a Experimental setup of the AI-TA. The spectra of ML1 and ML2 are shown in Supplementary Fig. 1. ML mode-locked laser, DC dichroic mirror, BBO beta barium-borate crystal, PD photodiode, BS beam-splitter, PM parabolic mirror, BPD balanced photodiode, LD linear delay stage. The self-referenced interferogram is acquired through a Mach–Zehnder configuration and detected using a balanced photodetector. b Definition of the time variables in relation to the laboratory time, t_lab. By using time-division multiplexing, the laboratory time can be processed into pump-probe delay time (t_pump-probe), probe autocorrelation time (t_auto), and reaction time (t_react). c An example of a transient signal depending on the t_pump-probe, which is converted from the laboratory time using the down-conversion factor, ??. Here, when ?? = 80.0 MHz and Δ? = 51.2 Hz, 187.5 μs of laboratory time is required to complete a single scan of pump-probe delay from 0 to 120 ps. d The interferogram was acquired by collecting signals sampled at the rate of Δ? = 51.2 Hz, while the linear delay stage is moving at a constant velocity of 2 μm/s. The laboratory time in this case can be converted to the t_auto using the Doppler shift factor ??Ds. Fast Fourier transform of the data yields the probe spectrum with respect to the wavelength, λ_probe. e Experimental schematic of the ultrafast pump-probe investigation of photoinduced reactions of perovskite nanomaterials. Since the pump and the probe lights work as reaction stimulants that induce photo-substitution or photon-driven transformation, the data acquisition measurement time can be considered as t_react.

The approximation in Eq. (1) is valid due to the significantly small detuning frequency Δ? compared to the repetition frequency (??), i.e., Δ????. An important point to note is that the down-conversion factor ??=Δ?/?? can be precisely controlled by a GPS-disciplined Rb atomic clock, ensuring consistent generation of time steps^37,38. Additionally, when the separated pump and probe pulses overlap in space and time on another BBO nonlinear crystal for sum-frequency generation, distinct trigger signals are produced at a rate of 1/Δ?, marking the initiation of the pump-induced probe response.

In conventional transient absorption techniques, obtaining wavelength-resolved information in the time domain requires a monochromator or spectrograph equipped with a charge-coupled device. In our approach, we accomplish this by generating a self-referenced interferogram (see Supplementary Fig. 1 for the pump and probe spectra). The position of the linear stage corresponds to the time delay between split probe pulses, so scanning the stage at a constant velocity produces an autocorrelation signal (Supplementary Fig. 2). Using frequency domain detection with a single point detector offers advantages in terms of effective noise filtration and signal collection efficiency.

In our setup, the probe intensity is detected by a fast digitizer operating at a sampling rate of ??, synchronized with the repetition rate of the probe laser. The probe signal contains information about both the linear absorbance of the sample and its pump-induced nonlinear responses. Since the raw data consists of an array of voltage signals that are time-tagged by optical trigger signals, time-division multiplexing can be applied to acquire information from segments of different temporal lengths^3,38. In Fig. 1b, to avoid confusion in the upcoming discussions, we define four temporal variables involved in the interpretation of AI-TA data. The time tag from the digitizer represents the real-world time measured during the experiment, also referred to as laboratory time (t_lab). Given that our system records data every 12.5 ns, the temporal span over a 1-h experiment encompasses eleven orders of temporal dynamic range. This extensive time span allows the analysis of additional time-dependent variables that influence the probe response if they operate in temporally distinct regimes. Based on this, laboratory time is divided into three segments in this study: pump-probe delay time, probe autocorrelation time, and reaction time.

First and foremost, using the relationship (Eq. 1), we can convert the laboratory time, recorded at an interval of 1/?? to the pump-probe delay time (t_pump-probe) by multiplying it with the down-conversion factor ??. In Fig. 1c, the pump-probe response of the PeNC solution is shown as a function of t_pump-probe. The starting time, or time zero, of the signal is defined by the optical trigger signal. Then, the pump-induced transmission change is calculated by dividing the voltage signals recorded after the trigger by those recorded before the trigger.

