WorksPhysics and Astronomy
Predictive sufficiency as an information microscope for time-resolved coherent scattering
Abstract
X-ray photon correlation spectroscopy (XPCS) measures dynamics through the evolution of coherent speckle patterns, most commonly summarized by intensity correlation functions. These averages describe relaxation but can conceal how speckle fluctuations evolve in time and across reciprocal space. Here we introduce an information microscope that tests whether a chosen set of XPCS observables retains the information needed to describe subsequent measured scattering dynamics. A learned representation of the present speckle sequence indicates when information remains outside the chosen description; an XPCS Alphabet then searches for experimentally accessible observables that recover it. In simulated Bragg scattering, speckle activity and anisotropy reveal signatures of correlated deformation and intermittent slip. In a controlled SAXS model, fluctuations of speckle activity provide sensitivity to the spatial extent of cooperative particle motion. In a two-dimensional glass former, retaining time-resolved speckle correlations and radial intensity recovers information lost in the average relaxation function. The fluctuations of these correlations become stronger during caging, without selecting a single dominant correlated mode. This approach uses learning to guide the physical analysis of coherent scattering, while keeping the resulting XPCS observables measurable and interpretable.
Introduction
A central problem in the analysis of time-resolved measurements is how to represent a complex, evolving signal by a manageable set of observables. The full measurement record is commonly reduced to correlation functions, relaxation times, fluctuation amplitudes, order parameters, or related descriptors. These quantities provide physical insight, but the reduction can discard information relevant to subsequent measured dynamics. The practical question is therefore whether the chosen observables retain the information required to describe those dynamics.
This issue becomes particularly important for heterogeneous, intermittent, aging, or otherwise complex dynamics. Two measurements can exhibit similar mean relaxation while differing substantially in the temporal organization, reciprocal-space distribution, or intermittency of their fluctuations. A high-dimensional measurement may therefore contain dynamically relevant information that is absent from a compact correlation function. Conversely, a suitably chosen set of experimentally accessible observables may already capture essentially all information relevant to a specified future measurement.
X-ray photon correlation spectroscopy (XPCS) provides a particularly direct setting in which to address this problem. Coherent X-ray scattering produces speckle patterns whose temporal evolution encodes the dynamics of the scattering structure [1, 2, 3, 4]. The recorded speckle movie contains the time-dependent scattering intensity, whereas conventional XPCS extracts reduced observables such as the intensity autocorrelation function , with scattering-vector magnitude and delay time , two-time correlations, and measures of temporal fluctuations and intermittent dynamics [5, 6, 7, 8, 9, 10]. These observables have provided detailed insight into relaxation and collective motion across a broad range of materials and time scales. Yet it is not generally tested whether a given reduction retains the information relevant to subsequent measured dynamics.
This question is relevant across the range of current XPCS experiments. Measurements of non-equilibrium and aging systems have revealed relaxation beyond simple exponential behavior [11]. Atomic-scale XPCS probes slow motion in amorphous materials [12, 13, 14, 15, 16], while resonant XPCS accesses fluctuations of ordered states [17, 18]. Split-pulse, MHz, and pump–probe methods extend coherent-scattering measurements towards faster and driven processes [19, 20, 21, 22]. Machine-learning methods have also been used to denoise scattering data, classify dynamical patterns, compare measurements with simulations, and extract dynamical information [23, 24, 25, 26, 28] [27]. These developments make it increasingly valuable to establish which information a conventional XPCS reduction preserves and which it leaves behind.
Here we address this question through predictive sufficiency. We use experimentally accessible observables , evaluated from a present measurement window, to describe a later scattering quantity . The vector may contain measurements at several scattering vectors and correlation delays; it does not imply averaging over . We then compare this description with one that also includes a learned representation of the present speckle sequence. If improves the description of the same future measurement, the chosen XPCS observables have left accessible information unused. This practical test is related to predictive-state and information-based approaches [29, 30, 31, 32, 33, 34, 35]. It does not require the speckle movie to determine a unique microscopic state or predict individual stochastic events.
We call this approach an information microscope because the learned representation is a diagnostic, not the intended final physical observable. Once a missing contribution is detected, we search an XPCS Alphabet of quantities constructed from scattering intensities: correlations resolved in wavevector and delay, short-time intensity changes, spatial variations of speckle activity, and measures of diffraction-pattern shape. Candidates are chosen for their ability to improve the description of subsequent XPCS measurements. The observable search uses scattering intensities; controlled simulations then allow us to examine the physical sensitivity of the selected quantities. The benchmark-specific encoder training and model selection are described separately in Methods and Supplementary Information.
We illustrate the approach in three scattering settings. Around a Bragg reflection, we examine how changes in speckle activity and anisotropy encode correlated lattice deformation and intermittent slip [36, 37, 38, 39, 40, 41, 42]. In a controlled SAXS model, we distinguish the amplitude of collective displacements from the spatial extent of cooperative motion. Finally, in a dense two-dimensional glass, we identify which time-resolved speckle correlations and intensity measures improve the description beyond the mean relaxation, and examine their fluctuations as caging develops. The three examples show how information retained in a coherent-scattering movie can guide the choice of physically interpretable XPCS observables.
Results
A predictive description of XPCS dynamics
We describe the present scattering measurement by a set of experimentally accessible XPCS observables and define a future target from scattering data acquired at later times. Here marks the end of the present observation window and specifies the separation to the later measured target, according to each benchmark. Depending on the system, the target may describe a later intensity correlation or another measured scattering quantity; its definition and time window are specified for each benchmark. A regression model is trained on one set of speckle sequences and tested on independent data. We quantify prediction quality by the coefficient of determination .
To assess whether the chosen observables capture the information relevant to the future measurement, we compare predictions from alone with those obtained after adding a learned representation of the same present speckle movie. The gain measures the additional predictive contribution of the movie representation under the tested models. A statistically resolved positive gain motivates a search for further measurable XPCS observables that can account for it.
Stationary and smoothly aging Brownian dynamics provide simple controls (Figure 1). Panel a compares predictions from the measured observables alone with those including the learned speckle representation , using the same later measured target. The two-time correlations in panels b and c illustrate stationary relaxation and gradual aging, respectively. In the stationary system, the ensemble-mean relaxation is reproducible, but fluctuations in a later finite measurement window are only weakly predictable from the preceding window at the tested horizon. In the aging system, the gradual slowing is already reflected in the measured intensity correlations and provides predictability. When the explicit description includes correlations at several wave vectors and delay times, their temporal fluctuations, and correlations evaluated over successive intervals of the present measurement, adding the learned speckle representation yields no statistically resolved improvement in either case (Figure 1d). These controls illustrate two distinct outcomes of the same test: little predictable information in the measured future, or predictable dynamics already captured by conventional XPCS quantities. They do not establish information-theoretic sufficiency beyond the tested models and data.
Figure 1. Predictive description of XPCS dynamics and Brownian controls. a, Present speckle patterns are characterized by XPCS observables and compared with subsequent scattering dynamics ; represents additional information from the same speckle sequence. b,c, Independently generated, illustrative two-time correlations for stationary and aging Brownian motion; these panels do not supply the quantitative prediction results in panel d. d, Ridge prediction of subsequent correlations using , comprising mean intensity correlations, fluctuations of speckle decorrelation, and their time-resolved averages. The additional speckle information gives no resolved improvement in either case. Error bars show archived 95% paired whole-movie confidence intervals.
Bragg speckles reveal information beyond a compact XPCS description
We first consider coherent scattering around a Bragg reflection, where even small lattice displacements modify the phase of the scattered wave and redistribute intensity within the speckle pattern (Figure 2a) [36, 37, 38, 39, 40, 41, 42]. For a reciprocal-lattice vector and a small deviation from the Bragg position, the simulated coherent amplitude is

Figure 2. Bragg speckles reveal signatures of correlated deformation and intermittent slip. a, Schematic correlated lattice displacements and localized slip, with the resulting coherent Bragg intensity and radial intensity profile before and after a slip. b, Prediction of later Bragg-XPCS correlations using the compact description and , the additional learned speckle representation, and two or five observables identified by the XPCS Alphabet. The five-addition state is selected separately from the two-addition state. Error bars indicate 95% whole-movie bootstrap intervals. c, Short-time intensity changes in the outer Bragg halo and anisotropy of the intensity distribution, illustrating the first two Alphabet observables. d, Representative evolution of simulated deformation amplitude , spatial correlation length , and short-delay speckle correlation around a slip event.
