arxiv: 2604.05312 · v1 · submitted 2026-04-07 · ⚛️ nucl-th · cs.LG· physics.atom-ph

Recognition: no theorem link

Predictions of charge density distributions for nuclei with Z geq 8

Yun Dong Wang , Tian Shuai Shang , Hui Hui Xie , Peng Xiang Du , Jian Li , Haozhao Liang

Authors on Pith no claims yet

Pith reviewed 2026-05-10 19:41 UTC · model grok-4.3

classification ⚛️ nucl-th cs.LGphysics.atom-ph

keywords nuclear charge densitydeep neural networkcharge radiirelativistic continuum Hartree-BogoliubovFourier-Bessel expansionnuclear structuremachine learning

0 comments

The pith

A deep neural network predicts nuclear charge density distributions with smaller errors than the relativistic calculations used to train it.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper trains a deep neural network on charge density distributions computed for nuclei with at least eight protons using relativistic continuum Hartree-Bogoliubov theory. The network learns to output Fourier-Bessel coefficients that describe these densities and reproduces the input data with root-mean-square deviations of 0.0123 fm for charge radii on the training set and 0.0198 fm on the validation set. A sympathetic reader would care because precise charge densities enter calculations of atomic energy levels, nuclear reaction rates, and astrophysical processes. If the network truly generalizes, it supplies fast, high-accuracy density profiles for many nuclei without repeating the full microscopic computation for each case.

Core claim

The central claim is that a deep neural network, trained on charge density distributions obtained from relativistic continuum Hartree-Bogoliubov theory for nuclei with proton number Z at least 8, reproduces those distributions with root-mean-square deviations of 0.0123 fm and 0.0198 fm for the charge radii on the training and validation sets respectively, thereby achieving higher precision than the original relativistic calculations.

What carries the argument

The deep neural network that maps nuclear structure features to the coefficients of a Fourier-Bessel series expansion of the nuclear charge density distribution.

If this is right

High-precision charge density data become available for atomic physics calculations without repeated microscopic runs.
Nuclear astrophysics models gain reliable input for reaction and decay processes involving many nuclei.
Rapid generation of density profiles for nuclei outside current computational reach becomes feasible.
The approach supplies a practical surrogate for systematic surveys of nuclear structure across the chart.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Retraining the same network on experimental charge density data, once sufficient measurements exist, could further reduce model dependence.
The method could be extended to predict related quantities such as neutron densities or electromagnetic form factors.
Machine-learning surrogates of this type may accelerate large-scale parameter scans in nuclear theory by replacing repeated full calculations.

Load-bearing premise

The relativistic continuum Hartree-Bogoliubov calculations supply sufficiently accurate reference data for both training the network and for claiming superior performance.

What would settle it

Comparison of the network's predicted charge radii or full density profiles against direct experimental measurements for nuclei with Z at least 8 that were held out from both training and validation.

Figures

Figures reproduced from arXiv: 2604.05312 by Haozhao Liang, Hui Hui Xie, Jian Li, Peng Xiang Du, Tian Shuai Shang, Yun Dong Wang.

**Figure 3.** Figure 3: Charge density distributions of 50Cr, and 70Zn. The results from the RCHB theory and the corresponding experimental data are also included for comparison. cision, with deviations from experimental values predominantly confined within the 0 — 0.01 fm range. In contrast, the RCHB predictions exhibit significantly broader dispersion in their deviations. RCHB has proven to be a successful DFT for investigat… view at source ↗

read the original abstract

A deep neural network (DNN) has been developed to accurately predict nuclear charge density distributions for nuclei with proton numbers $Z \geq 8$. By incorporating essential nuclear structure features, the model achieves a significant improvement in predictive accuracy over conventional methods. The charge density distributions are analyzed using a Fourier-Bessel (FB) series expansion, and the DNN is trained on a comprehensive dataset derived from relativistic continuum Hartree-Bogoliubov (RCHB) theory calculations. The model demonstrates exceptional performance, with root-mean-square deviations of 0.0123 fm and 0.0198 fm for charge radii on the training and validation sets, respectively, remarkably surpassing the precision of the original RCHB calculations. Beyond advancing nuclear physics research, this high-precision model provides critical data for applications in atomic physics, nuclear astrophysics, and related fields.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