The interferogram generated by scanning the linear stage contains information about the frequency of the electromagnetic field, which can be easily accessed through a fast Fourier transform (FFT). The method requires the data points of the interferogram to be evenly spaced in time. Provided that the linear stage moves at a fixed velocity, this can be achieved by using the trigger rate, Δ?, as the sampling frequency, as shown in Fig. 1d. The time interval between each sample (Δ?) can be calculated using the Doppler shift relation ?Ds=2??/?, where ω is the angular frequency of the probe field, ? is the stage speed, and ? denotes the speed of light. Then, we can deduce that Δ?=2?/?Δ?=??Ds/Δ?. The Doppler shift factor, ??Ds=2?/?, is multiplied by t_lab to calculate the effective time between split probe pulses, known as the probe autocorrelation time (t_auto).

From here, we can extract frequency-resolved light intensity via FFT, which can also be expressed in terms of probe wavelength (λ_probe) using the speed of light ?. Since each data point is annotated with a corresponding t_lab in relation to trigger time zero, our approach allows simultaneous measurement of decay lifetimes for each spectral component, provided the linear stage scanning speed is significantly slower than the rate of pump-probe delay generation time t_pump-probe. More detailed and advanced theoretical descriptions and comprehensive analyses of experimental parameters related to this technique have been presented and demonstrated in our previous works^3,7,39. Our implementation, AI-TA, also features a high duty cycle and low energy per pulse, rendering it particularly advantageous for scenarios involving photoreactions and photodamage.

Figure 1e presents two target systems for the ultrafast pump-probe investigation in this study and defines the final time parameter, i.e., reaction time (t_react). This experimental design is conceived to integrate both synthesis and analysis, with the pump and probe lights serving as both reaction stimulants and observation tools simultaneously. Meanwhile, the rapid DAQ procedure of our AI-TA setup enables data collection before significant changes occur in the perovskite nanomaterials. Since the data from each linear stage scan are saved as discrete files with corresponding timestamps, we can determine the interval between subsequent measurements, which is less than 10 s (Supplementary Fig. 3). Thus, this measurement time can be considered as the reaction time in our analysis. Note that t_react has fewer significant digits, as it is not governed by the precision of an atomic clock.

AI-TA measurement of CsPb(Br/Cl)₃ in toluene

Before proceeding to in situ analysis, we first analyzed the transient absorption spectrum of the nanocrystals dissolved in toluene to investigate the dynamics involved in the relaxation of PeNCs. Since this solvent does not contain halide agents, halide exchange is avoided. The resulting TA spectrum, shown in Fig. 2a, represents an average of 200 interferogram scans. We applied singular value decomposition (SVD) and selected the first two eigenmodes as the principal representations of the states contributing to the overall transient absorption signal (Supplementary Fig. 4).

Fig. 2: Photodynamics of PeNCs.

a The transient absorption spectrum of CsPb(Br/Cl)₃ PeNCs dissolved in toluene measured up to 1 ns. The spectrum is the averaged result of 200 scans. b Decay-associated spectra and c time-resolved decay profiles of the first two modes from singular value decomposition analysis. The second mode decays quickly within 1 ps, and the first mode persists over a longer period, which has been analyzed with a bi-exponential fit. d List of various dynamics observable with the AI-TA technique spanning a broad dynamic range from sub-picosecond to minutes. The initial pump pulse instantaneously generates holes and electrons. As the excitation energy exceeds the bandgap energy, any excess energy is quickly dissipated as the hot carriers relax to the band-edge state. This transition is accompanied by bandgap renormalization. Due to the presence of trap states near the band edge, the conduction band is partially depopulated with a short lifetime of ~20 ps. The shallow nature of this charge trap allows interaction with other excitons, resulting in a charged exciton recombination with a lifetime of 0.35 ns. During the experiment photoinduced halide substitution occurs with timescales of minutes.