Here denotes the spatial illumination envelope. The displacement field consists of reversible, spatially correlated distortions and permanent changes caused by localized slip events:
The elastic displacement field evolves according to
where is a normalized, spatially correlated Gaussian random vector field with correlation length . The amplitude sets the strength of the reversible displacements, while provides a transient enhancement following a slip event. The factor 0.82 describes the persistence of earlier deformations. Localized slip events additionally produce permanent increments in the plastic displacement field. The model thus generates continuous speckle fluctuations from correlated lattice motion and abrupt changes from intermittent rearrangements, as encountered in Bragg-XPCS. Predictions use only the simulated detector intensities; the displacement and event variables are used subsequently for physical interpretation.
We initially characterize the measured Bragg dynamics by two correlations at a delay of four frames: the usual intensity autocorrelation and a spatial Pearson correlation between speckle patterns. While measures intensity decorrelation, emphasizes changes in the spatial speckle configuration after removing the mean intensity and contrast of each frame. We ask how well these two quantities, measured over a short present window, predict their counterparts in a later, non-overlapping window separated by 16 frames.
The two correlations predict the future Bragg-XPCS observables with . Including information learned directly from the present speckle sequence increases the prediction score to (Figure 2b). Thus, the evolving speckle pattern contains information about subsequent dynamics that is not captured by conventional intensity decorrelation and spatial speckle similarity alone. This gain reflects information missing from the chosen experimental description, not an ability to predict every future rearrangement: stochastic deformation and slip events also limit the reproducibility of future speckle fluctuations (see Supplementary Information).
To recover this additional information in experimentally accessible terms, we use the XPCS Alphabet: a library of candidate observables built from speckle-intensity changes, temporal fluctuations, reciprocal-space variations, and pattern-shape measures. Starting from the two conventional correlations, we select additions according to how much they improve prediction of the same future XPCS observables. Importantly, this search uses only measured scattering intensities, not the known microscopic deformation or slip variables. It therefore asks which physical features of the speckle movie are missing from the compact correlation description.
The search first identifies two complementary quantities (Figure 2c). The first, , characterizes strong, short-time intensity changes in the outer Bragg halo relative to activity in other reciprocal-space regions. The second, , captures anisotropy of the intensity distribution during the observation window. Together, they retain information about where intensity changes occur and how the overall Bragg pattern changes shape—features not distinguished by the two initial correlation measures.
Adding these observables increases the prediction score from to 0.412, recovering much of the improvement obtained from the learned movie representation (). Three further short-delay and -resolved observables bring only a small additional improvement, to . The learned representation still improves prediction beyond this enriched XPCS description, indicating that some information in the speckle sequence remains unaccounted for.
Having identified scattering observables that improve the prediction of future Bragg-XPCS correlations, we next ask what these observables reveal about the underlying motion. The simulated deformation field and slip history are known, allowing a separate test of how well microscopic properties can be inferred from the measured XPCS description. These microscopic variables were not used in selecting the Alphabet observables. Figure 2d illustrates a representative slip event: the short-delay speckle correlation changes sharply near the event, while the deformation amplitude and correlation length evolve over the same interval. The figure illustrates the dynamics to be inferred, not the accuracy of the inference.
The enriched XPCS description is particularly sensitive to intermittent motion: the reconstruction score for the number of recent slip events increases from to 0.517, and that for the time since the last event from 0.229 to 0.317. For the separately selected five-observable description, the score for the reversible-deformation amplitude increases from 0.096 to 0.274, while the spatial correlation length , weakly represented by the initial correlations, reaches . These results do not establish a unique reconstruction of individual slip events or displacement fields. Rather, the spatial distribution of speckle activity and changes in the Bragg-pattern shape carry complementary signatures of correlated deformation and recent slip history.
This conclusion is not specific to a linear prediction model. Nonlinear prediction and conventional covariance-based descriptions also recover part of the missing information, but neither removes the additional contribution of the learned speckle representation. Details of these comparisons and their statistical uncertainties are given in the Supplementary Information.
A controlled SAXS model separates dynamical activity from cooperativity
We next consider a model of coherent SAXS from particles that undergo independent Brownian motion together with intermittent collective displacements (Figure 3a). The collective motion acts within square spatial domains of linear size : particles in the same active domain receive the same displacement, whereas particles in different domains move independently. Its amplitude is the standard deviation of each Cartesian component of this common displacement. Varying and independently therefore allows us to distinguish how far particles move from how many particles move together. In an XPCS experiment these effects can both alter speckle decorrelation, making their separate identification from conventional correlations nontrivial.
Figure 3. XPCS fluctuations reveal the amplitude and extent of collective motion. a, Model of intermittent collective displacements within domains of size and amplitude , together with a representative coherent SAXS pattern and intensity correlations . b, Prediction of subsequent XPCS correlations and fluctuation amplitudes from the initial and -like observables, an additional manually defined speckle-activity measure, and two or five observables selected by the XPCS Alphabet. The full learned speckle representation provides the comparison. c, Time-dependent speckle activity and its coefficient of variation (top), and the additional predictive contribution for future mean relaxation and -like fluctuations (bottom). d, Reconstruction of the collective displacement amplitude and cooperative domain size from the compact and enriched XPCS descriptions. Dashed lines indicate exact agreement. Error bars in the prediction and reconstruction panels represent the reported uncertainty estimates (see Methods).
As a compact description of the measured dynamics, we use the intensity autocorrelation and a -like measure of the fluctuations of the time-resolved correlation, both averaged over the measured range. We ask how well their values in a 72-frame observation window predict the same quantities in the following non-overlapping window. The mean correlation characterizes the overall relaxation, while its fluctuations reflect how strongly the dynamics vary during the measurement.
These conventional XPCS quantities predict the future correlations and fluctuation amplitudes with . Adding a representation learned from the full present speckle sequence improves the prediction to (Figure 3b). The improvement is markedly larger for the future -like quantity () than for the mean correlation (). Thus, much of the average relaxation is already captured by the compact XPCS description, whereas the speckle movie contains additional information about the future strength of dynamical fluctuations.
This points to the time dependence of speckle activity as a useful place to look for the missing information. We first make a simple, physically motivated test: from the changes in detector intensity between successive frames, we calculate an activity signal and measure how strongly it varies over the present observation window. Adding its coefficient of variation (defined in Methods) raises the prediction score to (Figure 3b); the activity trace is shown in panel c. The remaining predictive gain from adding to this enriched measured state decreases from 0.093 to 0.058. This shows that the intermittency of measured intensity changes already accounts for an appreciable part of the information missed by the mean correlation and its fluctuation amplitude.
We then broaden this targeted test into a systematic search using the XPCS Alphabet. Rather than prescribing one activity measure, the Alphabet generates candidate observables from experimentally available speckle intensities, their short-time changes, and their dependence on scattering vector and time. Each candidate is tested for whether it improves prediction of the same future XPCS quantities beyond the two initial correlations. The observable search uses the measured scattering intensities, not , , or particle trajectories; the separate historical choice of encoder architecture did use these simulated parameters (see SI). In this way, the learned movie representation first identifies an information deficit, and the Alphabet searches for measurable scattering quantities that can account for it.
The first selected observable combines temporal variations in the overall strength of frame-to-frame intensity changes with a fluctuation measure from a fixed region. It therefore retains both when strong speckle activity occurs and which part of reciprocal space contributes. The second characterizes how the spread of intensity changes across detector pixels varies over time. In contrast to a single averaged correlation, these quantities distinguish speckle sequences with similar mean decorrelation but different histories of fluctuating activity. They do not, by themselves, measure the real-space size of a cooperative domain.
Adding these two automatically selected observables increases the prediction to , leaving an additional movie contribution of only 0.014. An independently selected, larger set of five additions, which also retains spatial dispersion of intensity changes between detector patches and additional -resolved intensity-change summaries, gives and a remaining contribution of 0.008. Thus, most of the predictive improvement initially found in the full speckle movie can be recovered through a small set of explicit quantities calculated from the measured scattering intensities. The detailed definitions and selection tests are provided in the Supplementary Information.
The model now allows us to examine what physical information these observables have recovered. Only after the scattering-based selection is complete do we use the known values of and to test how well they can be inferred from the measured XPCS descriptions (Figure 3d). The collective displacement amplitude is already well constrained by the compact correlations (); the enriched descriptions raise this to 0.971 and 0.983. The contrast is much stronger for the cooperative domain size: the compact description gives only , whereas the two- and five-observable enrichments reach 0.800 and 0.896, respectively.