A DNN trained on RCHB outputs reproduces those outputs with low error but supplies no external check that it improves on them.

read the letter

The core of this paper is a standard feed-forward network that learns to output the Fourier-Bessel coefficients of nuclear charge densities for Z ≥ 8, using RCHB calculations as the sole training target. The reported root-mean-square deviations on charge radii (0.0123 fm training, 0.0198 fm validation) show that the network fits its training distribution tightly, which is the main concrete result on offer. That fit is useful for anyone who already trusts RCHB and wants a faster surrogate for generating densities across many nuclei without rerunning the full calculation each time. The architecture choice and the focus on FB coefficients are straightforward and reproducible from the description given. Beyond that, the work does not introduce new nuclear physics or a new theoretical framework; it is an application of existing machine-learning tools to an existing dataset. The central claim that the model “remarkably surpasses the precision of the original RCHB calculations” is not backed by any comparison to experiment or to an independent theoretical benchmark. All performance numbers are residuals against the same RCHB labels used for training, so they measure how well the network copies RCHB rather than whether it corrects systematic errors in RCHB or matches reality better. Without that external anchor the surpassing statement cannot be evaluated. The paper is therefore best read as a practical interpolation tool inside one theoretical model rather than as an advance in predictive power. Readers who need quick charge-density estimates for atomic or astrophysical calculations may find the surrogate convenient once the code and trained weights are released. Readers looking for new insight into nuclear structure or for predictions that stand on their own will not get much from it. The manuscript is coherent on its own terms and the authors engage honestly with the literature on RCHB, so it is worth sending to referees who can ask for the missing external validation. I would not cite it in its current form.

Referee Report

2 major / 1 minor

Summary. The manuscript develops a deep neural network (DNN) trained on relativistic continuum Hartree-Bogoliubov (RCHB) calculations to predict charge density distributions for nuclei with Z ≥ 8. Charge densities are represented via Fourier-Bessel expansion; the model reports RMS deviations of 0.0123 fm (training) and 0.0198 fm (validation) for charge radii and claims to surpass the precision of the original RCHB data.

Significance. If the reported accuracy were shown to reflect genuine improvement over RCHB (via external benchmarks) rather than internal fitting, the DNN could serve as an efficient surrogate for generating charge densities across many nuclei, aiding applications in nuclear astrophysics and atomic physics. The approach demonstrates the feasibility of ML interpolation within a single theoretical framework.

major comments (2)

[Abstract] Abstract: the claim that the DNN 'remarkably surpassing the precision of the original RCHB calculations' is unsupported. The quoted RMS values (0.0123 fm training, 0.0198 fm validation) quantify residual error between DNN output and RCHB targets; no comparison to experimental charge radii or an independent theoretical benchmark is supplied to demonstrate that the DNN lies closer to truth than RCHB itself.
[Results] Validation procedure (results section): no information is given on the size, selection criteria, or nuclear species in the validation set, nor on how 'surpassing RCHB precision' was quantified. Without an external reference standard, it remains unclear whether the network generalizes or simply reproduces RCHB-specific features.

minor comments (1)

[Abstract] The abstract refers to 'essential nuclear structure features' incorporated into the model but does not enumerate them or indicate how they are encoded as inputs.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which help clarify the scope and presentation of our work. We agree that the abstract phrasing regarding surpassing RCHB precision is unsupported and will be revised. We will also expand the validation description in the results section to include the requested details on set composition and to remove any implication of improvement over the underlying RCHB theory. Our responses to the major comments follow.

read point-by-point responses

Referee: [Abstract] Abstract: the claim that the DNN 'remarkably surpassing the precision of the original RCHB calculations' is unsupported. The quoted RMS values (0.0123 fm training, 0.0198 fm validation) quantify residual error between DNN output and RCHB targets; no comparison to experimental charge radii or an independent theoretical benchmark is supplied to demonstrate that the DNN lies closer to truth than RCHB itself.

