The exponential growth of experimental and theoretical data in atomic collision physics demands new strategies beyond traditional tabulations. We present a paradigm shift toward intelligent databases: curated, self-consistent, and predictive data collections powered by machine learning (ML). Our approach combines unsupervised learning for automated data cleaning with deep neural networks (NNs) for interpolation and prediction across broad parameter spaces.
We illustrate this framework with two recent open-source tools: ESPNN [1] and IKEBANA [2]. The code ESPNN uses a DBSCAN-based filtering algorithm to remove outliers from the IAEA stopping power database [3] and trains a deep NN to predict electronic stopping power cross sections for any ion–atomic target pair with a mean absolute percentage error (MAPE) below 6%.
The code IKEBANA, on the other hand, is a machine learning model trained on a recent and comprehensive compilation of experimental K-shell ionization
cross sections [4]. It uses only atomic number and overvoltage as inputs, achieving R2 > 0.997 on unseen test data, even for elements with no experimental measurements.
We extend this methodology to new domains: (i) an NN trained on extensive Continuum Distorted Wave (CDW) calculations reproduces ion-impact ionization cross sections with high fidelity; and (ii) a novel transformer-based architecture is under development to predict stopping power in molecular targets, overcoming limitations of atomic additivity rules. These ML-driven databases not only reconcile historical discrepancies but also guide future experiments by identifying regions of high uncertainty or missing data. By transforming passive compilations into active, predictive knowledge engines, intelligent databases represent a powerful tool for both fundamental research and applied simulations in atomic physics.
References:
[1] F. Bivort Haiek et al., J. Appl. Phys. 132, 245103 (2022).
[2] D.M. Mitnik et al., Atoms 13, 80 (2025).
[3] https://www-nds.iaea.org/stopping/
[4] S.P. Limandri et al., At. Data Nucl. Data Tables 166, 101756 (2024).