The central result is therefore not simply a better prediction of future correlations. Although the original observables already capture much of the displacement amplitude, the additional time- and -resolved speckle fluctuations provide sensitivity to the spatial extent of collective motion. This sensitivity is established through the controlled model comparison; it should not be confused with a direct measurement of from an individual scattering observable. The XPCS Alphabet thus turns information detected in the full speckle movie into interpretable experimental signatures of cooperative dynamics.
A dense two-dimensional glass exposes information beyond the mean XPCS relaxation
Our final example is a two-dimensional 50:50 mixture of soft disks with a diameter ratio of 1.4. Upon increasing the packing fraction from to 0.78, the static scattering remains liquid-like, while the intensity correlations show the characteristic slowing associated with caging. The relaxation time increases and the KWW exponent decreases, indicating a broader relaxation. Particle mean-squared displacements independently exhibit an intermediate-time cage regime (see Supplementary Information). The corresponding structure factors and XPCS relaxation curves are shown in Figure 4a,Figure 4b; the predictive-state comparison appears separately in panels c,d.
Figure 4. From conventional glass dynamics to an explicit XPCS-state description. a, Static structure factor across packing fractions. The dotted lines indicate the representative wavevectors used in the predictive XPCS analysis; is the fixed reciprocal-space reference of the simulation. b, Representative two-time Pearson speckle-pattern correlation at for the densest state (top); independently averaged Pearson relaxation curves across packing fractions (bottom). The inset shows the increasing relaxation time and decreasing KWW exponent . c, Cumulative measured XPCS-state ladder, starting from the mean (), followed by time-resolved speckle-pattern correlations () and reference-plus-current radial-intensity information (). d, Predictive information along the same ladder. Top: prediction score of the explicit XPCS state; adding the reference-plus-current radial-intensity block further improves prediction from to . Bottom: additional predictive contribution of the full present speckle history beyond each explicit state. The large additional contribution beyond disappears once the time-resolved speckle-pattern correlations are retained.
We first characterize the dynamics by the time-averaged intensity correlation at , where the fixed reference sets the reciprocal-space scale. This compact description captures the average relaxation but removes variations between individual time origins. In the prediction test of Figure 4, is a fixed-prefix mean- calibration vector, repeated within each movie, rather than a moving-window correlation estimate. Adding the learned representation of the present speckle history increases the prediction score by . This demonstrates that the fixed mean-correlation reference omits information relevant to subsequent scattering dynamics; by itself, it does not establish that every conventional time-resolved description is insufficient.
The XPCS Alphabet helps identify which additional parts of the measured speckle signal are useful. In a separate analysis using only earlier scattering frames to establish the mean- reference, it first selects the angular spread of the instantaneous speckle intensities at three scattering vectors. It subsequently selects combinations of speckle-pattern correlations at short and longer delays. These quantities describe two familiar properties of coherent scattering: intensity variations around a ring and the loss of speckle similarity with time. Their calculation requires neither particle coordinates nor a supplied packing-fraction label.
These five scattering quantities improve prediction of the later XPCS correlations in all three independent test cohorts. For the largest cohort, increases from for the mean-correlation description to 0.041 with the selected observables. The improvement persists when the future correlation is formed entirely from frames recorded after the present observation window. The full movie still provides a small additional improvement for the primary target. Across repeated searches, the precise selected quantities vary, but intensity-amplitude and time-resolved correlation measures recur (details in the Supplementary Information).
The selected correlations motivate retaining the time dependence of the speckle pattern rather than reducing it to a mean relaxation curve. We therefore examine the ring-wise Pearson correlation
and its corresponding decorrelation . The following comparison is the separate, physically defined XPCS-state analysis shown in Figure 4; its prediction scores are not combined with those of the Alphabet search.
Pearson normalization removes the instantaneous mean intensity and overall intensity scale of each ring, emphasizing changes in the registered speckle configuration. The angular coordinate labels detector positions along a constant- ring; it is not used here to determine orientational order. We calculate pairwise Pearson correlations among 15 speckle patterns sampled from the preceding 128 frames. Together with the fixed mean- reference, these values form , which retains the temporal organization of the recorded speckle correlations.
In the original Figure 4 test, gives , while adding the learned speckle representation gives , with no positive incremental gain. This indicates that the explicit time-resolved Pearson description recovers much of the predictive contribution missing from the fixed mean-correlation reference, although the absolute predictability of later correlations remains low.
To distinguish temporal resolution from Pearson normalization, we also examined moving-window conventional intensity correlations and matched pairwise conventional two-time correlations (Supplementary Information). Moving-window conventional correlations improve on the fixed-prefix reference, but the learned representation retains a positive incremental contribution in the tested summaries. At matched correlation endpoints and state dimension, Pearson correlations provide a resolved predictive advantage over conventional pairwise correlations when the instantaneous radial scattering intensity is excluded.
We next retain the instantaneous radial scattering intensity , whose fluctuating amplitude is removed by Pearson normalization. In the original Figure 4 analysis, adds both a density-dependent reference scattering profile and the instantaneous radial intensity relative to that reference, improving prediction from to . This comparison alone cannot separate the two contributions. Matched controls without a supplied density-dependent reference show that the instantaneous radial intensity itself improves prediction, without explicitly supplying a density-dependent mean profile. Once the same instantaneous radial-intensity information is supplied to both pairwise correlation descriptions, the Pearson advantage is no longer resolved. The remaining contribution of the learned representation to conventional pairwise correlations plus instantaneous radial intensity is also unresolved, although predictive equivalence is not established (Supplementary Information).
These comparisons show that predictive information absent from a fixed mean-relaxation reference can be recovered by explicit time-resolved correlations together with instantaneous radial scattering intensity. Pearson correlations are advantageous in a matched comparison without this intensity information, but their advantage is not resolved once the same instantaneous radial-intensity measurements are available to both representations. The results therefore do not identify a uniquely predictive Pearson-specific or hidden glass coordinate. Instead, they motivate examining how the fluctuations of the measured time-resolved scattering observables are organized across wavevector and internal delay as caging develops.
Fluctuations and correlations within the predictive XPCS state
We extend the Pearson-correlation analysis to a denser grid of scattering vectors and internal delays, resolving the fluctuations within the experimentally accessible XPCS state identified above.
For each time origin, we compare two nearby cuts through the two-time correlation function (Figure 5A). Using the Pearson speckle-pattern decorrelation
Figure 5. Fluctuations and correlations within the time-resolved XPCS state. a, Short-time changes in the Pearson speckle-pattern decorrelation , obtained by comparing two cuts through the TTC at time origins separated by , while keeping the internal delay fixed. The illustration uses frames; the analysis at uses frames. b, Fluctuation amplitude of these short-time decorrelation changes across scattering vector and internal delay, shown on a common color scale. The -dependence resembles the relaxation-rate modulation . c, Distribution of the leading correlated fluctuations across and short delays, after removing specified mean-relaxation contributions and normalizing channel variances. Colors indicate participation in the leading two correlation directions, not fluctuation amplitude. d, Evolution with packing fraction of the mean raw fluctuation variance (top) and the null-corrected squared residual cross-channel correlation strength (bottom). e, Reciprocal-space profile of the leading correlated sector, summed over six short internal delays. Its weight remains broadly distributed in , without a sharply selected wavevector.
we define its short-time change as
Here, is the internal XPCS delay, whereas denotes a short shift of the time origin. The latter follows the density-dependent relaxation time and ranges from one to seven frames over the investigated packing fractions. Thus, measures how a fixed-time cut through the two-time correlation pattern changes over a short interval.
The analysis contains 31 scattering-vector values and 14 internal delays, giving 434 - and delay-resolved XPCS channels. Repeating the measurement of over many time origins allows us to construct the covariance matrix
where and denote individual measured XPCS channels. The diagonal elements of give the variance of the short-time changes in each channel. Figure 5B shows this local increment fluctuation amplitude,
and therefore identifies where in reciprocal space and internal delay the time-resolved XPCS signal changes most strongly. The fluctuation landscape remains broad in both and , with pronounced structure below in the dense states and substantial weight extending over a wide range of internal delays. Its reciprocal-space dependence resembles the de-Gennes-modulated relaxation-rate scale . This correspondence is a comparison of observed -dependent trends, not an identification of fluctuation variance with the relaxation rate.
The off-diagonal elements of provide complementary information: they quantify whether short-time changes in different - and delay-resolved channels fluctuate together. The strongest raw covariance directions, however, are largely aligned with variations of the average relaxation pattern itself. To isolate correlations beyond these mean-relaxation contributions, we remove the corresponding frozen mean-relaxation directions and normalize the remaining channels by their own fluctuation amplitudes (see Methods). This yields a residual correlation matrix that measures how strongly the remaining short-time changes at different values and internal delays fluctuate together relative to their local variance.
