is time-delayed evaluation could be carried out and we consequently neglect the data
by Δt with an optical delay stage. After spatial recombination by another in this region in our discussion. In all other reliable datasets recorded
dichroic mirror, the pulses are focused to a spot diameter below 2 µm for temperatures larger than 307.15 K where a slight asymmetric
on the orthoferrite sample with a transmissive objective lens of a background is superimposed with the correlated spin noise, a slope
numerical aperture of 0.4. The sample is a d = 10 µm thick, c-cut plate of correction is employed to remove the asymmetry. To account for
single-crystal Sm0.7Er0.3FeO3. The two probe pulses are linearly polar- potential artefacts in the correlated spin noise introduced by this
ised along the a-axis. Upon transmission of the pulses through the procedure, we assign a relative uncertainty of 10% of the amplitude
sample of thickness d at times t and t + Δt, transient magnetisation determined at each time delay.
fluctuations δM c ðtÞ parallel to the propagation direction of the pulse
trains introduce polarisation noise δαðtÞ / V ðλÞ  d  μ0  δM c ðtÞ via the Atomistic spin model simulations
magneto-optic Faraday effect, where V ðλÞ is the wavelength depen- Sm0.7Er0.3FeO3 is modelled as a generic orthoferrite with magnetic
dent Verdet constant and μ0 is the vacuum permeability. After colli- moments on the Fe sites only. These are treated as classical vectors Si
mating them with an additional lens, the two probe beams are spatially on a simple cubic lattice with four sublattices. The rare earth moments
separated with a dichroic mirror and sent into individual analysers to order only at very low temperatures of typically T < 6 K and are, hence,
monitor their polarisation. Each detector consists of a half-wave plate neglected. In orthoferrites, the nearest neighbour exchange constant
(HWP), a Wollaston prism and a pair of balanced photo diodes. The J 1 is of the order of 20 meV leading to an antiferromagnetic order
angles of the HWPs are set to compensate for stationary components of with a Néel temperature in the range of 630 K, whereas the next
the polarisation rotation induced by, e.g., biaxial birefringence of the nearest neighbour exchange constant J 2 is much smaller and of the
sample45. The Wollaston prisms (WP) separate p- and s-polarised com- order of 1 meV46–48. A reorientation transition can be modelled by
ponents of the probes. The intensity difference of the polarisation different thermally stable anisotropies, as it is done in ref. 49. Here the
components is then detected in separate balanced photodetectors reorientation transition is due to a competition of second-order on-site
(BPD) and amplified with transimpedance amplifiers (Amp), respec- anisotropy, favouring the [001] direction, and a second-order two-site
tively. Subsequently, the signals pass a 20 MHz bandpass filter (BP) and anisotropy, favouring the [001] plane. The low-temperature state is
are demodulated at the first sub-harmonic of the laser repetition rate dominated by the on-site anisotropy whereas the high-temperature
(20 MHz) with a radio-frequency lock-in amplifier (UHFLI, Zurich state is dominated by the more thermally stable two-site anisotropy.
Instruments)21. The outputs of these demodulation channels now With these two anisotropies, one would obtain a first-order reor-
comprise the polarisation noise amplitudes δα (t) and δαðt + ΔtÞ, ientation transition. By adding a small cubic anisotropy preferring the
respectively, as well as uncorrelated components dominated by the [111] direction, one obtains a second-order reorientation transition in
shot noise of the photons in the probes. In a last step, both demodu- agreement with experiments. The strength of the cubic anisotropy
lation signals are multiplied in real-time inside the lock-in amplifier. This determines the width of the reorientation transition. The weak ferro-
product is averaged over approximately 108 pulse pairs per time delay magnetism caused by the canting of the antiparallelly aligned sub-
Δt

