A pattern inside AI sharpened a nuclear-mass calculation

A neural network gave nuclear physicists a clue they could turn into a conventional calculation. After finding a pattern in the model’s internal representations, researchers connected it to a mathematical relation from 1969 and used that relation to improve the widely used WS4 nuclear-mass model. Its root-mean-square error fell from about 280 to 115 kiloelectronvolts across the evaluated measured nuclei.[1]

Kate Richardson, Sokratis Trifinopoulos and Mike Williams report the work in Communications AI & Computing on 21 October. They studied 2,325 experimentally characterised nuclei from the AME2020 mass evaluation, with proton and neutron numbers each at least 12 and additional measurement-quality criteria. The project asked whether a predictive model could reveal structure that physicists could understand and use independently.[1]

Why the weight of an atomic nucleus matters

A nucleus contains protons and neutrons held together by the strong interaction. Its mass is slightly less than the summed masses of its unbound constituents, reflecting the energy that binds the system. Tiny differences in binding energy influence whether a nucleus is stable, which radioactive transformations are possible and which chains of reactions can produce heavy elements in stars and stellar explosions. Direct many-body calculations are difficult for heavy nuclei, so physicists rely on a mixture of theory, empirical regularities and measured neighbours.[1]

The team gave its network two distinct sets of learned numerical embeddings, one representing proton number and one representing neutron number. Each embedding had 1,024 dimensions before further processing. When they reduced those representations to their dominant directions using principal-component analysis, the points arranged themselves in a double-helix-like pattern. Nearby relationships between the two strands corresponded to particularly stable nuclear combinations. It was a visualization of learned coordinates, not a claim that atomic nuclei are shaped like helices.[1]

A second finding lay in the network's smaller components. Those contributions approximately split into a function of proton number added to a function of neutron number. That separability—a complicated mass correction behaving like two simpler components—echoed a factorization described by Arthur Jaffe in work related to Garvey–Kelson relations in 1969. The authors call their recovered local form Jaffe factorization. AI did not invent the physical relationship from nothing; it directed attention back to a documented one and suggested a practical way to apply it.[1]

From a learned pattern to a local correction

The correction estimates small proton- and neutron-dependent shifts using nearby measured nuclei. Its 115-keV root-mean-square residual describes the spread of errors against reference binding energies under the paper’s evaluation procedure. The improvement belongs to WS4 combined with this local, interpretable correction; it is not simply the neural network’s accuracy.[1]

Each target nucleus was withheld from the local measurements used for its own correction. That protects the immediate comparison from using the answer being predicted, but the method still depends on known neighbours. Far beyond the measured region, those neighbours may be unavailable. The authors leave open whether the local factorization remains useful for the very neutron-rich nuclei involved in astrophysical rapid neutron capture.[1]

A different use for interpretability

Researchers often discuss AI interpretation in terms of explaining why a model made an individual decision. This study explores a more ambitious possibility: search the model's learned structure for a regularity that can be rewritten as a conventional mathematical rule. Once the factorization was visible, the physicists could test it directly and build a correction that no longer relied on asking a neural network for every prediction. The authors have made development code available, allowing other groups to inspect the procedure.[1]

The sequence is the useful scientific result: inspect a learned representation, identify a candidate relationship, and test that relationship as mathematics. The double-helix-like visualisation drew attention to structure in the model, while the subsequent correction supplied a numerical test of what researchers could extract from it.[1]

Report an error or suggest a correction