Figure 5C shows where the leading correlated structure of is distributed in - and delay-space. The displayed quantity,
combines the two leading eigenvectors of the normalized residual correlation matrix. It is not a fluctuation amplitude, but measures how strongly each - and delay-resolved XPCS channel participates in the leading correlated pattern. This leading two-dimensional sector accounts for only a small fraction of the full residual correlation spectrum. Related short-delay organization persists over a broader set of correlation directions, while its specific -localization becomes progressively less pronounced.
The density evolution is summarized in Figure 5D. The mean variance of the short-time XPCS changes is larger in the dense regime, indicating stronger local fluctuations in the measured Pearson-decorrelation signal. The lower part of Figure 5D shows the equal-20-movie, null-corrected squared residual correlation strength, (see Methods). This matched analysis shows a small, null-corrected difference between the low- and high-density regimes. The data establish a modest difference in residual correlation strength between density regimes, without demonstrating monotonic growth of inter- organization. A separate finite-size analysis confirms that correlations between time-resolved XPCS channels persist for particle numbers up to , although the density dependence and full covariance structure are not demonstrably converged.
Figure 5e shows the reciprocal-space distribution of the leading correlated sector at short internal delays. The weight , summed over frames and normalized over , remains broadly distributed in reciprocal space. Contributions from wavevectors below the principal structure-factor peak coexist with those at intermediate and structural length scales, without the emergence of a sharply selected wavevector.
Taken together, the variance and covariance analyses reveal two complementary aspects of the time-resolved XPCS dynamics. The diagonal variance describes the amplitude of short-time decorrelation changes. Its reciprocal-space dependence resembles the structurally modulated relaxation-rate scale , suggesting a relation to de Gennes narrowing without identifying the variance with a relaxation rate. The off-diagonal covariance, by contrast, identifies correlations between fluctuations at different wavevectors and delay times. These correlations are concentrated predominantly at short delays and remain distributed across a broad residual correlation space, rather than forming a dominant collective mode. Although the dense regime exhibits larger absolute fluctuations and a small, resolved excess of residual cross-channel correlation, neither a monotonic increase in correlation strength nor a distinct onset of cross-scale organization is established.
Discussion
Conventional XPCS describes dynamics primarily through intensity correlation functions and their fluctuations. These quantities provide direct access to relaxation and its evolution, but they do not necessarily retain all the information contained in a time-resolved speckle movie. In particular, measurements with similar average relaxation may differ in the temporal organization and reciprocal-space distribution of their intensity fluctuations. Our information microscope addresses this limitation by testing whether a chosen set of XPCS observables captures the information relevant to subsequent scattering dynamics. When additional information is detected in the speckle sequence, the XPCS Alphabet provides a systematic route to identifying measurable quantities that recover it. The objective is therefore not to replace conventional XPCS observables with a learned representation, but to extend the physical description where conventional data reduction proves insufficient.
| System / benchmark | Measured XPCS description | ||
| Bragg | , | 0.347 | 0.112 |
| Alphabet: two additions | 0.412 | 0.047 | |
| Alphabet: five additions | 0.417 | 0.042 | |
| Cooperative SAXS | , -like | 0.523 | 0.093 |
| Alphabet: two additions | 0.602 | 0.014 | |
| Alphabet: five additions | 0.609 | 0.008 | |
| Glass | Mean | -0.173 | 0.196 |
| 0.041 | -0.005 | ||
| 0.056 | -0.010 | ||
| Glass: causal Alphabet | Mean | -0.182 | 0.201 |
| Five Alphabet additions | 0.041 | 0.014 | |
| 0.037 | 0.005 | ||
| 0.051 | -0.004 |
Table 1. Table 1 Prediction of future XPCS observables. is the prediction from measured XPCS observables and the additional contribution of the learned speckle representation. The two glass analyses use different prediction tasks and are reported separately.
The Bragg and cooperative SAXS examples demonstrate how this approach reveals different aspects of collective motion. In Bragg scattering, intensity changes across the diffraction pattern and changes in speckle anisotropy provide information about correlated lattice deformation and recent slip events that is not retained by average intensity decorrelation alone. In cooperative SAXS, the temporal fluctuations and reciprocal-space distribution of speckle activity provide additional sensitivity to the spatial extent of correlated particle displacements. In both cases, a small number of experimentally accessible observables recovers much of the information initially identified in the full speckle sequence. The microscopic interpretation is supported by the controlled simulations, while the selected scattering quantities themselves require only measured intensities.
The glass example shows that the same approach is useful for increasingly slow and caged dynamics. Much of the information missing from the time-averaged intensity correlation can be recovered by retaining the temporal evolution of the speckle correlations, with a further contribution from radial scattering intensity. This identifies a time-resolved XPCS description that goes beyond the average relaxation function without requiring a uniquely defined microscopic glass coordinate. Examining its fluctuations as packing fraction increases reveals a pronounced growth in the amplitude of short-time decorrelation changes. Their reciprocal-space dependence broadly follows the structurally modulated relaxation-rate spectrum, while the normalized correlations between different wavevector and delay channels change much less strongly. Thus, increasingly pronounced fluctuations during caging do not necessarily correspond to the emergence of a dominant correlated mode in the measured XPCS dynamics.
Predictive sufficiency is a practical criterion for evaluating an XPCS description, rather than a claim that the microscopic state of the sample can be reconstructed uniquely from coherent scattering. Its outcome depends on which future scattering quantity is considered and on the time scales accessible to the measurement. Similarly, improved prediction does not by itself identify a microscopic mechanism. The physical interpretation emerges from the additional scattering observables and their comparison with established XPCS quantities, structural information, and controlled models. The learned representation serves primarily to indicate where the measured description can be improved.
This perspective also changes the role of the recorded speckle movie. Rather than treating it only as the starting point for calculating a prescribed set of correlation functions, the full time-resolved intensity record becomes a resource for identifying new scattering observables as physical questions evolve. This is particularly relevant for heterogeneous, intermittent, aging, and driven systems, where temporal averaging may remove important aspects of the dynamics. Retaining such data in reusable form, consistent with FAIR practices and initiatives such as DAPHNE4NFDI [43], allows the physical information accessible through XPCS to be revisited and extended. The information microscope thus complements established correlation analysis by providing a route from the measured speckle evolution to a more informative, experimentally accessible description of complex dynamics.
Methods
Predictive-sufficiency analysis
For all systems, the analysis is formulated as a prediction problem between measured scattering observables. A present observation window ending at time is represented by a vector of experimentally accessible XPCS quantities,
Depending on the system, contains intensity correlations, two-time-correlation coordinates, fluctuation measures, or other quantities calculated directly from the measured speckle intensity.
The prediction target is calculated from later scattering measurements according to a benchmark-specific temporal support. The target always consists of measured scattering quantities and never of microscopic simulator variables. A regression model is fitted to predict
Prediction quality is quantified by the coefficient of determination on independent XPCS realizations.
To determine whether the explicit XPCS description leaves future-relevant information unused, we compare this baseline with a second model in which the same present speckle window is additionally represented by a machine-learning encoder,
where is the causal present support fixed for each benchmark. The encoder is trained to extract information predictive of future measured scattering quantities. Its training targets are benchmark-specific and may include related future observables or several prediction horizons rather than exactly the target used for the final readout. In every comparison, however, the predictors using and are evaluated against the same on identical test data. The augmented prediction is
and the additional predictive information accessible from the speckle movie is quantified by
This difference quantifies improvement in held-out prediction for the specified measured target and tested models. It is not a direct measurement of mutual information or proof of information-theoretic sufficiency.
A positive and statistically resolved indicates that the chosen XPCS reduction has omitted information that remains accessible in the present speckle sequence. Conversely, the absence of a resolved gain is interpreted only within the tested observation window, target, prediction model, and available data; it does not establish microscopic completeness. We then enrich the measured state with explicit scattering observables and repeat the same comparison, so that the recovered information can be analyzed without the learned representation.
The XPCS Alphabet searches themselves use measured scattering quantities and do not select observables by microscopic simulator labels. In the cooperative SAXS benchmark, simulated displacement amplitude and domain size were used earlier to choose among encoder architectures; this architecture-choice step is distinct from training the encoder on future scattering signals and from selecting the Alphabet observables. The selection history and a scattering-only architecture check are documented in the Supplementary Information.