The decay-associated spectrum (DAS) of the first mode exhibits a strong negative optical density near 435 nm, indicative of a bandgap photo-bleaching (PB) signal, with a possible contribution from stimulated emission (Fig. 2b). Additionally, distinctive peaks of excited-state absorption (ESA) appear below 425 nm, resulting from the quantum confinement effect in our PeNCs^40,41,42 (Supplementary Fig. 5).

A bi-exponential fit to the time profile of the first mode yields two lifetimes: 20.2 and 345.5 ps (Fig. 2c). The shorter component, decaying within tens of picoseconds, is typically associated with biexciton decay or the Auger recombination process^{40,43,44,45,46}. However, given our exceptionally low pump energy of 0.5 nJ per pulse, this is better described by the charge carrier trapping process originating from shallow defects on the surface^47,48,49. The intrinsically low power of our spectroscopic apparatus allows us to bypass the consideration of multi-exciton formation, which can occur when using high-power pumps and requires many-body analyses to extract one-exciton dynamics from the complicated TA signals⁵⁰.

On the other hand, lifetimes on the order of hundreds of picoseconds are primarily associated with trap-mediated decay or recombination of charged excitons, known as trions^44,46. It is important to note that the decay of unaffected excitons typically occurs on the order of a few nanoseconds^40,44. The decay of charged excitons is observed even under low fluence, as their formation is closely tied to charge traps in surface defects^51,52. Both observations are consistent with the high trap density of chloride-based perovskite nanomaterials, which creates a significant barrier to their application as blue light emitters⁵¹.

For the second mode, its time profile decays within one ps. Such a sub-picosecond transition is often a complex signal and difficult to identify due to the concurrent contributions of coherent artifacts, such as scattering signals, within this temporal range⁵³. The DAS of the second mode shows a negative peak at an energy level higher than the PB peak of the first mode, which is a characteristic of hot carrier thermalization and relaxation^45,54. The pronounced ESA feature observed here is linked to the bandgap renormalization process^41,43. Thus, we interpret the second mode as indicative of the initial excited state undergoing intra-band relaxation.

An overview of the discernible kinetics achieved through our technique is presented in Fig. 2d based on distinct lifetimes and spectral features. The shorter PB decay originates from the emptying of the conduction band-edge to shallow traps, while the longer PB decay is induced by the recombination of charge carriers to the ground state. The radiative decay lifetime is confirmed through time-resolved fluorescence measurements (Supplementary Fig. 6).

Our unique contribution to the study of photodynamics is shown in the final panel of Fig. 2d, which will be discussed in the following sections. This portion illustrates the photoinduced substitution of the halide composition of PeNCs, occurring over an extended duration, spanning minutes. The figure is based on reports suggesting that the impact of halide composition primarily affects the valence band, driven by the interaction between hot phonons and holes⁵⁵. Our AI-TA technique enables the observation and analysis of the distinct physical properties exhibited by each composition generated through light exposure.

Photoinduced halide substitution of CsPbX₃ nanocrystal

The first subject of our study involves PeNCs with a mixed halide composition, specifically CsPb(Br/Cl)₃, dissolved in a chloroform/toluene mixture. Upon irradiation, the bromide ions in the PeNCs are substituted with chloride ions. As described earlier, this anion-exchange reaction occurs as a result of the photoinduced electron transfer process, in which the PeNCs extract halides from the surrounding solvents, leading to the formation of a new composition. The overall process can now be summarized by the following photochemical equation:

where PeNC, denoted as CsPb(Br↓/Cl↑)3, differs from CsPb(Br/Cl)3 due to the substitution of Br with Cl.

Since the equilibrium between the nanocrystals and the solvated halide anions is reached almost instantly, within a few seconds¹⁸, continuous DAQ under these conditions allows for in situ observation of the transforming PeNCs. The high-speed DAQ enabled by AI-TA significantly mitigates the influence of the pump and probe on the system during each scan. Remember that each interferogram scan, including the data-saving process, takes less than 10 s, which is a substantial improvement over conventional techniques. However, due to inherent noise in the data from a single interferogram, averaging is necessary (Supplementary Figs. 7 and 8). The transient mapping shown afterward represents an average of a few scans, culminating in an approximate acquisition time of 1 min, which is required to delineate the transient features.

