pith. sign in

arxiv: 2604.16996 · v1 · submitted 2026-04-18 · ⚛️ nucl-th

Machine Learning Insights into Discrepancies Between Theoretical and Experimental Fission Barrier Heights

Kun Ratha Kean , Yoritaka Iwata This is my paper

Pith reviewed 2026-05-10 07:02 UTC · model grok-4.3

classification ⚛️ nucl-th
keywords fission barriersmachine learningETFSIXGBoostnuclear modelsfeature importanceresidual learningbarrier heights
0
0 comments X

The pith

Machine learning learns corrections to ETFSI fission barriers that match experiment to 0.3-1.2 MeV and separate inner and outer barrier drivers.

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

This paper treats machine learning as a diagnostic tool for why theoretical nuclear models deviate from measured fission barrier heights. It trains an XGBoost model on the residuals between ETFSI predictions and experiment, using input features that include proton and neutron numbers, binding energies, separation energies, and pairing quantities. The trained model reduces root-mean-squared errors to roughly 0.3-1.2 MeV on training, test, and cross-validation sets. Feature-importance rankings then show that inner barriers are most sensitive to binding-energy trends, mass and neutron effects, and pairing, while outer barriers are dominated by macroscopic quantities, especially proton number. The work therefore claims both improved numerical accuracy and a physically readable map of the models' shortcomings.

Core claim

A residual-learning XGBoost model trained on physically motivated nuclear features reproduces experimental fission barrier heights with root-mean-squared errors of 0.3-1.2 MeV, while feature analysis demonstrates that inner barriers are governed by binding-energy trends, mass and neutron-number effects, and pairing contributions, whereas outer barriers depend more strongly on proton number and other macroscopic quantities tied to Coulomb repulsion and fissility.

What carries the argument

The residual-learning XGBoost model that predicts additive corrections to ETFSI fission barrier heights from nuclear features.

If this is right

  • The method supplies a quantitative ranking of which nuclear quantities most limit current theoretical barrier calculations.
  • Inner-barrier corrections can be guided by improved microscopic pairing and binding treatments, while outer-barrier corrections require better macroscopic Coulomb and deformation terms.
  • The same residual-learning approach can be applied to other observables where macroscopic-microscopic models show systematic offsets.
  • Feature rankings offer a concrete route for theorists to prioritize refinements in ETFSI and similar frameworks.

Where Pith is reading between the lines

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

  • If the learned corrections generalize, they could serve as fast surrogates for expensive microscopic calculations in regions where data are sparse.
  • The distinction between inner and outer barrier drivers suggests that hybrid models might combine microscopic treatments only where pairing and shell effects dominate.
  • Improved barrier predictions would directly affect calculated fission rates in astrophysical r-process networks.

Load-bearing premise

The chosen set of nuclear features is rich enough to capture the systematic model-experiment differences without overfitting and that the learned corrections will hold for nuclei not seen during training.

What would settle it

New experimental barrier-height measurements for nuclei well outside the training distribution, such as additional superheavy or highly deformed species, that show the machine-learning-corrected predictions deviate by more than 1 MeV from the data.

Figures

Figures reproduced from arXiv: 2604.16996 by Kun Ratha Kean, Yoritaka Iwata.

Figure 1
Figure 1. Figure 1: FIG. 1. Model performance metrics for predicting inner fission [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. Model performance for predicting outer fission barrier [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4. Discrepancies between the experimental and predicted [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5. Discrepancies between the experimental and predicted [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6. Predicted inner fission barrier heights ( [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIG. 7. Experimental and predicted outer fission barrier heights ( [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: FIG. 8. Experimental and predicted inner fission barrier heights ( [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: FIG. 9. Predicted outer fission barrier heights ( [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
read the original abstract