Authors: We acknowledge that the abstract claim is imprecise and unsupported by the evidence presented. The reported RMS values measure the fidelity with which the DNN reproduces the RCHB charge densities used as training targets; because the DNN is trained to approximate RCHB, it cannot exceed the accuracy of RCHB relative to experiment or other benchmarks. In the revised manuscript we will remove the phrase 'remarkably surpassing the precision of the original RCHB calculations' and replace it with a statement that the DNN serves as an accurate surrogate model, reproducing RCHB results to within 0.0123 fm (training) and 0.0198 fm (validation) for charge radii. If space allows, we will add a short contextual comparison of RCHB predictions to available experimental charge radii to illustrate the overall level of theory accuracy. revision: yes
Referee: [Results] Validation procedure (results section): no information is given on the size, selection criteria, or nuclear species in the validation set, nor on how 'surpassing RCHB precision' was quantified. Without an external reference standard, it remains unclear whether the network generalizes or simply reproduces RCHB-specific features.

Authors: We agree that the validation procedure requires additional documentation. In the revised results section we will specify the validation-set size (as a fraction of the total dataset), the selection method (e.g., random or stratified split), and the range of nuclear species included (Z and N intervals). We will also clarify that the quoted RMS values quantify reproduction error relative to RCHB rather than any improvement over RCHB, and we will remove the 'surpassing' language. The validation RMS will be presented simply as evidence that the network generalizes within the RCHB framework without overfitting to the training nuclei. revision: yes

Circularity Check

1 steps flagged

DNN 'predictions' and precision claims reduce to reproduction error on RCHB training targets

specific steps

fitted input called prediction [Abstract]
"The model demonstrates exceptional performance, with root-mean-square deviations of 0.0123 fm and 0.0198 fm for charge radii on the training and validation sets, respectively, remarkably surpassing the precision of the original RCHB calculations."

The quoted RMS values are the differences between DNN outputs and the RCHB-computed charge radii that constitute both the training targets and the validation labels. Because the network is fitted directly to these RCHB values, the reported deviations measure reproduction fidelity to the input data rather than an external demonstration that the DNN achieves higher precision than RCHB theory.

full rationale

The paper trains a DNN on charge densities and radii computed by RCHB theory, then reports RMS deviations on training and validation subsets of the same RCHB data as evidence that the model 'remarkably surpass[es] the precision of the original RCHB calculations.' This performance metric is the residual between DNN output and the RCHB labels themselves; the low values therefore quantify successful approximation of the inputs rather than an independent result that improves upon RCHB. The architecture itself is not self-referential, but the headline claim of superiority is forced by the choice of benchmark.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on RCHB theory supplying reliable training targets and on standard DNN training assumptions; no independent experimental validation or first-principles derivation is described.

free parameters (1)

DNN hyperparameters and architecture
Number of layers, neurons per layer, activation functions, and training schedule chosen to minimize error on RCHB data.

axioms (1)

domain assumption RCHB calculations provide sufficiently accurate charge densities to serve as training targets and as a precision benchmark
All reported performance metrics are defined relative to RCHB output.

pith-pipeline@v0.9.0 · 5463 in / 1418 out tokens · 64885 ms · 2026-05-10T19:41:07.234827+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

9 extracted references · 9 canonical work pages

[1]

7398 19932.0111 53 86 4.8272 23.3017 699.6550 24852.4184 51 58 4.5935 21.1004 585.8778 19619.5905 53 87 4.8417 23.4419 707.0386 25194.1866 51 59 4.6103 21.2547 596.5676 20363.1465 53 88 4.8579 23.5990 718.7698 25967.4055 51 60 4.6135 21.2842 595.1257 20021.2178 53 89 4.8669 23.6869 722.5134 26094.5559 51 61 4.6300 21.4368 606.4959 20865.5610 53 90 4.8860 ...

work page arXiv 2035
[2]

Charge radii, 2nd moment, 4th moment and 6th moment obtained by DNN. Z N R c(fm) R2(fm2) R4(fm4) R6(fm6) Z N R c(fm) R2(fm2) R4(fm4) R6(fm6) 55 57 4.7327 22.3986 667.3721 24485.5555 57 86 4.9177 24.1840 748.9968 27360.8942 55 58 4.7319 22.3912 663.4629 23994.0801 57 87 4.9266 24.2717 751.8425 27343.9148 55 59 4.7505 22.5670 677.4417 25061.5502 57 88 4.947...

work page arXiv 2093
[3]