Figure 3a shows the absorption and emission spectra of PeNC in toluene/chloroform, before and after 20 min of AI-TA experimentation, and Fig. 3b depicts the spectral evolution of transient absorption spectra, reflecting the first and the last minute of the experiment. Both the absorption edge and the PB peak initially appeared around 440 nm but shifted to 430 nm, indicating a change in bandgap energy levels. The PB peak positions closely match the steady-state absorption and emission spectra measured before and after the experiment, as shown in the top panel of the figure. This measurement result provides evidence of a shifting bandgap while upholding the confined nature of the system.

Fig. 3: Photo-substitution of PeNCs.

a Comparison of absorption and emission spectra acquired before and after the AI-TA experiment of PeNC in chloroform. b Transient absorption of the first- and the last-minute scan of the experiment showing spectral evolution. The blueshift of the PB peak and corresponding steady-state spectra confirm the increased bandgap of the nanocrystal. c X-ray photoelectron spectroscopy (XPS) results for the PeNC sample before and after an AI-TA experiment of PeNC in chloroform. d TEM image of PeNCs measured before and after the AI-TA experiment of PeNC in chloroform, and the histogram showing size distribution. The scale bar of the image is 0.2 μm.

To verify compositional changes, the remaining stock solution and the solution collected after the AI-TA experiment were analyzed using other materials analysis methods. X-ray photoelectron spectroscopy (XPS) analysis, shown in Fig. 3c (wide scan in Supplementary Fig. 9), corroborates the compositional changes in the PeNCs before and after a separate 30-min AI-TA experiment (Supplementary Fig. 10). The peaks corresponding to the elements are magnified in the panel to highlight the compositional changes. The Cl/Br ratio shifts from 1.72:1.28 to 2.24:0.76 in the pre- and post-photo-substituted samples, respectively. Additionally, the size of the nanocrystals remains unchanged, suggesting that the observed shift in bandgap is not due to changes in the nanocrystal dimensions (Fig. 3d). The average sizes are 7.1 (±1.1) nm and 7.2 (±1.1) nm for the samples before and after photoinduced anion-exchange reaction, respectively. The results confirm that halide exchange occurred while preserving most of the macroscopic physical properties, although some physical degradation may have also occurred (Supplementary Fig. 11).

Analysis of the parameters involved in halide substitution

Before conducting an in-depth analysis of the AI-TA data scans, it is imperative to verify the orthogonality of the time parameters. The independence of these parameters is confirmed by monitoring the PB peak throughout the experiment, as depicted in Fig. 4a. The peak position of each successive data point falls within the error bars of both the preceding and following measurements, which validates the orthogonality between each transient absorption spectra map and the reaction time t_react. It is important to note that the reaction reaches completion after ~100 min, indicating a dynamic equilibrium between the halides in the PeNCs and those dissolved in the solvent. The photoreaction can also be characterized by an approximate lifetime of 46 min, as determined by fitting the peak position (Supplementary Fig. 12).

Fig. 4: Comprehensive mapping of AI-TA analysis results.