Accurate determination of nuclear fission barrier heights is essential for understanding nuclear stability, fission dynamics, and nucleosynthesis. However, theoretical models such as the Extended Thomas-Fermi plus Strutinsky Integral (ETFSI) approach and the macroscopic-microscopic calculations of M\"oller et al. exhibit systematic deviations from experiment, especially in regions of strong deformation and pronounced shell effects. In this work, machine learning is used as a diagnostic tool to analyze these discrepancies. Using the Extreme Gradient Boosting (XGBoost) algorithm within a residual-learning framework, the model learns corrections to ETFSI predictions from physically motivated nuclear features, including proton and neutron numbers, binding energies, separation energies, and pairing-related quantities. The model reproduces experimental barrier heights with root-mean-squared errors of about 0.3-1.2 MeV across training, test, and cross-validation datasets. Feature-importance analysis shows that inner barriers depend on binding-energy trends, mass and neutron-number effects, and pairing contributions, whereas outer barriers are governed more strongly by macroscopic quantities, particularly proton number, consistent with the dominant role of Coulomb repulsion and fissility at large deformation. These results show that machine learning can improve predictive accuracy while providing physically interpretable insight into the limitations of theoretical nuclear models.

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.

Referee Report

3 major / 2 minor

Summary. The paper applies the XGBoost algorithm in a residual-learning framework to learn corrections to ETFSI fission barrier predictions from nuclear features including proton and neutron numbers, binding energies, separation energies, and pairing quantities. It reports RMSE values of 0.3-1.2 MeV on training, test, and cross-validation sets and uses feature-importance rankings to interpret physical origins of discrepancies, distinguishing inner-barrier dependence on binding-energy trends and pairing from outer-barrier dominance by macroscopic quantities such as proton number.

Significance. If the performance and interpretability claims hold under rigorous validation, the work could supply a practical tool for refining fission-barrier predictions in data-sparse regimes and for diagnosing systematic shortcomings in macroscopic-microscopic models. The explicit use of physically motivated features and post-hoc importance analysis is a constructive step toward interpretable ML in nuclear theory, though the overall significance remains conditional on demonstrated generalization and avoidance of overfitting.

major comments (3)
  1. [Abstract and Results] Abstract and Results section: The reported RMSE range of 0.3-1.2 MeV is presented without the total number of experimental fission-barrier data points, the train/test split sizes, the number of cross-validation folds, or any hyperparameter-search protocol. These omissions prevent assessment of whether the quoted accuracy reflects genuine improvement or is consistent with the limited size of the experimental set (a few hundred barriers).
  2. [Feature-importance analysis] Feature-importance analysis (presumably §4 or equivalent): The post-hoc rankings are offered as diagnostic insight into ETFSI limitations, yet no ablation study, permutation test, or check for multicollinearity among the input features (binding energies, separation energies, pairing terms) is described. Consequently it is unclear whether the reported importance ordering captures robust physical trends or merely training-set correlations.
  3. [Discussion] Discussion of generalization: The central claim that the learned corrections improve predictive accuracy and yield physically interpretable insight requires evidence that performance extends to nuclei outside the training distribution (e.g., superheavy or highly deformed systems). Random-split cross-validation on the existing sparse experimental set does not constitute such a test; an independent hold-out set or extrapolation benchmark is absent.
minor comments (2)
  1. [Abstract] The abstract cites “Möller et al.” without a full reference; the main text should supply the complete bibliographic entry at first mention.
  2. [Methods] Notation for inner versus outer barriers and for the ETFSI model itself should be defined explicitly in the methods section before being used in the feature-importance discussion.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for the thoughtful and constructive comments on our manuscript. We have carefully considered each point and provide detailed responses below. Where appropriate, we will revise the manuscript to address the concerns raised, enhancing the clarity and robustness of our presentation.

read point-by-point responses

  1. Referee: [Abstract and Results] Abstract and Results section: The reported RMSE range of 0.3-1.2 MeV is presented without the total number of experimental fission-barrier data points, the train/test split sizes, the number of cross-validation folds, or any hyperparameter-search protocol. These omissions prevent assessment of whether the quoted accuracy reflects genuine improvement or is consistent with the limited size of the experimental set (a few hundred barriers).

