Accelerating Lanthanide and Actinide Opacity Calculations with Machine-Learning Models

Lanthanides and actinides are heavy elements produced in neutron-star mergers that
dominate the opacity of kilonova ejecta through complex f-shell line spectra, but calculating their bound–bound opacities with atomic-structure codes is computationally
extremely costly. This work introduces a two-stage machine-learning approach that replaces those calculations. To address this, the NIST-LANL opacity database — comprising over 12000 high-resolution spectra spanning 28 elements, 27 temperatures and 17
densities — was collected into a unified dataset for machine-learning. A convolutional
autoencoder first reduced a down-sampled 200-bin opacity spectrum to a low-dimensional
latent space, which a gradient-boosted regressor then mapped back to opacity. This surrogate was able to reproduce Planck-mean values with errors on the order of 103
cm2/g
and reduced spectral RMSE by more than 50% relative to a nearest-neighbour baseline.
Next, a convolutional classifier analyzed these spectra, inferring element identity with
99.7% accuracy, plasma temperature within 0.19 eV RMSE, and mass density within
8.35×105 RMSE. Although aggressive down-sampling smoothed out the narrowest resonances and performance degraded in the most highly ionized regimes, the surrogate
evaluated opacities orders of magnitude faster than conventional atomic-structure codes.
Future work on attention-based models to recover the narrow peaks, physics-informed
loss functions to enforce atomic constraints, and uncertainty-driven active learning could
sharpen accuracy and enable transparent, uncertainty-aware opacity calculations in nextgeneration astrophysical simulations.