Machine-learning representations and XPCS Alphabet
A learned representation is used to test whether the present speckle sequence contains information about later scattering observables that is absent from the explicit XPCS state. The benchmark-specific encoders are: a supervised linear projection for the Brownian controls; a five-member CNN–Transformer ensemble for Bragg scattering; five CNN encoders of frame-to-frame intensity changes for cooperative SAXS; and a 48-dimensional multilayer-perceptron representation for the glass. Encoder input windows contain only present scattering measurements; encoder-training targets are future scattering quantities and may differ from the final Ridge-readout targets. Training and selection are separated from evaluation on held-out movies. The architectures, input transformations, training targets, cohort sizes, and regularization protocols are specified in Supplementary Methods.
The XPCS Alphabet consists of measured intensity correlations, time-resolved intensity changes, their temporal and reciprocal-space fluctuations, and pattern-shape measures. For Bragg and cooperative SAXS, stable candidate expressions are screened on TRAIN, selected on VALIDATION, and fixed before TEST evaluation. The guarded primary selection jointly rewards better prediction from explicit scattering observables and a reduction of the remaining encoder gain, penalizing additional coordinates and expression complexity. Selection stops at at most five additions or when no admissible validation improvement remains. The secondary state is selected separately by a direct-future forward-selection procedure. The complete candidate grammar, score function, thresholds, and selection history are given in Supplementary Methods. This stopping rule is not evidence of universal predictive sufficiency.
The cooperative SAXS encoder architecture was historically selected using simulated displacement amplitude and domain extent on validation data. This choice was distinct from the scattering-only observable search; an additional scattering-selected architecture control is reported in the Supplementary Information. The glass Alphabet is a separate, future-prediction-based search and is not identical to the physically motivated Pearson-state ladder.
Brownian control models
Stationary and smoothly aging Brownian dynamics serve as controls for the predictive-sufficiency analysis. Each simulated speckle movie contains 96 non-interacting particles in a periodic square of side , propagated for 2176 frames with a time step and reference diffusivity . The stationary system has constant diffusivity, whereas the aging system slows according to
with , 50, or 60. Coherent scattering intensities are sampled at 16 registered angular positions on each of three reciprocal-space rings near , 8, and 11. Neither particle positions nor the simulation time or aging parameters are supplied to the prediction models.
Each present measurement comprises 128 frames, followed by a 16-frame gap and a separate 128-frame future measurement. The future XPCS target contains nine time-averaged angular Pearson-decorrelation values, measured at delays , 8, and 24 frames on the three rings. At each time origin, the Pearson decorrelation is
The measured-state hierarchy separates conventional intensity correlations from their temporal fluctuations. contains nine time-averaged -like intensity-product correlations, one for each ring and delay. adds nine fluctuation amplitudes, each defined as 16 times the sample variance of over the valid time origins. additionally retains four successive temporal averages of for every ring and delay, giving 36 further quantities and 54 measured coordinates in total. The four intervals partition the valid correlation-pair sequence at each delay, rather than the full 128-frame movie into identical segments. Exact averaging conventions and feature ordering are given in Supplementary Information.
For comparison, a nine-dimensional representation is obtained from the present speckle sequence using , followed by training-set standardization and supervised cross-covariance singular-value decomposition. Predictions from and use Ridge regression trained on independent movies. The encoder-training, regression-training, validation, and test sets contain 20, 40, 15, and 20 independent movies, respectively, for each system. Each test movie provides eight non-overlapping present/future pairs. Prediction scores are calculated over the nine target coordinates and all 160 test windows, centering each target coordinate on its own test-set mean. The additional contribution of the speckle representation is evaluated from the paired difference in . Uncertainty intervals are those of the archived whole-movie bootstrap analysis; the frozen test predictions and point scores were independently checked against the recovered analysis records. The displayed two-time-correlation examples are illustrative and have separate, incompletely recovered figure provenance.
Bragg-scattering model
The Bragg model consists of a two-dimensional crystalline lattice with reference positions and time-dependent displacements
where
The reversible elastic field undergoes spatially correlated fluctuations whose slowly varying amplitude is denoted by and whose spatial correlation length is denoted by . In addition, an internal load variable controls the probability of localized slip events. A slip produces a permanent shear-like contribution to , partially releases the stored load, and transiently enhances the subsequent mobility. The model therefore combines continuous correlated deformation with intermittent irreversible rearrangements.
The coherent intensity around a Bragg reflection is calculated as
with . Only the resulting detector intensities enter the measurement-based analysis. The compact Bragg state is
Here, is the conventional intensity autocorrelation and is the Pearson correlation between log-transformed detector intensity patterns separated by . Whereas measures the usual intensity correlation, emphasizes changes of the normalized spatial speckle pattern after centering and normalization of the log-transformed intensities. Because the transformation uses , this observable is not exactly invariant to multiplication of the raw intensity by a constant. The future target consists of the same two quantities in a later, non-overlapping window.
The observable search produces additional quantities describing -dependent activity and deformation of the Bragg pattern. The leading activity observable is based on frame-to-frame logarithmic intensity increments,
and measures a cross- contrast between strong activity in the outer Bragg halo and the broader fluctuation level. The second observable measures anisotropy of the Bragg-speckle pattern. From the intensity-weighted covariance matrix
is constructed from the separation of its principal values. Exact quantile definitions of –, together with the complete frozen feature grammar, are given in the Supplementary Information.
Controlled collective-motion SAXS model
The controlled SAXS model is designed to separate the amplitude of collective motion from its spatial range. Particle positions evolve according to
where represents independent Brownian motion and all particles belonging to the same active spatial domain receive the same collective displacement
The parameter controls the amplitude of collective motion, whereas denotes the linear extent of the square cooperative domains.
The coherent SAXS intensity is
The compact measured state is
where characterizes the mean intensity relaxation, while quantifies temporal fluctuations of the correlation function and thus the heterogeneity of the dynamics. Both quantities are averaged over the measured reciprocal-space range.
Temporal fluctuations of the instantaneous speckle activity are characterized from frame-to-frame intensity changes. Their variability within the present window is summarized by
The activity-CV comparison is separate from the guarded XPCS Alphabet search, which starts again from the compact state. The primary selected quantities describe global increment-energy fluctuations, fixed-ring amplitude and temporal variation of the spatial increment variance. The secondary forward-selection state additionally retains across-patch dispersion and -resolved intensity-change summaries. Their exact expressions are retained in the frozen analysis records.
Two-dimensional glass former
We simulate an overdamped two-dimensional 50:50 binary mixture of soft disks with particles and diameter ratio 1.4. The principal density series is
The additional densities were generated using the same dynamical and sampling protocol as the original long-film ensemble and were assigned to training, validation, and test sets before analysis. Structural diagnostics are reported in the Supplementary Information. Lengths are expressed in units of the root–mean square particle diameter and we use the fixed reciprocal-space reference
The coherent scattering amplitude and intensity are
The mean XPCS relaxation is characterized by the correlation function and by KWW fits of the form
yielding the relaxation time and stretching exponent . Particle coordinates are used only for independent physical diagnostics such as the mean-squared displacement and are not used for construction of the XPCS state or for observable selection.
The glass analysis contains three measurement-based steps. First, an XPCS information ladder tests which parts of the measured scattering signal recover information that is absent from a compact mean-correlation description. Second, the - and delay-dependent fluctuation amplitude and the organization of these fluctuations across XPCS channels are characterized. Third, a strictly non-overlapping later XPCS state is used to test how much of the detailed subsequent scattering dynamics can be predicted from the present measurement.
Glass predictive-state ladder
The Figure 4 benchmark compares three nested states constructed from measured intensities at . The baseline contains nine conventional intensity-product correlation excesses, at delays frames:
Each coordinate is averaged over a fixed causal calibration prefix ending 129 frames before the earliest prediction origin. Thus is constant within each movie: it represents a fixed-prefix mean-correlation reference, not a moving-window conventional . Here denotes the contrast excess without normalization by speckle contrast, rather than an angular Pearson correlation.
The second state retains time-resolved correlations of the registered speckle pattern. At each prediction origin , we use the offsets
and calculate, for each unordered pair and each ring,
where is the angular Pearson correlation between the two intensity patterns. The pairs per ring give decorrelation coordinates in . The measured state is .
The third state additionally retains radial scattering intensity. The original Figure 4 benchmark uses six archived coordinates: three frozen, density-dependent TRAIN reference intensities and three instantaneous radial-intensity residuals,
Here is the fluctuating radial intensity at the present observation time, not the time-averaged structure factor . The reference and residual coordinates use archived TRAIN scaling. This block therefore combines state-point reference information with instantaneous intensity variations; it is not an intensity history. The state dimensions are
Each state is evaluated with and without the same frozen 48-dimensional speckle-history representation , using identical origins and targets.
