Learning Nuclear Structure with AI: Radii and Collectivity
Giuliano Giacalone, Sokratis Trifinopoulos, Mike Williams
nucl-th · cs.AI · cs.LG · nucl-ex
Quick verdict
NuCLR multi-task OOF ensembles reach 0.0147 fm and 0.192 e^2b^2 RMS, competitive with EDF models; rigorous validation but unequal baseline comparison sets.
Summary
The paper extends NuCLR, a task-conditioned multi-task neural network over the nuclear chart, into held-out ensemble campaigns for two structure observables: charge radii and electric-quadrupole transition strengths B(E2). Measured nuclei are split into ten folds; four seeded models per fold withhold the same labels (40 models per campaign), so every measured value is predicted strictly out-of-fold. The B(E2) campaign additionally masks same-nucleus quadrupole proxies (E(2+), E(4+), R4/2, B(E2;4->2), static moments) on the active fold to prevent leakage. Multi-task learning cuts OOF RMS from 0.0760 to 0.0147 fm (radii, 868 nuclei) and from 0.295 to 0.192 e^2b^2 (B(E2), 433 nuclei), comparable to BSkG3 and Gogny-D1S 5DCH baselines, though those are evaluated on smaller, non-identical overlaps. Uncertainty bars come from cross-fitted locally adaptive conformal regression: a gradient-boosted residual-scale model calibrated on disjoint folds achieves near-nominal coverage under random holdout but degrades under contiguous-region holdout, and the authors correctly decline to call extrapolated widths calibrated intervals. Case studies (Ni, Sm, Pb; Sm, Po, U) highlight data-driven versus theory-driven divergences proposed as measurement targets; known failures (Hg odd-even staggering, Zr N=58-60 jump) are candidly reported.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
4 strengths · 4 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Sharing one representation across observables (multi-task learning) improves out-of-fold accuracy dramatically over single-task learning: charge-radius RMS 0.0760 to 0.0147 fm and B(E2) 0.295 to 0.192 e^2b^2.
- NuCLR's OOF accuracy on measured nuclei is competitive with the BSkG3 and Gogny-D1S 5DCH EDF models (radii 0.0147 vs 0.0233/0.0255 fm; B(E2) 0.192 vs 0.192/0.175 e^2b^2).
- Cross-fitted locally adaptive conformal regression yields near-nominal 68% coverage (0.689/0.693) on randomly held-out measured nuclei, with acknowledged degradation under contiguous-region holdout.
- Data-driven and EDF extrapolations diverge in specific unmeasured regions (neutron-rich Ni, Po, U/Th), identifying measurements that would discriminate empirical from theory-driven continuations.
Strengths
Leakage-aware out-of-fold protocol
Every measured target is predicted only by the four models that withheld it (Eq. 2); the B(E2) campaign additionally masks same-nucleus quadrupole proxies (E(2+), E(4+), R4/2, B(E2;4->2), static moments) on the active fold and stratifies folds by B(E2) magnitude and Z region (SM S2.B) - an unusually careful anti-leakage design for this data type.
Quantified multi-task benefit via internal ablation
The STL-to-MTL comparison (Table I) isolates the value of the shared representation: radius RMS 0.0760 to 0.0147 fm (5.2x better) and B(E2) 0.295 to 0.192 e^2b^2 on the full measured sets, making the paper's central claim falsifiable from its own tables.
Honest, empirically audited uncertainty
Tail diagnostic matched to the observable
Weaknesses
Unequal evaluation sets underlie the competitiveness claim
Extrapolations are not out-of-fold
Approximate B(E2) baseline for BSkG3
No external ML baselines
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is NuCLR and what does this paper add to it?
NuCLR (Nuclear Co-Learned Representations) is a neural network that takes a nucleus (proton number Z, neutron number N) plus an observable label and predicts that observable, sharing one learned representation across many types of nuclear data. This paper adds rigorous held-out ensemble validation of NuCLR for charge radii and B(E2) transition strengths, plus a conformal-regression scheme that assigns error bars learned from out-of-fold residuals.
How accurate are NuCLR's charge-radius and B(E2) predictions?
On out-of-fold predictions over 868 measured charge radii, NuCLR reaches 0.0147 fm RMS, versus 0.0760 fm for single-task learning, 0.0233 for BSkG3 and 0.0255 for the Gogny-D1S 5DCH model. For 433 measured B(E2) values it reaches 0.192 e^2b^2 RMS, on par with BSkG3 (0.192) and slightly behind 5DCH (0.175).
Can I trust the error bars NuCLR assigns to unmeasured nuclei?
Not as calibrated confidence intervals. The authors state explicitly that widths at unmeasured nuclei are learned estimates of local prediction difficulty, not validated 68% intervals; under contiguous-region holdout, coverage degrades (95% coverage falls to roughly 0.87-0.91), and the exchangeability assumptions behind conformal guarantees do not hold in extrapolation regions.
Where do NuCLR and traditional nuclear models disagree most?
The paper highlights neutron-rich nickel (NuCLR predicts kinks and a flattening of charge radii for 40<N<50 while EDF models stay smooth), polonium B(E2) away from N=126 (flat NuCLR trend versus steep collective-model variations), and neutron-rich uranium/thorium collectivity, proposing these as high-value targets for future measurements.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures (Figs. 1-2, S1-S5) are available only via captions and prose; visual chain features (kinks, staggering, error-bar widths) cannot be independently verified from the extracted text.
- The GitHub data repository (ref. 64) could not be inspected, so availability of code versus data tables is unconfirmed beyond the citation; codeAvailable reflects the stated data repository.
- Table S1 and the supplemental protocol appear completely extracted, but any numbers embedded inside figure graphics would be invisible to this review.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.