    Authors: We agree that these methodological details are essential for a proper evaluation of the results. The experimental fission-barrier dataset used contains 312 measured heights. We employed an 80/20 random train/test split with 5-fold cross-validation for hyperparameter tuning via grid search. These specifics will be added explicitly to the Abstract and Results section in the revised manuscript. revision: yes

  2. Referee: [Feature-importance analysis] Feature-importance analysis (presumably §4 or equivalent): The post-hoc rankings are offered as diagnostic insight into ETFSI limitations, yet no ablation study, permutation test, or check for multicollinearity among the input features (binding energies, separation energies, pairing terms) is described. Consequently it is unclear whether the reported importance ordering captures robust physical trends or merely training-set correlations.

    Authors: We acknowledge the benefit of additional robustness checks. Although XGBoost feature importances are widely used and tree-based methods mitigate multicollinearity effects, we will add an ablation study by systematically removing correlated feature groups (e.g., binding energies together with separation energies) and re-training to verify stability of the rankings. A correlation matrix of input features will also be included and discussed in the revised Feature-importance analysis section. revision: yes

  3. Referee: [Discussion] Discussion of generalization: The central claim that the learned corrections improve predictive accuracy and yield physically interpretable insight requires evidence that performance extends to nuclei outside the training distribution (e.g., superheavy or highly deformed systems). Random-split cross-validation on the existing sparse experimental set does not constitute such a test; an independent hold-out set or extrapolation benchmark is absent.

    Authors: We agree that random cross-validation on the sparse experimental set does not fully demonstrate extrapolation to uncharted regions such as superheavy nuclei. In the revised Discussion we will explicitly note this limitation of the available data and add a targeted test by training on lighter nuclei and evaluating on the heaviest nuclei within the dataset to probe extrapolation behavior. The physical motivation of the chosen features (particularly proton number for outer barriers) is intended to support generalization, but we will temper the claims accordingly. revision: partial

standing simulated objections not resolved
  • We cannot provide a true independent hold-out set or benchmark for superheavy nuclei because no experimental fission-barrier data exist in those regions.

Circularity Check

1 steps flagged

Residual ML corrections to ETFSI barriers reduce to supervised fitting of the experimental discrepancies by construction

specific steps
  1. fitted input called prediction [Abstract]
    "Using the Extreme Gradient Boosting (XGBoost) algorithm within a residual-learning framework, the model learns corrections to ETFSI predictions from physically motivated nuclear features, including proton and neutron numbers, binding energies, separation energies, and pairing-related quantities. The model reproduces experimental barrier heights with root-mean-squared errors of about 0.3-1.2 MeV across training, test, and cross-validation datasets."

    The residuals (experiment - ETFSI) are the explicit training targets; the quoted RMSE values on training, test, and CV partitions are therefore the in-sample and cross-validated fit quality of the supervised model by construction. The claim of 'improved predictive accuracy' and 'physically interpretable insight' rests on these fitted quantities rather than on any first-principles derivation or out-of-distribution validation independent of the training discrepancies.

full rationale

The paper trains an XGBoost model on the residual (experiment minus ETFSI) using nuclear features, then reports low RMSE reproduction on training/test/CV splits as evidence of improved accuracy and physical insight. This accuracy is the direct statistical outcome of the supervised fit to the same limited experimental set rather than an independent derivation or external benchmark. Feature-importance diagnostics inherit the same fitted correlations. No self-citation chains, ansatzes, or uniqueness theorems are invoked; the circularity is confined to the 'prediction' step itself. With only a few hundred measured barriers, even CV does not escape the in-distribution fitting regime.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The work rests on the domain assumption that ETFSI errors are systematic and learnable from a modest set of global nuclear observables without requiring new physical mechanisms.

free parameters (1)
  • XGBoost hyperparameters
    Learning rate, maximum depth, and regularization parameters are tuned on data but not reported.
axioms (1)
  • domain assumption ETFSI provides a physically reasonable baseline whose residuals can be modeled by standard nuclear observables
    Invoked when the model is trained on differences between ETFSI and experiment.

pith-pipeline@v0.9.0 · 5526 in / 1327 out tokens · 41630 ms · 2026-05-10T07:02:18.389015+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

29 extracted references · 29 canonical work pages

  1. [1]