To distinguish the fixed-prefix baseline from time-resolved conventional XPCS, separate matched controls construct moving-window conventional correlations and pairwise conventional two-time intensity-product excesses on the same rings, endpoint pairs, and prediction origins. The full pairwise conventional and Pearson descriptions each contain 324 coordinates, including the common nine-coordinate prefix. Pearson correlations give a resolved predictive advantage over the matched conventional pairwise correlations when instantaneous radial intensity is not supplied. When both descriptions are augmented with the same three measured instantaneous log-radial-intensity coordinates, without an explicit density-dependent reference, that advantage is no longer statistically resolved. The augmented conventional description also has no statistically resolved additional contribution from , although predictive equivalence under the prespecified small-gain criterion is not established. These controls distinguish the contribution of temporal resolution from the effects of Pearson normalization and radial-intensity amplitude; numerical contrasts and confidence intervals are given in the Supplementary Information. Their conclusions concern the tested linear readouts, not information-theoretic sufficiency.
The prediction target contains nine future Pearson-decorrelation values at the three rings and relative delays . It compares the pattern at with later patterns; fractional delays use linear interpolation of decorrelation values at and . The present reference pattern is shared with the predictor history, but no future partner pattern enters any predictor. This target differs from the strictly frame-disjoint, 434-channel future-state analysis below.
All comparisons use Ridge regression, with input standardization and final fitting on TRAIN only. Regularization is selected by pooled VALIDATION mean-squared error over 0.1, 1.0, 100, 1000; the selected penalty is 1000 for each original Figure 4 head. At each of , the original cohort provides ten TRAIN and two VALIDATION movies. The principal evaluation uses 36 independent confirmatory TEST movies, twelve per density. Target scaling is fixed from TRAIN. Scores are equal-movie averages of multivariate , evaluated within each movie over nine target coordinates. Pointwise 95% intervals use 5,000 paired whole-movie bootstrap draws, conditioning on the frozen encoder and fitted predictors. The matched-control design, scaling, regularization selection, bootstrap contrasts, and limitations are detailed in Supplementary Information.
Causal XPCS Alphabet search for the glass
The glass XPCS Alphabet search was performed independently of the physically motivated Pearson-correlation hierarchy in Figure 4. Its baseline contains conventional intensity-product correlations, , at three scattering-vector rings and 14 delays. These mean correlations were calculated over a fixed earlier portion of each movie, preceding all present observation windows. The additional candidate quantities use only coherent intensities recorded up to the prediction time. No particle coordinates, packing-fraction labels, or density-dependent reference scattering profiles enter the search.
The predefined library contains 975 scalar candidates constructed from angular speckle-pattern correlations, current-frame intensity amplitudes, centered intensity covariances, angular intensity moments, cross-ring quantities, and combinations of correlations at different delays. Starting from the mean-correlation baseline, candidates are screened against the training prediction residual; the leading candidates are evaluated for their improvement of the same future XPCS target on validation movies. A new observable is accepted when the validation prediction score increases by at least 0.002, and no more than five quantities are added. Candidate standardization and predictor fitting use training movies only. This future-prediction search differs from the encoder-gain-guided selection used for Bragg and cooperative SAXS.
The future target contains Pearson speckle-pattern decorrelations at three rings and three horizons scaled to the relaxation time. In the primary test, the present frame serves as the reference for the later intensity pattern. A separate control forms both correlation endpoints from frames after the present observation window. Selection uses the original training and validation movies; the fixed observables are evaluated separately on original, regenerated, and confirmatory test cohorts. Prediction differences and their uncertainties are assessed by paired resampling of complete movies.
The automatically selected quantities are compared with systematic - and delay-resolved Pearson correlations and the current radial scattering intensity, without supplying a density-dependent structural reference. The search identifies useful observable families rather than uniquely specifying the full Pearson correlation field. Exact formulas, selection order, stability checks, and results for both future targets are provided in the Supplementary Information. This causal search is a separate prediction benchmark: its scores are not directly interchangeable with the original Figure 4 hierarchy.
Fluctuation amplitude and correlations in the glass
The dense-grid analysis evaluates the Pearson speckle-pattern decorrelation on 31 equally spaced rings over and 14 internal delays frames. At each time origin, a short increment compares two correlation cuts at fixed ,
where frames ranges from one to seven frames. For each movie, the 434-channel increments are centered across 48 sampled origins. We calculate the within-movie sample covariance and average the covariance matrices with equal movie weight. The diagonal increment-variance map is ; the factor 16 reflects the angular-sampling convention, and is not interpreted as a conventional .
We remove a fixed set of mean-relaxation directions estimated from independent TRAIN movies, forming . The projector removes 95 specified template directions from the 434-dimensional channel space, not every relaxation-related fluctuation. Normalization by the residual diagonal yields
The leading two eigenvectors characterize the channel participation , rather than local fluctuation amplitude. The RMS off-diagonal strength of is compared with a matched channel-wise circular-shift null, with subtraction performed on squared strengths,
The null preserves each channel’s circular autocorrelation but disrupts cross-channel temporal alignment. Exactly 20 independent movies per density enter this matched comparison, using 20-of-30 subsampling where applicable. The variance profiles are compared descriptively to ; no causal identification of variance with a relaxation rate is implied. Full projector templates, null construction, finite-size checks, spectral measures, and uncertainty procedures are documented in Supplementary Methods.
High-resolution future XPCS state and persistence controls
A separate benchmark tests a later speckle-pattern decorrelation on the same 31-ring, 14-delay grid. Its target is
with . Every future endpoint lies in , strictly after the present support . Unlike the nine-coordinate ladder target, both correlation endpoints belong to the future measurement, and delays are rounded rather than interpolated. The target construction, present states and Ridge predictors are separate from those used for Figure 4.
At each density, ten TRAIN, two VALIDATION and four TEST movies form the frozen split. Regularization and present-state choices use whole-movie cross-validation within TRAIN. We first test prediction of all 434 future coordinates. No robust prediction of the complete later pattern was established on independent TEST movies. We also examine lower-rank combinations of the target selected from TRAIN predictions made while holding out whole movies; TEST scores assess their support without selecting rank.
Persistence controls compare these predictions with the present values, a separate linear autoregression for each channel, and a local autoregression allowing the neighboring ring and delay channels. The fitted controls use TRAIN only. Cross-movie pairing at the same density and additional temporal shifts distinguish realization-specific prediction from persistence of the scattering pattern. Target sampling, rank-selection and support criteria, and the separate resampling protocols are specified in Supplementary Information. This test concerns prediction of a later measured pattern; it does not constitute a matched test of adding the covariance pattern to the Figure 4 state.
Statistical Analyses
Independent simulated speckle movies, rather than temporally overlapping windows, define the sampling units for training, validation, TEST evaluation, and uncertainty estimation. Normalization and model fitting use TRAIN only; regularization and observable choice use VALIDATION, and reported scores use held-out TEST movies. Prediction is quantified by for identical later scattering targets in the explicit and augmented comparisons. Brownian scores pool target-coordinate errors over TEST windows; Bragg and SAXS scores average coordinate-wise scores after pooling windows; glass scores are averaged with equal weight per test movie. Scores are therefore not directly comparable across benchmarks. Differences in use paired whole-movie resampling, retaining every window in each sampled movie. Where specified, 95% confidence intervals use 5,000 bootstrap draws and condition on frozen models and observable selections, rather than rerunning the complete selection procedure. The original Brownian archived intervals and separate causal-glass controls use their own sample counts, detailed in Supplementary Methods. Previous inspection of some archived TEST cohorts limits the Alphabet search results to exploratory held-out evaluations rather than prospective confirmation. The measured-state covariance analysis uses independent ensembles, matched movie counts, and a channel-wise temporal-shift null as described above.
Microscopic interpretation and inference
Microscopic variables are introduced only after the measurement-based observable search has been completed. For the Bragg and controlled SAXS models, the selected measured state
is paired with the corresponding known simulator parameters , allowing an inverse relation
to be learned. The mapping is not assumed to be one-to-one: several measured scattering observables may jointly constrain several microscopic degrees of freedom.
Data availability
The data supporting the findings of this study, including the processed scattering observables, predictive performance measures, and numerical results underlying the figures, are available from the corresponding author upon reasonable request. The simulated coherent scattering datasets were generated using the numerical models described in the Methods and Supplementary Information. Owing to their size, the complete time-resolved speckle datasets are not included with the manuscript.
