Code and Dataset for Deep-learning-Enabled Resolution-Adaptive Nuclear Spectroscopy with Synchrotron Mössbauer Sources

Nuclear resonant spectroscopies at synchrotron Mössbauer sources are fundamentally constrained by a trade-off between spectral resolution and photon intensity, which limits access to weak signals and extreme sample environments. Here, we show that a deep learning–based inverse model can effectively relax this trade-off by recovering high-resolution nuclear response functions from broadened and noisy synchrotron Mössbauer spectra. A resolution-adaptive neural network is trained entirely on physics-based synthetic data spanning a broad range of instrumental linewidths and noise conditions, and learns to reconstruct the intrinsic hyperfine spectra without explicit knowledge of the instrumental function. Applied to ⁵⁷Fe synchrotron Mössbauer measurements on metallic iron, the method retrieves the transition spectrum in quantitative agreement with high-resolution reference data while operating at up to 17-fold higher count rates. Its effectiveness is further confirmed on complex oxides. The approach remains robust across mixed noise statistics and uncalibrated linewidths, and can be combined with learned denoising networks for very low signal-to-noise regimes. Our results demonstrate that data-driven inverse modeling can transform the effective performance of nuclear resonant spectrometers and provide a general framework for overcoming resolution–throughput constraints in photon-counting spectroscopies.