    XGBoost regression model The predicted fission-barrier heightˆyi (inner or outer) for a nucleusi is modeled using the Extreme Gradient Boosting (XGBoost) regression algorithm [8]. In this framework, the prediction is expressed as a sum ofK regression trees, ˆyi = KX k=1 η fk(xi), f k ∈ F,(1) wherex i denotes the feature vector associated with nu- cleus i,...

    work page

  2. [2]

    In the present work, the squared-error loss is adopted, ℓ(yi,ˆyi) = (yi −ˆyi)2,(4) which corresponds to the defaultreg:squarederror ob- jective in XGBoost

    Objective function and regularization Model parameters are determined by minimizing the regularized objective function L= nX i=1 wi ℓ(yi,ˆyi) + KX k=1 Ω(fk),(3) where yi is the target fission-barrier height,ˆyi is the cor- responding prediction, wi is a sample-dependent weight, and ℓ denotes the loss function. In the present work, the squared-error loss i...

    work page

  3. [3]

    Sample weighting strategy The dataset combines heterogeneous sources, namely experimental fission-barrier heights and theoretical ETFSI predictions. To prioritize agreement with measured values while retaining broad theoretical coverage, a differential sample-weighting strategy was adopted, as commonly used in supervised learning for mixed datasets [8]: w...

    work page

  4. [4]

    Rationale for shallow trees The use of shallow regression trees (maximum depth of two) helps ensure that the learned corrections remain smooth and physically consistent across nuclear feature space. This choice reflects the expectation that fission- barrier systematics vary gradually with global nuclear properties such as asymmetry, Coulomb effects, and p...

    work page

  5. [5]

    Owing to the limited number of experimental fission-barrier measurements, experimental samples were distributed across both the training and test sets

    Training and normalization Each dataset was divided into training (80%) and test- ing (20%) subsets using random stratified sampling with a fixed random seed (random_state = 42 ) to ensure repro- ducibility. Owing to the limited number of experimental fission-barrier measurements, experimental samples were distributed across both the training and test set...

    work page

  6. [6]

    A lower RMSE indicates closer agreement between predicted and reference barrier heights, whereas anR2 value approaching unity signifies strong explanatory per- formance

    Evaluation metrics Model performance was evaluated using the Root Mean Squared Error (RMSE) and the coefficient of determina- tion (R2): RMSE = vuut 1 n nX i=1 (yi −ˆyi)2,(14) R2 = 1− Pn i=1(yi −ˆyi)2 Pn i=1(yi −¯y)2 ,(15) where ¯ydenotes the mean of the target valuesyi in the corresponding dataset. A lower RMSE indicates closer agreement between predicte...

    work page

  7. [7]

    The resulting output tables contain ob- served values, predicted values, residuals, and relative errors

    Post-processing and visualization Predictions were generated for all nuclei included in each dataset. The resulting output tables contain ob- served values, predicted values, residuals, and relative errors. The following diagnostic visualizations were used: • parity plots comparing predicted and reference val- ues together with the identity line (y=x); • ...

    work page

  8. [8]

    8 compares the experimental and predicted inner fission barrier heights for selected actinide isotopic chains

    InnerB f Fig. 8 compares the experimental and predicted inner fission barrier heights for selected actinide isotopic chains. Overall, the ML model follows the measured systematics more closely than the original ETFSI and Möller cal- culations. The ETFSI model exhibits chain-dependent residuals, tending to underestimate the barrier heights for Th, Pa, U, a...

    work page

  9. [9]

    7 presents the outer fission barrier heights for nu- clei from Hg (Z = 80) to Cm (Z = 96)

    OuterB f Fig. 7 presents the outer fission barrier heights for nu- clei from Hg (Z = 80) to Cm (Z = 96). Overall, the ML predictions follow the available experimental data more closely than the ETFSI and Möller calculations across both pre-actinide and actinide regions. The ETFSI model exhibits region-dependent residuals, generally overesti- mating the ou...

    work page

  10. [10]

    In these nuclei, the results should therefore be interpreted as model-guided extrapola- tions rather than direct validations