Charge radii, 2nd moment, 4th moment and 6th moment obtained by DNN. Z N R c(fm) R2(fm2) R4(fm4) R6(fm6) Z N R c(fm) R2(fm2) R4(fm4) R6(fm6) 60 77 4.9111 24.1190 755.3367 28280.5810 63 69 4.9747 24.7480 822.1770 33727.1486 60 78 4.9166 24.1731 758.9091 28523.8909 63 70 4.9819 24.8195 823.1856 33518.9348 60 79 4.9108 24.1159 750.3034 27714.6937 63 71 4.978...

work page arXiv 1972
[4]

Charge radii, 2nd moment, 4th moment and 6th moment obtained by DNN. Z N R c(fm) R2(fm2) R4(fm4) R6(fm6) Z N R c(fm) R2(fm2) R4(fm4) R6(fm6) 65 101 5.1252 26.2674 893.7146 37337.9886 68 97 5.2187 27.2348 972.3014 42895.1895 65 102 5.1842 26.8757 948.3296 41998.5660 68 98 5.2553 27.6183 1010.1192 46301.3208 65 103 5.1332 26.3501 900.6257 38012.5966 68 99 5...

work page arXiv 1933
[5]

Charge radii, 2nd moment, 4th moment and 6th moment obtained by DNN. Z N R c(fm) R2(fm2) R4(fm4) R6(fm6) Z N R c(fm) R2(fm2) R4(fm4) R6(fm6) 71 96 5.2841 27.9219 1021.6153 45928.5930 74 93 5.3212 28.3153 1043.5119 46466.4436 71 97 5.3071 28.1654 1049.7670 48920.3938 74 94 5.2981 28.0699 1017.7883 43908.1437 71 98 5.3017 28.1080 1037.7748 47322.1521 74 95 ...

work page arXiv 2025
[6]

Charge radii, 2nd moment, 4th moment and 6th moment obtained by DNN. Z N R c(fm) R2(fm2) R4(fm4) R6(fm6) Z N R c(fm) R2(fm2) R4(fm4) R6(fm6) 76 123 5.4341 29.5295 1115.7361 50453.9622 79 107 5.4134 29.3046 1107.4789 50150.9281 76 124 5.4486 29.6868 1135.7338 52758.5805 79 108 5.4124 29.2936 1097.2533 48456.6810 76 125 5.4424 29.6194 1124.3005 51241.4286 7...

work page arXiv 1961
[7]

Charge radii, 2nd moment, 4th moment and 6th moment obtained by DNN. Z N R c(fm) R2(fm2) R4(fm4) R6(fm6) Z N R c(fm) R2(fm2) R4(fm4) R6(fm6) 81 133 5.5513 30.8165 1188.1464 51732.6615 84 123 5.5430 30.7247 1199.3100 54593.5789 81 134 5.5741 31.0709 1219.9071 55346.9903 84 124 5.5525 30.8306 1212.3231 56110.8460 81 135 5.5686 31.0098 1203.5459 52829.5937 8...

work page arXiv 1940
[8]

Charge radii, 2nd moment, 4th moment and 6th moment obtained by DNN. Z N R c(fm) R2(fm2) R4(fm4) R6(fm6) Z N R c(fm) R2(fm2) R4(fm4) R6(fm6) 87 121 5.5777 31.1108 1229.6886 56830.8520 90 134 5.7082 32.5841 1322.6057 61048.7437 87 122 5.5803 31.1396 1231.0119 56896.3541 90 135 5.7186 32.7023 1333.3684 61874.3387 87 123 5.5869 31.2131 1233.3065 56594.1047 9...

work page arXiv 2093
[9]

Charge radii, 2nd moment, 4th moment and 6th moment obtained by DNN. Z N R c(fm) R2(fm2) R4(fm4) R6(fm6) Z N R c(fm) R2(fm2) R4(fm4) R6(fm6) 94 134 5.7144 32.6549 1305.9068 57262.9409 98 158 5.9474 35.3714 1541.3770 76151.5896 94 135 5.7233 32.7561 1317.0960 58342.6432 99 140 5.7702 33.2951 1346.4879 58738.0470 94 136 5.7389 32.9349 1328.6674 58954.2361 9...

work page arXiv 1984