The atomic nucleus rarely sits still: it often trembles, rotates, or pulses like a plucked string. Each such motion is an excited state, and describing it usually requires extremely complex calculations. Now neural networks come to the rescue: they shape the nucleus the way a musician identifies a note in a chord by ear.

The authors took the method of neural-network quantum states (NQS)—where the wave function of particles is represented not by a formula but by a trained network, and added two ways to find excitations. The first, overlap penalty (OP), penalizes the network for resembling already found states, forcing it to seek new ones. The second, natural excited states (NES), directly varies the energy for each level. For hypernuclei — nuclei that, besides protons and neutrons, host a strange particle called lambda (Λ)—they had to solve the problem of "spin contamination": the network could confuse states with different spins. Here they devised quantum number targeting (QNT), an additional filter that keeps the network in the desired spin channel.

The result is impressive: the energies and spatial shapes computed by both methods match exact solutions. Moreover, for the first time, they calculated from first principles the strength of the M1 magnetic transition in hyperhydrogen-4 (⁴_ΛH). It turns out the transition is suppressed by about 1.3%—as if the lambda particle slightly dampens the nucleus's magnetic response. This agrees with the picture of weak coupling, where the strange guest barely interferes with the nucleons' dance.

Now neural networks not only find the ground state of a nucleus but also decompose its spectrum into notes. Such a tool is especially valuable for rare hypernuclei that live for billionths of a second and elude experimentalists. Computations replace direct experiment, and do so with musical precision.