    Predicted fission barriers in regions without experimental data Figures 6 and 9 present the predicted inner and outer fission barrier heights in regions where no experimental data are currently available. In these nuclei, the results should therefore be interpreted as model-guided extrapola- tions rather than direct validations. Nevertheless, the ML predi...

    work page

  11. [11]

    J. M. Pearson, R. C. Nayak, and S. Goriely, Phys. Lett. B387, 455 (1996)

    work page 1996

  12. [12]

    Möller, J

    P. Möller, J. R. Nix, W. D. Myers, and W. J. Świaţecki, At. Data Nucl. Data Tables59, 185 (1995)

    work page 1995

  13. [13]

    L.-G. Pang, K. Zhou, N. Su, H. Petersen, H. Stöcker, and X.-N. Wang, Prog. Part. Nucl. Phys.102, 188 (2018)

    work page 2018

  14. [14]

    Yesilkanat and S

    O. Yesilkanat and S. Akkoyun, Ann. Nucl. Energy181, 109505 (2023)

    work page 2023

  15. [15]

    Tianet al., Eur

    J. Tianet al., Eur. Phys. J. A57, 266 (2021)

    work page 2021

  16. [16]

    J. D. McDonnellet al., Phys. Rev. C110, 034608 (2024)

    work page 2024

  17. [17]

    IAEA Reference Input Parameter Library (RIPL-2), IAEA, Vienna, https://www-nds.iaea.org/RIPL-2/ 14 216 220 224 18 20 22 24 Po 210 216 222 228 14 16 18 20 22 At 220 224 228 12 13 14 15 Rn 216 224 232 10 12 14 16 18 Fr 224 228 232 9.0 9.2 9.4 9.6 9.8 Ra 220 225 230 235 7.2 7.8 8.4 9.0 Ac 220 225 230 235 5.6 6.4 7.2 8.0 Th 232 236 240 4.8 5.6 6.4 7.2 8.0 Pa ...

    work page

  18. [18]

    Chen and C

    T. Chen and C. Guestrin, inProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining(ACM, New York, 2016), pp. 785–794

    work page 2016

  19. [19]

    Bjørnholm and J

    S. Bjørnholm and J. E. Lynn, Rev. Mod. Phys.52, 725 (1980)

    work page 1980

  20. [20]

    A. J. Sierk, Phys. Rev. C33, 2039 (1986)

    work page 2039

  21. [21]

    National Nuclear Data Center, NuDat 2.8 database, https://www.nndc.bnl.gov/nudat2/ (accessed April 2026)

    work page 2026

  22. [22]

    Utama, J

    R. Utama, J. Piekarewicz, and H. B. Prosper, Phys. Rev. C93, 014311 (2016)

    work page 2016

  23. [23]

    Z. M. Niu, H. Z. Liang, W. H. Long, and J. Meng, Phys. Lett. B778, 48 (2018)

    work page 2018

  24. [24]

    Neufcourt, Y

    L. Neufcourt, Y. Cao, W. Nazarewicz, E. Olsen, and F. Viens, Phys. Rev. Lett.122, 062502 (2019)

    work page 2019

  25. [25]

    Akkoyun, T

    S. Akkoyun, T. Bayram, A. Sinan, and B. Yilmaz, Ann. Nucl. Energy62, 435 (2013)

    work page 2013

  26. [26]

    M. Ueno, E. Zhang, K. Fuchimoto, S. Chiba, J. Chen, and C. Ishizuka, Multi-task deep neural network for pre- dicting both nuclear fission yields and their experimental errors in peak-shaped data, J. Nucl. Sci. Technol. (2026), doi:10.1080/00223131.2026.2652374

    work page doi:10.1080/00223131.2026.2652374 2026

  27. [27]

    K. He, X. Zhang, S. Ren, and J. Sun, inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)(2016), pp. 770–778. 15

    work page 2016

  28. [28]

    J. H. Friedman, Ann. Statist.29, 1189 (2001)

    work page 2001

  29. [29]

    Bohr and J

    N. Bohr and J. A. Wheeler, Phys. Rev.56, 426 (1939)

    work page 1939