Systematic correction of core-loss spectra via machine learning: bridging the gap between simulated and experimental spectra.

Core-loss spectroscopy, including Energy Loss Near Edge Structures (ELNES) and X-ray Absorption Near Edge Structures (XANES), is a vital tool for materials characterization. However, its full potential is often hindered by the limited accuracy of computationally efficient simulations. Specifically, conventional methods like density functional theory with the generalized gradient approximation (DFT-GGA) often fail to reproduce experimental spectra perfectly. The DFT-GGA calculation is known to systematically underestimate energy separations, leading to a compression of spectral features compared to experimental results. To bridge this gap, we developed a machine learning framework utilizing the Random Forest algorithm, based on a dataset of 295 pairs of DFT-GGA and experimental K-edge spectra (C, N, O) from organic molecules. Our model demonstrated high fidelity in correcting the systematic errors of GGA calculations, accurately reproducing experimental spectral line shapes by learning a non-linear transformation. Analysis using artificial spectra reveals that the model has learned to numerically compensate for systematic discrepancies, producing spectral expansions that resemble self-energy corrections. Most significantly, the model exhibits remarkable generalization capabilities. Despite being trained exclusively on organic molecules, it successfully corrects the K-edge spectra of inorganic solid-state materials, for which it was not trained. These findings indicate that our model functions as a computationally efficient "Simulation-to-Experiment" corrector, providing a powerful tool for obtaining high-accuracy theoretical spectra at a fraction of the traditional cost, thereby accelerating high-throughput materials screening and providing reliable theoretical benchmarks to facilitate the interpretation of complex experimental data.