a Extended PB peak tracking over 3 h of the experiment. The inset zooms in to show that each sequential peak lies within the error bar of the prior measurement. b Tracking PB peaks of PeNCs in toluene over the course of the experiment, showing that halide solvent is required for the peak shift. c Effect of light on PB peak. Peak positions shift when a measurement is ongoing but remain the same when the experiment is paused, signaling that the shift is a light-induced process. The error bars in (a–c) are obtained by calculating the standard deviation of PB peak wavelengths collected from the t_pump-probe delay within 1 to 5 ps. d Analysis of the initial 1-min data acquired from AI-TA measurement. The trend along t_pump-probe axis provides insight into the kinetic parameters and lifetime of the photophysical processes of PeNCs, while the λ_probe-dependent spectrum provides spectral information of excited states. Along each axis, the first mode acquired from singular value decomposition is plotted. e Decay profiles of the first mode acquired from SVD of data segmented into 1-min intervals of t_react, showing its relationship with t_pump-probe. f λ_probe-t_react correlation map showing the evolution of spectral characteristics with the duration of the experiment. g Reduced core tensor for extended 2-hours-long measurement, which is the result of the high-order singular value decomposition (HOSVD) analysis of the AI-TA experiment data. Here, the index of each axis represents singular values arranged in descending order from the highest to the lowest, which are truncated to show only major modes involved in the dynamics.

The effects of the halide solvent and irradiation are validated through multiple control experiments. When toluene is used as the sole solvent, the peak position remains unchanged throughout the experiment (Fig. 4b). However, when chloroform is introduced to the solvent, the PB peaks exhibit a blueshift, highlighting the effect of solvent composition on the rate of photo-substitution within the PeNCs. The rate and final halide ratio of photoinduced exchange have been reported to depend on the ratio of toluene to halide solvents²³. Additionally, the concentration of PeNCs, and consequently their optical density, is shown to influence the kinetic rates (Supplementary Fig. 13a). The critical role of halide solvents is further demonstrated by the reverse substitution from chloride to bromide when dibromomethane is used as the halide solvent instead of chloroform (Supplementary Fig. 13b).

The role of light as a reaction trigger is confirmed through a temporary pause in measurements (Fig. 4c). While the peak position shifts during continuous irradiation, it remains unchanged in the absence of light. The observations corroborate both halide solvents and irradiation as key factors in the photoconversion process between halides in PeNCs and those bound to the solvent.

Furthermore, the use of a high-repetition-rate oscillator may induce heating in addition to irradiation for photoinduced chemical reactions. To assess the potential contribution of this effect to the substitution process, we conducted a control experiment by heating a PeNC/chloroform solution in the dark. The steady-state spectra (Supplementary Fig. 13c) remained unchanged, indicating that thermal effects are negligible.

Figure 4d presents processed data from the initial data segment, which is the averaged result of 10 interferogram scans and corresponds to less than 1 min of t_react, where the nonlinear response is plotted with respect to t_pump-probe and λ_probe. This yields information typically acquired from conventional transient absorption.

The inclusion of the t_react axis emphasizes the advantage of our approach over conventional methods, as it enables the identification of variations in spectral or temporal signatures that depend on t_react, which in this case is associated with photo-substitution. Figure 4e presents the decay profiles of the first mode, obtained from the SVD of data averaged over every ten scans, illustrating the relationship between t_pump-probe and t_react. The plot clearly shows that the ultrafast rise of the PB signal, identified earlier as hot carrier relaxation, becomes shorter as reaction time progresses. The thermalization of hot carriers is known to depend on the excess energy provided by the pump above the bandgap⁵⁴. Since our pump energy and fluence remain constant throughout the experiment, the observed behavior of the rise signal can be attributed to the bandgap energy increasing until it aligns with the pump range. It is important to note that this trend appears to plateau near 60 min, well before the PB peak shift halts. The early cessation of the change in thermalization rate may be due to the bandgap entering the pump spectrum range (Supplementary Fig. 1).

From these results, we can infer that having excess energy results in a longer thermalization and bandgap renormalization process in the sub-picosecond decay. Also, the exponential fitting of PB decay shows a slight decrease in lifetime as chloride photo-substitution progresses (Supplementary Fig. 14). Given that we have linked this parameter to the charge trapping of electrons in shallow traps, this relationship suggests that higher chloride content accelerates charge trapping. Interestingly, similar observations have been reported, where a faster charge-induced trapping rate was correlated with higher chloride content in PeNCs⁴⁷. Since vacancies are more abundant in higher chloride compositions, the increased trap density facilitates better separation of charge carriers such as electrons. Moreover, because PeNCs are continually exchanging ions with surrounding solvent molecules, the high variability in trap density and consequently the lifetimes can also be explained.