to effectively  yield the cross-correlation of Faraday noise lattice magnetisations originates from the oxygen-mediated
δαðt Þδαðt + Δt Þ . In this way, we extract the tiny portion of correlated Dzyaloshinskii–Moriya interaction (DMI).
fluctuations originating from the sample response out of a much larger Consequently, the Hamiltonian of our model reads
uncorrelated background. At the same time, the two-colour scheme
  X υ X υ X
avoids detrimental interference effects at short delay times, providing H Si =  J 1 Si Sjυ  J 2 Si Sjυ  ευηλ Dυij Sηi Sλj
sensitive access to high frequencies. Furthermore, it enables the two hi, ji hi, ji hi, ji
X X υη X υυηη ð1Þ
probe beams to collinearly overlap before the objective lens, allowing  κ υη Siυ Sjη  κ2 Siυ Siη  κ 4 Siυ Siυ Siη Siη
for the beam spots on the sample surface to maximally mode matched. hi, ji i i
This is crucial to obtain magnon correlation signals with sufficiently
strong amplitude at measurable levels. using the Einstein notation where i and j denote the site indices
Contrary to previous works21, the two-colour scheme also allows and υ,η and λ the Cartesian directions. The first two double sums
us to sample the Faraday rotation with two separate BPDs. By this way, correspond to the nearest neighbour interaction with J 1 = 22.32 meV
our time resolution is not limited by the deadtime of the photo- and the next nearest neighbour interaction with J 2 = 1.4 meV. The
detector, and we can exploit the full sub-picosecond resolution given DMI is included in the nearest neighbour shell with DMI vectors having
by the duration of the probe pulses. This fact enables us to analyse the a length of 0.18 meV for in-plane interactions and 0.25 meV for out-of-
changes of spin noise over a magnetic phase transition in a correlated plane interactions. The directions of the DMI vectors can be obtained
magnetic material with significantly shorter spin dephasing times. using the symmetry rules of ref. 26. In the Hamiltonian Dυij corresponds
The raw waveforms are post-processed with a third-order to the υ-component of the DMI vector describing the interaction
Savitzky-Golay filter for smoothing, as well as a linear baseline sub- between the spins on lattice site i and j. The two-site anisotropy is also
traction, where the average values of the first and last 5 ps serve as a included in the first shell with κ zz =  0:1255 meV, leading to an easy

Nature Communications | (2023)14:7651 6

Article https://doi.org/10.1038/s41467-023-43318-8

xy-plane. The second-order on-site anisotropy is κ zz 2 = 0:905 meV, It should be noted that because of the complexity of our simula-
which results in an easy z-axis. There is also a small in-plane con- tion, it is practically challenging to find a set of spin model parameters
yy xy yx
2 = κ 2 = 0:015 meV and κ 2 = κ 2 =  0:015 meV. The
tribution with κ xx that precisely fit all experimental measurements simultaneously, like
latter contribution is not necessary for the reorientation transition but the Néel temperature, reorientation temperatures and eigen-
lifts the degeneracy of the spins in the xy-plane. The fourth-order on- frequencies. Furthermore, additional effects that should be present in
site anisotropy is κ xxyy
4 = κ xxzz
4 = κ yyzz
4 = 0:036 meV. With these para- the experiment, such as the influence of domain states and possible
meters, the model undergoes a reorientation transition between 302 K crystal defects, are not covered by our simulations. The quantitative
and 322 K, where the Néel vector rotates continuously from the discrepancy found in the simulation results from the experimental
z-direction to the x y-direction and the magnetisation from the values, such as the SRT temperature region and the qF mode fre-
x y-direction to the z-direction (see Supplementary Information SI2). quency, stems from the above technical limitation.
Note that the x, y and z coordinate axes of the Hamiltonian are parallel
or antiparallel to the connection lines of the iron atoms, but not par- Z-scan
allel to the crystallographic axes of an orthoferrite a,b,c. The crystal- When a sample of thickness d is placed into the focus of a Gaussian
lographic unit vectors are given by: laser beam, the illuminated volume forms a position-dependent
hyperboloid:
1   1
ea = pffiffiffi ex  ey , eb = pffiffiffi ðex + ey Þ, ec = ez , ð2Þ " #
3
2 2 d =12 + dz 2
Ωðz Þ = πw20 d + , ð9Þ
z 2R
so that the a-direction corresponds to the x y-direction, the b-direction
to the xy-direction and the c-direction equals the z-direction. where z is the position of the sample along the axis of light propaga-
To simulate the dynamics of the magnetic moments, the sto- tion relative to the laser focus, w0 is the beam radius at its narrowest
πw2
chastic Landau-Lifshitz-Gilbert equation31,32 is used which reads: point and z R = λ 0 is the wavelength λ dependent Rayleigh length. We
estimate z R to be 7 µm. In conventional SNS, the statistical fluctuation
d γ   of N spins in the probing volume Ω results in Faraday noise of
the order
S =   S × Hi + α d Si × Hi ð3Þ 
dt i μS 1 + α 2d i δα / p1ffiffiffi
Ω
. For a correlated noise amplitude, we hence expect δα2 / Ω1
at Δt = 0.
with Inserting the formula for ΩðzÞ then yields the following fitting
function for the correlated noise amplitude at Δt = 0 as a function of
∂H lateral sample position (Fig. 2b):
Hi =  + ξi ð4Þ
∂Si
A
h 3
i ð10Þ
and a thermal Gaussian white noise term with d =12 + dz 2
πw20 d+ z 2R


