A Fully Convolutional Approach to Denoising 2D Correlation Spectra

Encoder–decoder architectures are powerful tools for denoising scientific data, yet many existing models are constrained by fixed input dimensions and poor generalization outside their training domain. Here, we address these challenges in two-dimensional representations of dynamic correlations, which are fundamental to techniques like X-ray photon correlation spectroscopy (XPCS), electron microscopy, and IR spectroscopy. Using XPCS two-time intensity correlation functions (C2) as a primary benchmark, we tackle photon-limited noise and detector artifacts where conventional image-processing filters fail. Because noise in correlation data is highly structured and non-Gaussian, our approach leverages domain-informed machine learning to preserve subtle, physically meaningful dynamic features without introducing artifacts.

We present a fully convolutional denoising autoencoder (FC-DAE) consisting entirely of convolutional layers. The model is trained on experimentally derived C2 matrices from NSLS-II beamlines, employing data augmentation to enhance diversity and mitigate overfitting. By eliminating fixed fully connected layers, the architecture exploits translation-invariant spatial features, allowing inference on C2 maps of arbitrary dimensions without resizing or cropping. This flexibility enables denoising strategies learned from training data to generalize across diverse dynamic patterns. Furthermore, we evaluate reconstruction fidelity using quantitative reliability metrics to detect potential model bias, establishing a robust framework for trustworthy AI-assisted analysis.

The results demonstrate that the FC-DAE effectively suppresses noise while preserving fine correlation features that are typically distorted or oversmoothed by standard latent-space autoencoders. The resulting gain in signal-to-noise ratio enables more reliable parameter extraction, reduces required temporal averaging, and expands the accessible experimental window under low-signal and low-dose conditions. Although established on XPCS data, the architecture is inherently generalizable to two-dimensional dynamic correlations across other domains, including electron microscopy. Overall, this work highlights fully convolutional models, paired with quantitative reliability assessments, as a robust, scalable framework for scientific data recovery across experimental disciplines.