In Fig. 4f, a similar analysis of the DAS of the first mode against t_react shows that spectral characteristics are evidently influenced by the duration of the experiment. This λ_probe-t_react correlation map reveals a blueshift in the PB peak, reaching 403 nm, consistent with the peak analysis shown in Fig. 4a. From the evidence presented so far, we can confidently assert that the real-time evolution of spectral characteristics is a result of the substitution of bromide with chloride. Overall, this illustrates the multidimensional nature of a single AI-TA measurement under photo-responsive conditions, a capability made possible by our technique.

Building upon this, we can parameterize three orthogonal variables based on the temporal order of the time segments defined earlier. First, P₁ represents the transient time-resolved response after the incidence of the pump pulse, which is related to t_pump-probe. P₂ corresponds to bandgap energy-resolved information of PeNCs in terms of λ_probe, obtained from the FFT of the interferogram scan. Lastly, P₃ reflects the halide ratio of PeNCs, based on the unidirectional photo-substitution from bromide to chloride, and is associated with the total DAQ time divided into 1-min intervals, corresponding to t_react. By examining the dynamic responses between any pair of these parameters, we can identify correlations within the system. Since the raw data arrays can be structured as a rank-three tensor using three-fold time-division multiplexing, we can isolate states or modes related to a single parameter of P_n (n = 1, 2, 3) and establish the relationships between these states through tensor decomposition analysis. In other words, the decomposed tensors can be projected into corresponding 2-dimensional planes.

To demonstrate and simplify the picture of parameter correlations, we apply higher-order SVD (HOSVD) to 2-h-long experimental results^56,57. The truncated core tensor, or decomposed projection, of this analysis is portrayed in Fig. 4g (Supplementary Table S1). The visualization summarizes and highlights the relationships discussed earlier. The prominent diagonal values along the P₂–P₃ basis correspond to the progression of different wavelength modes along the t_react axis. In other words, it shows the shift in bandgap energy as the chloride composition increases and the bromide composition decreases. Conversely, the distinctly lower values, apart from the first mode of the P₁ basis, indicate that a single decay profile dominates, regardless of λ_probe or t_react.

Through an iterative fabrication approach, often referred to as compositional engineering, remarkable strides have been achieved in developing efficient perovskite solar cells^58,59,60,61 and light-emitting devices⁶². This concept also extends to the quantum yield efficiency of PeNCs⁵⁵. The importance of halide composition has been emphasized in studies highlighting the close link between the purity of a single-photon source and its halide composition⁶³. Halide exchange, usually induced by salt-based mechanisms, has been a topic of great interest for those who study the effects of compositional modifications while preserving morphology and minimizing other structural influences^43,47,54,64. In particular, chloride substitution, despite its favorable bandgap for blue or ultraviolet light sources, has faced challenges in practical applications due to low quantum yield and relatively poor power conversion efficiency^51,65. Although compositional engineering has proven to be effective, it remains a time-consuming and complex process. As a result, considerable efforts have been directed toward evaluating photoluminescence lifetimes through in-line synthesis⁶⁶. Our strategy, utilizing AI-TA, offers a faster and more precise method for halide composition engineering.

AI-TA measurement of photon-driven transformation process

In the next phase of our AI-TA investigation, we observed changes in the transient features of PeNPLs dispersed in hexane as the measurement progressed. PeNPLs are characterized by stronger exciton binding energy due to enhanced quantum confinement from thickness control. The thickness is quantified by the number of lead halide octahedra layers, referred to as monolayers (MLs), in the quantum-confined dimension. In addition to varying the halide composition mentioned earlier, the bandgap energy of PeNPLs can also be tuned by controlling the confinement thickness. This approach is effective for only a few MLs, as the effective Bohr radius for exciton in CsPbBr₃ is ~7 nm, and the size of a cubic unit cell is 0.58 nm^13,31,67.