 2μ α k T
ξ i ðtÞ = 0, ξ iυ ðt Þξ jη ðt 0 Þ = S d B δ ij δ υη δðt  t 0 Þ ð5Þ
γi where A is a proportionality constant.

with the dimensionless damping parameter α d = 0:0002. The gyro- Spectra

magnetic ratio is set to the value of a free electron γi = 1:76086 × 1011 Hz
T,
The experimental noise waveforms are interpolated, zero-padded and
the magnitude of the magnetic moments is μS = 3:66μB . On this basis, smoothed. Subsequently, a FFT algorithm is harnessed to obtain high-
the time evolution of a system of 1923 spins is numerically calculated resolution spectra.
via the stochastic Heun method. Our main output is the magnetisation The spectral features are analysed using a double Lorentzian peak
mðtÞ, which reads: fit. In spectra where the qF mode and the LF feature (see Fig. 4b)
overlap, the centre frequencies are obtained by estimating the zero-
1X crossings of the second derivative.
mðt Þ = S: ð6Þ
N i i The simulated spectra are interpolated and smoothed, as well.
The spectral features are then analysed using a triple Lorentzian peak
N is the number of spins and Si denotes the spin at lattice site i. fit where the peaks correspond to the LF feature, the qF mode and the
Thus, mðt Þ is dimensionless and normalised to unity in the following. The qAF mode, respectively (see Suppl. Fig. 4a–c). We note that the
spectral noise amplitude P η (spectral power density) is calculated via: experimental and simulated spectra are acquired through different
analytical procedures. In the experiment, we directly measure the
* 2+ time-domain autocorrelation waveform and obtain the spectra via
mTη ð f Þ
Pηð f Þ = ð7Þ Fourier transforming after filtering and background subtraction as
T described above. On the other hand, in the simulation, we first calcu-
late the magnetisation dynamics and Fourier transform them into the
with the Fourier transform: spectral domain. The spectra shown in Fig. 4c are the averages of
20 spectra simulated in such a way.
Z T
mTη ð f Þ = mη ðtÞei2πf t dt: ð8Þ
Data availability
0
The datasets generated in the present study are available from the
Furthermore, the spectral amplitude is averaged over 20 to corresponding author upon reasonable request.
30 simulation runs. The time-dependent correlated noise amplitude
(autocovariance) is determined by the inverse Fourier transform of the References
spectral noise amplitude, taking advantage of the Wiener-Khinchin 1. Moriya, T. Spin Fluctuations in Itinerant Electron Magnetism
theorem. (Springer International Publishing, Cham, 20).

