Two-dimensional (2D) X-ray absorption near-edge structure (XANES) imaging offers the potential to probe local chemical reactions at the nanoscale with high spatial resolution. However, achieving high sensitivity and accuracy in 2D XANES, particularly for low-concentration elements, remains a significant challenge due to poor signal-to-noise ratios (SNR) in the measured images. This limitation hinders quantitative analysis and accurate determination of elemental valence states and local chemical environments. Two main techniques for acquiring 2D XANES data are scanning fluorescence X-ray microscopy (XRF-XANES) and full-field transmission X-ray microscopy (TXM-XANES). XRF-XANES offers high sensitivity but suffers from slow data acquisition times, limiting its applicability to large samples and dynamic systems. TXM-XANES, while efficient and providing a large field of view, has relatively poor sensitivity due to inherent low SNR associated with full-field imaging. This study aims to address the sensitivity limitations of TXM-XANES by developing a novel post-processing method based on physics-informed machine learning (ML) to significantly improve SNR and enable more accurate quantitative analysis of chemical states, even for low-concentration elements. The improved sensitivity is crucial for numerous applications in materials science, catalysis, and energy storage research where nanoscale chemical heterogeneity plays a vital role.

Conventional XANES spectroscopy provides bulk-averaged information, neglecting the inherent heterogeneity within materials. Spatially resolved XANES imaging techniques, such as scanning fluorescence X-ray microscopy (XRF-XANES) and full-field transmission X-ray microscopy (TXM-XANES), have been developed to overcome this limitation. XRF-XANES achieves nanoscale resolution by scanning a focused X-ray beam across the sample, but the slow data acquisition poses significant limitations. TXM-XANES, on the other hand, provides high efficiency and large field of view but suffers from relatively low signal-to-noise ratios. Existing image processing techniques, such as non-local means, total variation regularization, and BM3D, can effectively reduce random noise, but they struggle to eliminate systematic noise associated with beam profile variations, impacting accurate analysis. Machine learning (ML), specifically deep neural networks, has emerged as a powerful tool for image processing tasks, often outperforming traditional algorithms in terms of accuracy and efficiency. The application of ML in enhancing XANES imaging has been limited, emphasizing the need for advanced ML methods tailored to the unique challenges of XANES data.

This work presents a physics-informed deep learning model based on the residual-in-residual dense block (RRDB) network architecture for background normalization in TXM-XANES images. The model uses a RRDB network that takes noisy XANES images as input and predicts the beam backgrounds. Instead of directly predicting the denoised images, the model focuses on accurately estimating the background profile, which varies less than the sample foreground. The training dataset is a fusion of synthetic and experimental data. Synthetic Ni K-edge absorption images are generated using the Lambert-Beer law and absorption coefficients from the Xraylib database, combined with experimentally acquired beam intensity profiles. A Gaussian filter removes high-frequency noise from the background images, which are used as ground truth during training. The training process aims to extract the background from the input images and eliminate both the background and random noise to obtain the denoised XANES images. Multiple loss functions are employed to optimize the model performance. Mean squared error (MSE) loss and feature loss (VGG loss) quantify the difference between predicted and ground truth backgrounds. Additional loss functions are included to incorporate domain knowledge of the XANES spectrum's stepwise increase across the absorption edge. Equation (1) which describes a relationship between attenuated intensity (I), energy (E), and absorption thickness (A) among other parameters, allows for thickness consistency fitting across predicted and ground-truth XANES spectra. Consistency in fitted images, ensuring alignment between the low-frequency part of spectroscopic images reconstructed from model parameters and the denoised images, also improve the model. This multi-loss function strategy accelerates convergence, enhances model stability, and improves denoising performance. The trained model is then applied to experimental XANES data to evaluate its performance.

The proposed physics-informed ML method significantly improves the accuracy of chemical state mapping, particularly for low-concentration elements, as demonstrated through analysis of Ni valence states in LiNiO2 precursor samples. The method successfully reduces noise in the XANES spectra, particularly in low-absorption regions, leading to more consistent and robust results. Applying the model to two separate measurements on the same sample area produced almost identical mappings of the Ni fraction, in contrast to the vastly different results obtained using traditional methods. Analysis of pixel-wise Ni³⁺ fraction ratios revealed a narrow Gaussian distribution with a mean of 1 and a standard deviation of 0.006, indicating excellent consistency between the measurements (less than 1.8% difference for 99.7% of the particle area). The model's generalizability was demonstrated by applying the pre-trained model to correct noisy experimental XANES images for different elements (Ni, Co, Mn, Fe). While the model showed good transferability to most elements, a self-supervised learning approach was introduced to address cases with low signal response. This self-supervised method uses a subset of images from the pre-edge and post-edge of the XANES spectrum to further train the model without requiring additional labeled data. Through iterative self-supervised learning, the model significantly improved the signal-to-noise ratio in the Co spectrum of LiNi0.85Co0.1Mn0.05O2, enabling the successful characterization of Co valence states in this material. This analysis revealed a non-uniform distribution of Co valence states, varying from +2.7 to +3, challenging the common assumption of +3 valence for Co in LiNixCoyMnzO2. K-Means clustering identified three major clusters with distinct spatial distributions and correlations between Co and Ni valence states and atomic concentrations, highlighting the importance of considering this nanoscale chemical heterogeneity.

The results demonstrate the significant improvement in sensitivity and accuracy achieved by the physics-informed ML method for 2D XANES analysis. The ability to accurately determine valence states, even for low-concentration elements, opens new possibilities for studying nanoscale chemical heterogeneity in various materials. The consistent and robust results obtained from multiple measurements of the same sample highlight the reliability of the method. The successful application of the model to various elements and the development of the self-supervised learning approach enhance its generalizability and applicability to different experimental datasets. The findings on the non-uniform distribution of valence states in LiNi0.85Co0.1Mn0.05O2 provide new insights into the complex interplay between composition and valence state, with potential implications for optimizing material synthesis and performance.

This study presents a novel physics-informed machine learning method for enhancing the sensitivity and accuracy of 2D XANES imaging. The method successfully addresses the limitations of traditional methods by effectively reducing noise and enabling the characterization of low-concentration elements. The self-supervised learning approach further improves the model's transferability and adaptability to new datasets. Future research could focus on expanding the model's application to other types of X-ray imaging and exploring its potential for real-time data analysis during dynamic experiments, like in operando battery studies.

The current study focuses on TXM-XANES data, and the generalizability to other imaging modalities may require further investigation. The accuracy of the method relies on the quality of the reference spectra used for fitting. While self-supervised learning improved the model's transferability, it may not always achieve the same level of accuracy as supervised learning with extensive labeled data. The synthetic training data, while designed to be as realistic as possible, may not fully capture all aspects of real experimental data, potentially limiting the method's performance in certain cases.