The composition investigated here consists of cesium lead bromide with a 2 ML confinement layer. This is the thinnest variant, other than 1 ML, which consists of a single layer of lead halide octahedra without cesium cations and is stabilized with ligands. A TEM image confirms the formation of PeNPLs with an approximate size of 20 nm in the non-confined dimension (Fig. 5a), which is known not to affect the peak energy significantly⁶⁸. The absorption and emission spectra of the synthesized CsPbBr₃ PeNPLs show peaks near 430 nm, consistent with reported values, whereas bulk materials with the same composition show peaks near 530 nm^68,69. The sharp excitonic peak confirms the large binding energy that is characteristic of PeNPLs. After the AI-TA measurement, the steady-state spectra reveal the appearance of additional peaks in the longer wavelength region (Fig. 5b). The redshifted spectrum matches the characteristic multiple peaks of higher-order PeNPLs with more MLs, indicating the occurrence of photon-driven transformation during the AI-TA measurement. Compared to photoluminescence, the higher-order peaks are less prominent in the absorption spectrum because they have higher quantum yields and therefore show stronger photoluminescence^25,68. The transformation happening over the course of the experiment, visualized earlier in Fig. 1e, can be summarized with the following Eq. (3).

Fig. 5: In situ and real-time analysis of 2ML PeNPLs kinetics using AI-TA.

a TEM image of PeNPLs standing face down and side up. The average size in unconfined dimensions is ~20 nm. b Normalized absorption and emission spectra of PeNPLs before and after the AI-TA experiment. The probe range is marked with a dotted red square. c A transient absorption profile of 2ML PeNPLs acquired by averaging 400 scans of AI-TA. d Projection by t_react of individual scans participating in the averaged spectrum. e A collection of transient spectra at the t_pump-probe of 1 ps for different reaction times. f A collection of time profiles at the probe wavelength of 431.6 nm for different reaction times.

Another important consideration in this experimental scheme is that the bandgap energies of 3ML and higher-order structures, including bulk material, are much lower and fall outside the spectral range of our near-ultraviolet probe⁶⁸. Therefore, only the first half of the reaction involving only the 2ML PeNPLs and their agglomerate structures is selectively detected. Also note that other perovskite phases, which can form in cesium-poor environments similar to PeNPLs, have higher band gaps and are not considered in this investigation.

Typically, nonlinear responses are small, so transient absorption signals are usually averaged over extended DAQ periods. Here, it should be emphasized that the effect of sample coalescence is not very evident in the averaged TA measurement of PeNPLs shown in Fig. 5c, which represents the collective response of 400 interferogram scans acquired over ~40 min of t_react. However, AI-TA enables us to analyze the data with fewer averages and varying t_react durations (Fig. 5d). Unlike traditional TA spectroscopy, we can analyze the spectrum while adjusting the desired measurement time.

By separately analyzing the data in 1-min segments of t_react, we observe a blueshift in the transient signals. Additionally, the peak intensity decreases, likely due to sample degradation, as similar trends are observed in the AI-TA scan of PeNCs in toluene (Supplementary Fig. 11), where the intensity diminishes over the course of the scans. Such sample changes would likely interfere with time-profile measurements in conventional systems, where TA spectra are averaged over fixed time intervals.

The complete disappearance of AI-TA signal, marking the complete consumption of 2ML variants, takes longer than 2 h. However, in this study, extended measurements after 400 scans are affected by aggregation at the focus and the formation of higher-order structures, and thus these data are discarded from the analysis (Supplementary Fig. 15). Consequently, this result cannot be directly compared with the post-experiment steady-state spectra. TEM imaging of the measurement sample after collection reveals some mosaic-like structures of PeNPLs (Supplementary Fig. 16), consistent with previously reported observations.