Code availability
The custom code used for numerical simulations, coherent scattering calculations, and predictive representation learning in this study is available from the corresponding author upon reasonable request. The computational methods and model parameters are described in the Methods and Supplementary Information.
Acknowledgements
C.G. and A.T. acknowledge funding by DAPHNE4NFDI (DFG NFDI 40/1) and BMFTR (ErUM-Data projects 05D25PS2 and 0525PS1).
Author contributions
C.G. conceived and designed the project, performed the analyses, and wrote the manuscript. A.T., N.D.A., J.L., and M.M. contributed to scientific discussions, interpreted the results, and critically revised the manuscript. All authors approved the final version.
Competing interests
The authors declare no competing interests.
Declaration of AI use
Generative AI tools (ChatGPT and Codex, OpenAI) were used to assist with software development, including code generation, debugging, and implementation of numerical simulations and data analysis workflows, as well as language editing. The scientific concepts, methodological design, interpretation of results, and conclusions were developed by the authors. The computational results were assessed through physical consistency checks, benchmark calculations, and comparisons with established methods where applicable. The authors take full responsibility for the scientific content and conclusions of this work.
References
- [1]Sutton, M., Mochrie, S.G.J., Greytak, T., Nagler, S.E., Berman, L.E., Held, G.A., Stephenson, G.B.: Observation of speckle by diffraction with coherent x-rays. Nature 352, 608–610 (1991) https://doi.org/10.1038/352608a0
- [2]Grübel, G., Madsen, A., Robert, A.: X-ray photon correlation spectroscopy (xpcs). In: Borsali, R., Pecora, R. (eds.) Soft Matter Characterization, pp. 953–995. Springer, Dordrecht (2008). https://doi.org/10.1007/978-1-4020-4465-6_18
- [3]Perakis, F., Gutt, C.: Towards molecular movies with x-ray photon correlation spectroscopy. Physical Chemistry Chemical Physics 22, 19443–19453 (2020) https://doi.org/10.1039/D0CP03551C
- [4]Lehmkühler, F., Roseker, W., Grübel, G.: From femtoseconds to hours—measuring dynamics over 18 orders of magnitude with coherent x-rays. Applied Sciences 11(13), 6179 (2021) https://doi.org/10.3390/app11136179
- [5]Sutton, M., Laaziri, K., Livet, F., Bley, F.: Using coherence to measure two-time correlation functions. Optics Express 11(19), 2268–2277 (2003) https://doi.org/10.1364/OE.11.002268
- [6]Duri, A., Bissig, H., Trappe, V., Cipelletti, L.: Time-resolved-correlation measurements of temporally heterogeneous dynamics. Physical Review E 72, 051401 (2005) https://doi.org/10.1103/PhysRevE.72.051401
- [7]Trappe, V., Pitard, E., Ramos, L., Robert, A., Bissig, H., Cipelletti, L.: Investigation of q-dependent dynamical heterogeneity in a colloidal gel by x-ray photon correlation spectroscopy. Physical Review E 76(5), 051404 (2007) https://doi.org/10.1103/PhysRevE.76.051404
- [8]Perakis, F., Amann-Winkel, K., Lehmkühler, F., Sprung, M., Mariedahl, D., Sellberg, J.A., Pathak, H., Späh, A., Cavalca, F., Schlesinger, D., Ricci, A., Jain, A., Massani, B., Aubree, F., Benmore, C.J., Pettersson, L.G.M., Nilsson, A.: Diffusive dynamics during the high-to-low density transition in amorphous ice. Proceedings of the National Academy of Sciences 114, 8193–8198 (2017) https://doi.org/10.1073/pnas.1705303114
- [9]Wochner, P., Gutt, C., Autenrieth, T., Demmer, T., Bugaev, V., Diaz Ortiz, A., Duri, A., Zontone, F., Gruebel, G., Dosch, H.: X-ray cross correlation analysis uncovers hidden local symmetries in disordered matter. Proceedings of the National Academy of Sciences 106, 11511–11514 (2009) https://doi.org/10.1073/pnas.0905337106
- [10]Sanborn, C., Ludwig, K.F., Rogers, M.C., Sutton, M.: Direct measurement of microstructural avalanches during the martensitic transition of cobalt using coherent x-ray scattering. Physical Review Letters 107, 015702 (2011) https: //doi.org/10.1103/PhysRevLett.107.015702
- [11]Madsen, A., Leheny, R.L., Guo, H., Sprung, M., Czakkel, O.: Beyond simple exponential correlation functions and equilibrium dynamics in x-ray photon correlation spectroscopy. New Journal of Physics 12(5), 055001 (2010) https://doi.org/10.1088/1367-2630/12/5/055001
- [12]Ruta, B., Baldi, G., Monaco, G., Chushkin, Y.: Compressed correlation functions and fast aging dynamics in metallic glasses. The Journal of Chemical Physics 138(5), 054508 (2013) https://doi.org/10.1063/1.4790131
- [13]Ruta, B., Baldi, G., Chushkin, Y., Rufflé, B., Cristofolini, L., Fontana, A., Zanatta, M., Nazzani, F.: Revealing the fast atomic motion of network glasses. Nature Communications 5, 3939 (2014) https://doi.org/10.1038/ncomms4939
- [14]Giordano, V.M., Ruta, B.: Unveiling the structural arrangements responsible for the atomic dynamics in metallic glasses during physical aging. Nature Communications 7, 10344 (2016) https://doi.org/10.1038/ncomms10344
- [15]Dallari, F., Martinelli, A., Caporaletti, F., et al.: Stochastic atomic acceleration during the x-ray-induced fluidization of a silica glass. Proceedings of the National Academy of Sciences 120(2), 2213182120 (2023) https://doi.org/10.1073/pnas.2213182120
- [16]Shen, J., Yang, F., Cornet, A., Pineda, E., Martens, K., Chushkin, Y., Zontone, F., Festi, I., Ronca, A., Neuber, N., Frey, M., Busch, R., Cammarata, M., Michiel, M., Vaughan, G., Sprung, M., Ruta, B.: Length-scale dependence of the anomalous atomic motion in metallic glasses. Nature Physics 22, 672–679 (2026) https://doi.org/10.1038/s41567-026-03228-0
- [17]Chen, X.M., Farmer, B., Woods, J.S., Dhuey, S., Hu, W., Mazzoli, C., Wilkins, S.B., Chopdekar, R.V., Scholl, A., Robinson, I.K., De Long, L.E., Roy, S., Hastings, J.T.: Spontaneous magnetic superdomain wall fluctuations in an artificial antiferromagnet. Physical Review Letters 123, 197202 (2019) https://doi.org/10.1103/PhysRevLett.123.197202
- [18]Klose, C., Büttner, F., Hu, W., Mazzoli, C., Beach, G.S.D., Eisebitt, S., Pfau, B.: Photon correlation spectroscopy with heterodyne mixing based on soft x-ray magnetic circular dichroism. Physical Review B 105, 214425 (2022) https://doi.org/10.1103/PhysRevB.105.214425
- [19]Seaberg, M.H., Holladay, B., Lee, J.C.T., Sikorski, M., Reid, A.H., Montoya, S.A., Dakovski, G.L., Koralek, J.D., Coslovich, G., Moeller, S., Schlotter, W.F., Streubel, R., Kevan, S.D., Fischer, P., Fullerton, E.E., Turner, J.L., Decker, F.-J., Sinha, S.K., Roy, S., Turner, J.J.: Nanosecond x-ray photon correlation spectroscopy on magnetic skyrmions. Physical Review Letters 119, 067403 (2017) https://doi.org/10.1103/PhysRevLett.119.067403
- [20]Anth uparambil, N.D., Dargasz, M., Timmermann, S., Girelli, A., Retzbach, S., Möller, J., Jo, W., Raza, A.M., Leonau, A., Wrigley, J., Unger, F., Bin, M., Rajai ah, P.P., Andronis, I., Chèvremont, W., Hallmann, J., Rodriguez-Fernandez, A., Pudell, J.-E., Brausse, F., Boesenberg, U., Youssef, M., Shayduk, R., Rysov, R., Madsen, A., Lehmkuhler, F., Paulus, M., Zhang, F., Perakis, F., Schreiber, F., Gutt, C.: Lipoprotein diffusion in dense yolk plasma is governed by softness, hydrodynamics, and caging: Insights from MHz-XPCS. Proceedings of the National Academy of Sciences of the United States of America 123(8), 2519681123 (2026) https://doi.org/10.1073/pnas.2519681123