Nature Communications | (2023)14:7651 7

Article https://doi.org/10.1038/s41467-023-43318-8

2. Weller, D. & Moser, A. Thermal effect limits in ultrahigh-density 29. Zhao, X. et al. Crystal growth and spin reorientation transition in
magnetic recording. IEEE Trans. Magn. 35, 4423–4439 (1999). Sm0.4Er0.6FeO3 orthoferrite. Solid State Commun. 231-232,
3. Bauer, G. E. W., Saitoh, E. & van Wees, B. J. Spin caloritronics. Nat. 43–47 (2016).
Mater. 11, 391–399 (2012). 30. Yamaguchi, K., Kurihara, T., Minami, Y., Nakajima, M. & Suemoto, T.
4. Camsari, K. Y., Salahuddin, S. & Datta, S. Implementing p-bits With Terahertz time-domain observation of spin reorientation in ortho-
Embedded MTJ. IEEE Electron. Device Lett. 38, 1767–1770 (2017). ferrite ErFeO3 through magnetic free induction decay. Phys. Rev.
5. Falk, H. & Bruch, L. W. Susceptibility and fluctuation. Phys. Rev. 180, Lett. 110, 137204 (2013).
442–444 (1969). 31. Gilbert, T. L. Classics in magnetics a phenomenological theory of
6. Eckel, S., Sushkov, A. O. & Lamoreaux, S. K. Magnetic susceptibility damping in ferromagnetic materials. IEEE Trans. Magn. 40,
and magnetization fluctuation measurements of mixed gadolinium- 3443–3449 (2004).
yttrium iron garnets. Phys. Rev. B 79, 014422 (2009). 32. Nowak, U. Classical Spin Models. In Handbook of magnetism and
7. van Loosdrecht, P. H. M., Boucher, J. P., Martinez, G., Dhalenne, G. & advanced magnetic materials, edited by H. Kronmüller & S. S. P.
Revcolevschi, A. Magnetic fluctuation spectrum of CuGeO3: Raman Parkin (John Wiley & Sons, Chichester, England, 2007).
scattering. J. Appl. Phys. 79, 5395 (1996). 33. Borders, W. A. et al. Integer factorization using stochastic magnetic
8. Weber, M. C. et al. Raman spectroscopy of rare-earth orthoferrites tunnel junctions. Nature 573, 390–393 (2019).
RFeO3 (R =La, Sm, Eu, Gd, Tb, Dy). Phys. Rev. B 94, 214103 (2016). 34. Debashis, P. et al. Hardware implementation of Bayesian network
9. Yakhou, F., Létoublon, A., Livet, F., Boissieu, Mde & Bley, F. Magnetic building blocks with stochastic spintronic devices. Sci. Rep. 10,
domain fluctuations observed by coherent X-ray scattering. J. 16002 (2020).
Magn. Magn. Mater. 233, 119–122 (2001). 35. Kaiser, J. & Datta, S. Probabilistic computing with p-bits. Appl. Phys.
10. Berski, F., Kuhn, H., Lonnemann, J. G., Hübner, J. & Oestreich, M. Lett. 119, 150503 (2021).
Ultrahigh bandwidth spin noise spectroscopy: detection of large 36. Balakrishnan, V. Mathematical Physics: Applications and Problems
g-factor fluctuations in highly-n-doped GaAs. Phys. Rev. Lett. 111, (Springer Nature, 2020).
186602 (2013). 37. Machlup, S. Noise in semiconductors: spectrum of a two‐parameter
11. Crooker, S. A., Rickel, D. G., Balatsky, A. V. & Smith, D. L. Spectro- random signal. J. Appl. Phys. 25, 341–343 (1954).
scopy of spontaneous spin noise as a probe of spin dynamics and 38. Suemoto, T., Nakamura, K., Kurihara, T. & Watanabe, H.
magnetic resonance. Nature 431, 49–52 (2004). Magnetization-free measurements of spin orientations in ortho-
12. Hübner, J. et al. Rapid scanning of spin noise with two free running ferrites using terahertz time domain spectroscopy. Appl. Phys. Lett.
ultrafast oscillators. Opt. Express 21, 5872–5878 (2013). 107, 42404 (2015).
13. Römer, M., Hübner, J. & Oestreich, M. Spin noise spectroscopy in 39. Hung, K. K., Ko, P. K., Hu, C. & Cheng, Y. C. Random telegraph noise
semiconductors. Rev. Sci. Instrum. 78, 103903 (2007). of deep-submicrometer MOSFETs. IEEE Electron Device Lett. 11,
14. Zapasskii, V. S. et al. Optical spectroscopy of spin noise. Phys. Rev. 90–92 (1990).
Lett. 110, 176601 (2013). 40. Ralls, K. S. et al. Discrete resistance switching in submicrometer
15. Zapasskii, V. S. Spin-noise spectroscopy: from proof of principle to silicon inversion layers: individual interface traps and low-
applications. Adv. Opt. Photon. 5, 131 (2013). frequency (f1?) noise. Phys. Rev. Lett. 52, 228–231 (1984).
16. Bloch, F. Nuclear Induction. Phys. Rev. 70, 460–474 (1946). 41. Safranski, C. et al. Demonstration of nanosecond operation in sto-
17. Keffer, F. & Kittel, C. Theory of antiferromagnetic resonance. Phys. chastic magnetic tunnel junctions. Nano Lett. 21,
Rev. 85, 329–337 (1952). 2040–2045 (2021).
18. Jungwirth, T., Marti, X., Wadley, P. & Wunderlich, J. Anti- 42. Hayakawa, K. et al. Nanosecond random telegraph noise in in-plane
ferromagnetic spintronics. Nat. Nanotechnol. 11, 231–241 (2016). magnetic tunnel junctions. Phys. Rev. Lett. 126, 117202 (2021).
19. Siegrist, F. et al. Light-wave dynamic control of magnetism. Nature 43. Kanai, S., Hayakawa, K., Ohno, H. & Fukami, S. Theory of relaxation
571, 240–244 (2019). time of stochastic nanomagnets. Phys. Rev. B 103, 094423 (2021).
20. Riek, C. et al. Direct sampling of electric-field vacuum fluctuations. 44. Wood, D. L., Remeika, J. P. & Kolb, E. D. Optical spectra of rare‐earth
Science 350, 420–423 (2015). orthoferrites. J. Appl. Phys. 41, 5315–5322 (1970).
21. Riek, C. et al. Subcycle quantum electrodynamics. Nature 541, 45. Usachev, P. A. Optical properties of thulium orthoferrite TmFeO3.
376–379 (2017). Phys. Solid State 47, 2292 (2005).
22. Benea-Chelmus, I.-C., Settembrini, F. F., Scalari, G. & Faist, J. Elec- 46. White, R. L. Review of recent work on the magnetic and spectro-
tric field correlation measurements on the electromagnetic scopic properties of the rare‐earth orthoferrites. J. Appl. Phys. 40,
vacuum state. Nature 568, 202–206 (2019). 1061–1069 (1969).
23. Yang, L. et al. Two-colour spin noise spectroscopy and fluctuation 47. Treves, D. Studies on orthoferrites at the Weizmann Institute of
correlations reveal homogeneous linewidths within quantum-dot Science. J. Appl. Phys. 36, 1033–1039 (1965).
ensembles. Nat. Commun. 5, 4949 (2014). 48. Shapiro, S. M., Axe, J. D. & Remeika, J. P. Neutron-scattering studies
24. Fitzky, G., Nakajima, M., Koike, Y., Leitenstorfer, A. & Kurihara, T. of spin waves in rare-earth orthoferrites. Phys. Rev. B 10,
Ultrafast control of magnetic anisotropy by resonant excitation of 4f 2014–2021 (1974).
electrons and phonons in Sm0.7Er0.3FeO3. Phys. Rev. Lett. 127, 49. Donges, A. et al. Magnetization compensation and spin reorienta-
107401 (2021). tion transition in ferrimagnetic DyCo5: Multiscale modeling and
25. Dzyaloshinsky, I. A thermodynamic theory of “weak” ferromagnet- element-specific measurements. Phys. Rev. B 96, 024412 (2017).
ism of antiferromagnetics. J. Phys. Chem. Solids 4, 241–255
(1958). Acknowledgements
26. Moriya, T. Anisotropic superexchange interaction and weak ferro- This research was supported by the Overseas Research Fellowship of the
magnetism. Phys. Rev. 120, 91–98 (1960). Japan Society for the Promotion of Science (JSPS), Zukunftskolleg Fel-
27. Herrmann, G. F. Resonance and high frequency susceptibility in lowship from the University of Konstanz, JSPS KAKENHI (JP21K14550,
canted antiferromagnetic substances. J. Phys. Chem. Solids 24, JP20K22478, JP20H02206) and by the Deutsche For-
597–606 (1963). schungsgemeinschaft (DFG, German Research Foundation)—Project-ID
28. Herrmann, G. F. Magnetic resonances and susceptibility in ortho- 425217212-SFB 1432. T.K. acknowledges Tohru Suemoto for fruitful
ferrites. Phys. Rev. 133, A1334–A1344 (1964). discussion.

