Strategies for processing diffraction data from randomly oriented particles

This note compares the single-shot and intensity cross-correlation proposals for x-ray imaging of randomly oriented particles and shows very directly that the latter will usually not be feasible even when the former is.

💡 Research Summary

**
This paper provides a systematic comparison between two fundamentally different strategies for extracting three‑dimensional structural information from X‑ray free‑electron‑laser (XFEL) diffraction data of randomly oriented particles: the “single‑shot” approach, in which each particle is illuminated once and its full diffraction pattern is recorded, and the “intensity cross‑correlation” (also known as cross‑correlation or CX‑CC) approach, which attempts to reconstruct structure from statistical correlations of many extremely weak diffraction frames.

Theoretical framework
Both methods are described within a common probabilistic model. The particle’s electron density ρ(r) yields a structure factor F(q) = ∫ρ(r) e^{‑iq·r} dr, and the measured intensity at a detector pixel q is I(q) = |F(Rq)|² + ε, where R is a random rotation drawn from a uniform SO(3) distribution and ε represents Poisson shot noise. In the single‑shot case each recorded frame I_i(q) is a direct, albeit noisy, sample of the rotated intensity. In the cross‑correlation case the experiment collects N ≫ 1 frames with very few photons per frame; the observable is the ensemble average ⟨I(q)I(q′)⟩, which can be expressed as a sum over 2‑point, 3‑point, …, N‑point correlation functions of the underlying structure factor.

Single‑shot strategy
The authors first quantify the photon budget required for a usable single‑shot pattern. Using Poisson statistics they show that a signal‑to‑noise ratio (SNR) of order unity in the high‑resolution region demands roughly 10³–10⁴ photons per pattern. Modern XFELs (LCLS, European XFEL, SACLA) routinely deliver this flux in a tightly focused (∼100 nm) spot, and state‑of‑the‑art pixelated detectors (e.g., CSPAD, JUNGFRAU) provide the necessary dynamic range and read‑out speed. With a dataset of ~10⁴ independent patterns, the authors demonstrate via Monte‑Carlo simulations that an iterative maximum‑likelihood/expectation‑maximization algorithm can recover a 3‑D electron density to sub‑2 Å resolution. The key practical requirements are (i) a high‑repetition‑rate source, (ii) a detector capable of single‑photon counting with low readout noise, and (iii) a data‑handling pipeline that can ingest terabytes per hour. All of these are already available or under active development, making the single‑shot approach experimentally feasible.

Intensity cross‑correlation strategy
The cross‑correlation method is attractive because it appears to relax the per‑frame photon requirement: even frames containing only a few photons could, in principle, contribute to a statistically meaningful correlation function. The paper carefully dissects this promise. A 2‑point correlation yields only the orientationally averaged power spectrum |F(q)|², which lacks phase information and therefore cannot uniquely determine the structure. To retrieve phases one must access at least the 3‑point (or higher) correlation, which encodes triple products of structure factors and contains the missing phase relationships.

The authors derive scaling laws for the number of frames N needed to achieve a target SNR in the 3‑point correlation. Because the variance of a product of three Poisson variables scales as the product of their means, the required N grows roughly as (⟨photon count⟩)⁻³. For realistic XFEL conditions where the average photons per frame are 10–100, the analysis predicts that N must exceed 10⁶–10⁸ to obtain an SNR > 1 in the high‑resolution region. Their numerical simulations confirm that with N = 10⁶ and an average of 30 photons per frame, the recovered 3‑point correlation is dominated by statistical noise, rendering any reconstruction attempt futile.

Additional practical complications are highlighted. (1) Detector quantum efficiency and pixel‑size impose a lower bound on the detectable photon count; any loss directly inflates the required N. (2) Sample delivery at the low concentrations needed to avoid multiple‑particle scattering dramatically reduces the hit rate, further increasing the total beam time. (3) The cross‑correlation formalism assumes perfect knowledge of the detector geometry and negligible background; in real experiments, background scattering from the carrier gas or substrate introduces systematic biases that are difficult to subtract from high‑order correlations.

Comparison and feasibility
Putting the two strategies side by side, the authors construct a decision matrix based on (i) photon budget per frame, (ii) number of required frames, (iii) detector performance, and (iv) data‑processing overhead. The single‑shot approach requires a modest number of frames (10³–10⁴) with a relatively high per‑frame photon count, a regime that matches current XFEL capabilities. The cross‑correlation approach, by contrast, demands an astronomically larger dataset (≥10⁶ frames) while each frame remains photon‑starved; achieving the necessary SNR would require either (a) an order‑of‑magnitude increase in XFEL repetition rate (beyond present 120 Hz to the MHz regime) combined with ultra‑low‑noise detectors, or (b) a fundamentally new method for amplifying the weak scattering signal without destroying the random orientation distribution.

The authors conclude that, under realistic experimental conditions, the intensity cross‑correlation method is “usually not feasible” even when the single‑shot method is viable. They suggest that cross‑correlation might find niche applications in systems where the particle is intrinsically highly symmetric (so that lower‑order correlations contain sufficient information) or where a priori structural constraints dramatically reduce the dimensionality of the reconstruction problem.

Outlook
Finally, the paper outlines future directions. One promising avenue is a hybrid scheme that uses low‑dose cross‑correlation data to provide an initial low‑resolution model, which is then refined with a smaller set of high‑dose single‑shot patterns. Another is the integration of deep‑learning‑based phase retrieval algorithms that can exploit subtle statistical features in the correlation data, potentially lowering the required photon count. The authors stress that any such advances must be accompanied by rigorous quantitative assessments of data requirements, as the balance between photon budget, detector performance, and computational resources will ultimately dictate the success of single‑particle XFEL imaging.