Spectra of transient signals at a fixed t_pump-probe of 1 ps (Fig. 5e) and a fixed probe wavelength of 431.6 nm (Fig. 5f) were collected to compare the evolution of the signals over reaction time, revealing clear and definite changes. The shifts in the PB signals to lower wavelengths are subtle but noticeable in the AI-TA spectra. The bandgap bleach peak shifts by ~1 nm, accompanied by changes in the distribution of excited carrier populations, as indicated by shifts in both sidebands (Supplementary Fig. 17). This effect becomes more pronounced when comparing normalized spectra acquired at different reaction times. This observation can be attributed to the preferential degradation of PeNPLs with larger non-confined dimensions. However, the impact of this parameter on the shift of the excitonic absorption peak is less significant compared to the number of layers, which serves as a more discrete parameter⁶⁸ (Supplementary Fig. 18).

The collection of time profiles at a fixed wavelength (Fig. 5f) shows that the decay trend also changes. The triexponential fit of the time profiles yields three different time constants (Supplementary Fig. 19), which can be interpreted in terms of kinetics similar to those identified earlier in Fig. 2, representing hot carrier thermalization, charge trapping, and carrier recombination⁷⁰.

The shortest component, corresponding to the hot carrier thermalization process, provides a unique insight into the dynamics of PeNPLs. The spectral and temporal signature comprises a sub-picosecond peak redshift, appearing as a decay at shorter wavelengths and as a rise at longer wavelengths. The excess energy of the hot state is lost via electron-phonon coupling. As the measurement progresses, the hot carrier relaxation lifetime of PeNPLs gradually increases from ~0.08 to 0.12 ps, indicating a bottleneck in the relaxation dynamics.

Although the hot phonon bottleneck is reported to have a negligible impact on highly confined CsPbBr₃ PeNCs⁷¹, such a slowdown in hot carrier cooling has been observed in superlattice PeNPLs⁷². Similarly, the efficiency of hot phonon cooling is much better in colloidal samples compared to film counterparts with the same number of MLs⁷³. In both cases, carrier cooling has been compared between colloidally dispersed and dried samples and confirmed through power dependence measurements. Another study on a sample aged for 80 days and measured with time-resolved upconversion spectroscopy also reports an increased lifetime³³.

Likewise, our results can be interpreted as an agglomeration-induced suppression of hot carrier relaxation. The excess energy of the hot state is dissipated via electron-phonon coupling to form longitudinal optical phonons^72,74. The thermal energy is subsequently transformed into out-of-plane modes and dissipated into the environment, particularly efficiently in highly confined small PeNCs and PeNPLs. However, the formation of stacked structures, known as superlattices, suppresses this out-of-plane coupling and causes a bottleneck in thermal relaxation. In our case, the observation of the hot phonon bottleneck was achieved not by measuring power dependence, but through fast DAQ of the progressing photoinduced agglomeration of PeNPLs.

The other lifetime, associated with the charge-trapping process, remains almost consistent. This lifetime is assumed to be intrinsic to the state of PeNPLs undergoing AI-TA measurement, where femtosecond laser pulses create a photochemically dynamic and highly stressed environment. Another interesting observation is that the lifetimes of the longest component are also increasing. This component reflects exciton recombination, which may involve trapped charges, as observed in the case of PeNCs. This could be related to the size reduction observed alongside the PB peak shift, which involves the removal of incomplete unit cells with vacancies at the edges of the nanoplatelets.

This trend can also be visualized in collections of highly averaged and normalized time profiles at longer wavelengths. One advantage of AI-TA is that it allows control of the length of DAQ for averaging, enabling us to validate the minute changes observed here. Shorter acquisition times result in a lower signal-to-noise ratio, so integrating longer reaction times and analyzing the data over these periods helps clarify the trend observed in the experiment. This is demonstrated by averaging over different reaction times. Comparing the averaged results for 1, 5, and 10 min of t_react (Supplementary Fig. 20) confirms the observations made with 1-min averages. Again, we emphasize the strength of the AI-TA technique, particularly its ability to account for changes occurring throughout the measurement process.