- [21]Dargasz, M., Anthuparambil, N.D., Retzbach, S., Girelli, A., Timmermann, S., Möller, J., Jo, W., Leonau, A., Raza Agha, M.B., Bin, M., Savvelkous, J., Andronis, I., Unger, F., Brausse, F., Hallmann, J., Boesenberg, U., Pudell, J.-E., Rodriguez-Fernandez, A., Wrigley, J., Shayduk, R., Youssef, M., Zozulya, A., Madsen, A., Lehmkuhler, F., Perakis, F., Zhang, F., Schreiber, F., Paulus, M., Gutt, C.: Depletion-induced interactions modulate nanoscale protein diffusion in polymeric crowder solutions. Proceedings of the National Academy of Sciences 123(31), 2524733123 (2026) https://doi.org/10.1073/pnas.2524733123
- [22]Mangu, A., Stoica, V.A., Zheng, H., Yang, T., Zhang, M., Wang, H.H., Zu, R., Nguyen, Q.L., Song, S., Das, S., Meisenheimer, P., Donoway, E., Chollet, M., Sun, Y., Turner, J.J., Freeland, J.W., Wen, H., Martin, L.W., Chen, L.-Q., Gopalan, V., Zhu, D., Cao, Y., Lindenberg, A.M.: Hidden domain boundary dynamics toward crystalline perfection. Proceedings of the National Academy of Sciences 122(2), 2407772122 (2025) https://doi.org/10.1073/pnas.2407772122
- [23]Konstantinova, T., Wiegart, L., Rakitin, M., DeGennaro, A.M., Barbour, A.M.: Noise reduction in x-ray photon correlation spectroscopy with convolutional neural networks encoder–decoder models. Scientific Reports 11, 14756 (2021) https://doi.org/10.1038/s41598-021-93747-y
- [24]Konstantinova, T., Wiegart, L., Rakitin, M., DeGennaro, A.M., Barbour, A.M.: Machine learning for analysis of speckle dynamics: quantification and outlier detection. Physical Review Research 4, 033228 (2022) https://doi.org/10.1103/PhysRevResearch.4.033228
- [25]Horwath, J.P., Lin, X.-M., He, H., Zhang, Q., Dufresne, E.M., Chu, M., Sankaranarayanan, S.K.R.S., Chen, W., Narayanan, S., Cherukara, M.J.: Ai-nerd: Elucidation of relaxation dynamics beyond equilibrium through ai-informed x-ray photon correlation spectroscopy. Nature Communications 15, 5945 (2024) https://doi.org/10.1038/s41467-024-49381-z
- [26]Cheng, M., Yu, B., Fu, C.-L., Andrejevic, N., Agne, M.T., Hanus, R., Wan, Q., Drucker, N.C., Nguyen, T., Fluerasu, A., Wiegart, L., Chen, X.M., Pajerowski, D., Cheng, Y., Turner, J.J., Snyder, G.J., Li, M.: Probing Non-Equilibrium Grain Boundary Dynamics with XPCS and Domain-Adaptive Machine Learning (2026). Timmermann, S., Starostin, V., Girelli, A., Ragulskaya, A., Rahmann, H., Reiser, M., Begam, N., Randolph, L., Sprung, M., Westermeier, F., Zhang, F., Schreiber, F., Gutt, C.: Automated matching of two-time x-ray photon correlation maps from phase-separating proteins with cahn–hilliard-type simulations using auto-encoder networks. Journal of Applied Crystallography 55, 751–757 (2022) https://doi.org/10.1107/S1600576722004435
- [28]Andrejevic, N., Zhou, T., Zhang, Q., Narayanan, S., Cherukara, M.J., Chan, M.K.Y.: Data-driven discovery of dynamics from time-resolved coherent scattering. npj Computational Materials 10, 225 (2024) https://doi.org/10.1038/s41524-024-01365-9
- [29]Shalizi, C.R., Crutchfield, J.P.: Computational mechanics: Pattern and prediction, structure and simplicity. Journal of Statistical Physics 104, 817–879 (2001) https://doi.org/10.1023/A:1010388907793
- [30]Littman, M.L., Sutton, R.S., Singh, S.: Predictive representations of state. In: Advances in Neural Information Processing Systems, vol. 14, pp. 1555–1561 (2001)
- [31]Tishby, N., Pereira, F.C., Bialek, W.: The information bottleneck method. In: Proceedings of the 37th Annual Allerton Conference on Communication, Control, and Computing, pp. 368–377 (1999)
- [32]Bialek, W., Nemenman, I., Tishby, N.: Predictability, complexity, and learning. Neural Computation 13(11), 2409–2463 (2001) https://doi.org/10.1162/089976601753195969
- [33]Creutzig, F., Globerson, A., Tishby, N.: Past-future information bottleneck in dynamical systems. Physical Review E 79(4), 041925 (2009) https://doi.org/10.1103/PhysRevE.79.041925
- [34]Wang, D., Tiwary, P.: State predictive information bottleneck. The Journal of Chemical Physics 154(13), 134111 (2021) https://doi.org/10.1063/5.0038198
- [35]Schmitt, M.S., Koch-Janusz, M., Fruchart, M., Seara, D.S., Rust, M., Vitelli, V.: Information theory for data-driven model reduction in physics and biology. bioRxiv (2024) https://doi.org/10.1101/2024.04.19.590281 . Preprint
- [36]Robinson, I., Harder, R.: Coherent x-ray diffraction imaging of strain at the nanoscale. Nature Materials 8, 291–298 (2009) https://doi.org/10.1038/nmat2400
- [37]D’Anna, N., Lamb, E.S., Glefke, R., Ham, D., Nihal, I., Lee, S.Y., Takamura, Y., Shpyrko, O.G.: Dynamic nanoscale spatial heterogeneity in a perovskite-to-brownmillerite topotactic phase transformation. ACS Applied Materials & Interfaces 18(23), 32795–32803 (2026) https://doi.org/10.1021/acsami.6c01097
- [38]Carr, A., Bowlan, J., Mazzoli, C., Barbour, A., Hu, W., Wilkins, S., Walker, C., Ding, X., Kim, J.H., Lee, N., Choi, Y.J., Lin, S.-Z., Sandberg, R.L., Zapf, V.S.: Evidence for the alternating next-nearest neighbor model in the dynamic behavior of a frustrated antiferromagnet. Physical Review B 103, 060401 (2021) https://doi.org/10.1103/PhysRevB.103.L060401
- [39]Sheyfer, D., Zheng, H., Krogstad, M., Thompson, C., You, H., Eastman, J.A., Liu, Y., Wang, B.-X., Ye, Z.-G., Rosenkranz, S., Phelan, D., Dufresne, E.M., Stephenson, G.B., Cao, Y.: X-ray-induced piezoresponse during x-ray photon correlation spectroscopy of PbMg1/3Nb2/3O3. Journal of Synchrotron Radiation 31, 55–64 (2024) https://doi.org/10.1107/S1600577523009116
- [40]Carnis, J., Gao, L., Labat, S., Kim, Y.Y., Hofmann, J.P., Leake, S.J., Schülli, T.U., Hensen, E.J.M., Thomas, O., Richard, M.-I.: Towards a quantitative determination of strain in bragg coherent x-ray diffraction imaging: artefacts and sign convention in reconstructions. Scientific Reports 9, 17357 (2019) https://doi.org/10.1038/s41598-019-53774-2
- [41]Clark, J.N., Ihli, J., Schenk, A.S., Kim, Y.-Y., Kulak, A.N., Campbell, J.M., Nisbet, G., Meldrum, F.C., Robinson, I.K.: Three-dimensional imaging of dislocation propagation during crystal growth and dissolution. Nature Materials 14, 780–784 (2015) https://doi.org/10.1038/nmat4320
- [42]Orr, K.W.P., Diao, J., Lintangpradipto, M.N., Batey, D.J., Iqbal, A.N., Kahmann, S., Frohna, K., Dubajic, M., Zelewski, S.J., Dearle, A.E., Selby, T.A., Li, P., Doherty, T.A.S., Hofmann, S., Bakr, O.M., Robinson, I.K., Stranks, S.D.: Imaging light-induced migration of dislocations in halide perovskites with 3d nanoscale strain mapping. Advanced Materials 35(46), 2305549 (2023) https://doi.org/10.1002/adma.202305549
- [43]Murphy, B.M., Götz, A., Gutt, C., McGuinness, C., Rønnow, H.M., Schneidewind, A., Deledda, S., Pietsch, U.: Fair data – the photon and neutron communities move together towards open science. IUCrJ 12(1), 8–15 (2025) https://doi.org/10.1107/S2052252524011941