Nature Communications | (2023)14:7651 8

Article https://doi.org/10.1038/s41467-023-43318-8

Author contributions reviewer(s) for their contribution to the peer review of this work. A peer
T.K. and A.L. conceived the experiment. M.A.W., A.H. and T.K. developed review file is available.
the experimental system, performed the measurements, and analysed
the data. J.S., T.D., M.E. and A.D. performed the numerical simulations Reprints and permissions information is available at
under the supervision of U.N. M.N. produced the specimen. T.K., A.L. and http://www.nature.com/reprints
S.T.B.G. co-supervised the project. M.A.W. and T.K. wrote the manu-
script with the help of all co-authors. Publisher’s note Springer Nature remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.
Funding
Open Access funding enabled and organized by Projekt DEAL. Open Access This article is licensed under a Creative Commons
Attribution 4.0 International License, which permits use, sharing,
Competing interests adaptation, distribution and reproduction in any medium or format, as
The authors declare no competing interests. long as you give appropriate credit to the original author(s) and the
source, provide a link to the Creative Commons licence, and indicate if
Additional information changes were made. The images or other third party material in this
Supplementary information The online version contains article are included in the article’s Creative Commons licence, unless
supplementary material available at indicated otherwise in a credit line to the material. If material is not
https://doi.org/10.1038/s41467-023-43318-8. included in the article’s Creative Commons licence and your intended
use is not permitted by statutory regulation or exceeds the permitted
Correspondence and requests for materials should be addressed to use, you will need to obtain permission directly from the copyright
T. Kurihara. holder. To view a copy of this licence, visit http://creativecommons.org/
licenses/by/4.0/.
Peer review information Nature Communications thanks Ileana-Cristina
Benea-Chelmus, Dmitry Turchinovich and the other, anonymous, © The Author(s) 2023

Nature Communications | (2023)14:7651 9