Nuclear-mass prediction swaps the gated recurrent candidate state: complex product units bring interpolation error to 0.2698 MeV
Synopsis
Machine-learning predictions of nuclear masses and charge radii usually feed neighboring nuclides into a gated recurrent network as an ordinary sequence. AM-PU-GRU embeds learnable product-unit mappings in the GRU candidate state and works in the complex domain to keep both amplitude and phase, reaching a mean mass interpolation RMSE of 0.2698 MeV and a temporal RMSE of 0.1905 MeV across three data partitions with three training seeds each. Without mass inputs, MI-PU-GRU predicts charge radii with an RMSE of 0.019947 fm, 14.40% below the fixed WS∗ prior.
Interpretation
The candidate state of a gated recurrent unit no longer performs only an affine transformation; learnable product mappings are added so multiplicative dependencies among neighboring nuclei can be trained. Earlier recurrent mass predictions used standard GRU, RNN or LSTM structures with eleven features per nucleus, computing the candidate state through conventional gating. Across three data partitions with three training seeds each, nine runs in total, complex AM-PU-GRU reaches a mean mass interpolation RMSE of 0.2698 MeV and a temporal RMSE of 0.1905 MeV, the lowest among the models in Tables 1 and 3.
Each input is raised to a learnable exponent and multiplied together, and this product path is placed in the complex domain so amplitude scaling and phase rotation both enter the candidate state. Product units had mostly been used in feedforward networks, and their complex versions mainly in feedforward structures; here they enter the recurrent candidate computation while the standard update gate still mixes states. The mass task builds a five-position sequence from four nearest measured neighbors plus the target's WS4 prior; the charge-radius task instead uses four measured neighbor radii plus a WS∗ target estimate, with no mass features in the input.
On charge radii the gain depends on which candidate path carries the product mapping, rather than the product unit helping uniformly. Against the fixed WS∗ prior, MI-PU-GRU lowers radius RMSE by 14.40%; yet in architecture-matched replacements, swapping the MI path for Tanh gives a 0.21% lower RMSE. AM's product mapping lowers RMSE by 9.98% relative to Tanh and 7.07% relative to SiLU, improving all five seed pairs; MI lowers it by 1.28% relative to SiLU, also improving all five pairs, while MI versus Tanh has an interval spanning zero.
Perspective
Evaluation is limited to AME2020 mass labels and the 2013 experimental charge radii; the mass task uses random 70/15/15 interpolation and temporal extrapolation settings, and the radius task uses a 70/15/15 held-out split with the fixed four-parameter WS∗ prior. The intended reader is a researcher interested in machine-learning prediction of nuclear properties; the stated next direction is extending complex product-unit recurrent networks to other nuclear observables and broader extrapolation settings.
For radii, MI's paired gain intervals against both GRU baselines and the MLP include zero; omitting the three most favorable nuclides, 24Ne, 32Ar and 246Cm, reverses the ordering against real GRU, and 152Tb, 151Tb and 149Tb account for 44.6% of MI's total squared error. The authors also note that different data partitions change errors and model rankings.
