This paper reviews how machine learning (ML) can help recover high-precision predictions from ab initio nuclear calculations that are stuck in reduced model spaces. Ab initio means “from first principles” and refers to theories that start from the basic forces between protons and neutrons. The author surveys recent methods that learn how calculated results change as the model space grows, and then extrapolate those results toward the full, unreachable solution with quantified uncertainties.

The central technical problem is that modern many-body calculations must cut the full quantum space down to a tractable size. One widely used method is the no-core shell model (NCSM). In NCSM, two key settings control the calculation: Nmax, which limits how many excitations are included, and the harmonic-oscillator frequency ℏΩ, which fixes the single-particle basis. Energies in these truncated spaces converge toward the true value from above (because of the variational principle). Other observables, such as nuclear radii, can converge more irregularly and may approach the correct value from either side. The review discusses results computed with several interaction models, for example chiral effective field theory forces (EMN[Λ] and SMS[Λ]) and the Daejeon16 interaction.

The review compares traditional, physics-motivated extrapolation formulas with data-driven ML approaches. A common simple fit for ground-state energies is an exponential form E(Nmax)=E∞+a exp(−b Nmax), where E∞ is the extrapolated energy. One improvement is an “ensemble fit” that forces data at different ℏΩ values to share the same E∞. Machine learning approaches, especially artificial neural networks (ANNs), learn the mapping from truncated-space inputs to full-space outputs from examples. The paper includes a short introduction to ANNs: they combine simple computational “neurons” in layers, tune connection weights during supervised training, and then predict unseen cases. Other ML tools mentioned in the review include Gaussian-process emulators (statistical surrogates) and Bayesian optimization for parameter searches.