{"id":"83bb92b4-06f6-4596-aa8a-ab23652276a0","arxiv_id":"2506.20604","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"IKEBANA is a neural network that reproduces experimental K-shell electron-impact ionization cross sections from atomic number and overvoltage, from H to U.","lead":"Researchers trained a neural network that predicts how often an incoming electron knocks out an atom's innermost K-shell electron, using only atomic number and the electron's energy relative to the binding energy. The model, IKEBANA, matches available experimental data across the periodic table, but its predictions for the 27 elements with no measurements are unvalidated.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Random-split test set does not validate the claim of reliable predictions for unmeasured elements; a leave-one-element-out test is required.","rationale":"The paper's strongest advertised capability is prediction for elements with no measurements, and the load-bearing assumption is that the network interpolates smoothly in Z. The most serious problem is not the choice of activation function or data cleaning; it is that the reported test metrics cannot detect failure for unseen elements. A random split of points from all elements lets the network see every element in training, so R^2 and residual histograms in Table I and Fig. 4 are dominated by interpolation along U for known elements. Leave-one-element-out is the minimal experiment that would test the actual claim. The comparison with Bote et al. for Li, F, P, Re cannot substitute because that theory is known to be inaccurate near threshold for lighter elements (Fig. 5). The reader's conditional verdict already identifies this gap; my read supports the same verdict. I would add the leave-one-element-out test as an explicit condition.","tokens_in":8981,"tokens_out":4336,"duration_ms":48916,"concrete_test":"Perform leave-one-element-out cross-validation: for every element with at least 10 experimental data points, retrain the identical architecture and hyperparameters on all data from all other elements, then compute log-sigma_K residuals on the held-out element. Report per-element mean and standard deviation. If the held-out-element errors are comparable to the random-split test errors (log-sigma std near 0.5%), the cross-element generalization claim is supported; if low-Z or high-Z elements show systematic offsets larger than the experimental scatter, the unmeasured-element claim should be softened or supplemented with uncertainty estimates.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Section IV) is that ikebana 'delivers reliable predictions in energy regions where no measurements exist, and for a significant number of atomic targets for which no experimental data are currently available.' The Section II evaluation uses a random 20% split of individual data points, not an element-wise split. Because the same elements appear in both training and test sets, the reported R^2=0.997 measures interpolation within familiar elements at familiar overvoltages, not generalization to unseen elements. For the 27 elements without any data, e.g., Li, F, P, Re (Fig. 11), the predictions are pure interpolation in Z with no empirical check; the paper provides only a qualitative comparison with Bote et al. (6), a theory that itself deviates from data near threshold for H, He, C, and O (Fig. 5). No uncertainty bounds accompany the extrapolated values. The advertised reliability for unmeasured elements is therefore unsupported by the current test protocol.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper trains a fully connected neural network (IKEBANA) with two inputs, the atomic number Z and the overvoltage U = E_i / I_k, on the authors' 2024 compilation of experimental K-shell electron-impact ionization cross sections for elements from H to U. The target is log σ_K, and the model is trained on a random 20% test split, reporting R² = 0.997 on the test set. The authors claim that the model provides reliable predictions for the 27 elements with no experimental data and for energy regions without measurements, and they release the code as an open Google Colab notebook. Comparisons with the theoretical calculations of Bote et al. are shown for selected elements.","tokens_in":9139,"tokens_out":7096,"duration_ms":82900,"significance":"If the extrapolation claim holds, IKEBANA would be a practically valuable empirical estimator, filling gaps in a sparse experimental landscape and complementing theoretical calculations. The paper's strengths include a reproducible and openly available code, an up-to-date compilation of experimental data, a thoughtful discussion of architecture choices (e.g., tanh versus ReLU activation to ensure smoothness), and an honest acknowledgment of the scatter in the underlying data. However, the validation protocol does not support the core extrapolation claim: the reported metrics are based on a random split of individual measurements, so they quantify interpolation within well-measured elements, not generalization to unseen atomic numbers or energy gaps. The manuscript needs additional element-wise and energy-gap validation, or a substantial qualification of its conclusions, before the claim of 'reliable predictions' for unmeasured elements can be accepted.","major_comments":[{"comment":"The central claim of reliable predictions for unmeasured elements and energy regions is not supported by the random 20% test split described in Section II. In that split, the same elements appear in both training and test sets, so the reported R²=0.997 measures the network's ability to interpolate among measurements for known elements. The statements in Section III (Fig. 11) that adjacency in Z \"enhances the credibility\" of the predictions, and the Section IV conclusion that the model \"delivers reliable predictions ... for a significant number of atomic targets for which no experimental data are currently available,\" require a leave-one-element-out (or at least a Z-split) validation. Please perform such a cross-validation and report the prediction errors for held-out elements. Without this, the extrapolation claim is not demonstrated; the current test protocol only validates interpolation within measured elements.","section":"Section III and IV (evaluation protocol)"},{"comment":"The claim that IKEBANA gives reliable predictions \"in energy regions where no measurements exist\" (Section IV) is likewise untested. For elements with sparse data (e.g., molybdenum, xenon, barium, uranium in Figs. 8 and 10), the model's predictions across the cross-section peak rely on interpolation in Z and U, not on empirical constraints from those elements themselves. The random split does not create energy gaps, because every test point is surrounded by training points at nearby energies for familiar elements. A validation that masks a contiguous energy interval per element (for example, the peak region or the relativistic tail) and compares the model predictions with the excluded measurements would provide the necessary evidence. Until such a test is reported, the energy-extrapolation claim should be regarded as unsupported.","section":"Section III (energy extrapolation)"},{"comment":"The comparison with Bote et al. (6) is used to lend credibility to the unmeasured-element predictions, but it is only qualitative. Figure 11 shows curves for Li, F, P, and Re with no quantitative agreement metric. Moreover, Fig. 5 shows that Bote et al. deviates substantially from the experimental data near threshold for H, He, C, and O, so agreement with Bote et al. is not a reliable proxy for experimental accuracy in the near-threshold regime. The authors should provide a systematic quantitative comparison between IKEBANA and Bote et al. over all elements (e.g., median relative deviation in energy regions where experimental data exist), and explicitly state the known limitations of the Bote et al. theory where it fails to reproduce measured cross sections.","section":"Section III (comparison with theory)"},{"comment":"No uncertainty estimates accompany the predictions. For a model intended to supply cross sections for elements and energy ranges with no measured data, the absence of confidence intervals is a significant practical limitation. The authors could provide an ensemble spread (e.g., by varying the random seed or using Monte Carlo dropout) or bootstrap-based error bars. At a minimum, the term \"reliable predictions\" in Section IV should be qualified with the observed scatter of the training data and the lack of direct validation for extrapolated regions. Without such a qualification, readers may overinterpret the point predictions as exact values.","section":"Section IV (uncertainty quantification)"}],"minor_comments":[{"comment":"The hyperparameter selection process is described only qualitatively. Please specify how many configurations were tested and whether the test set was used at any point during model selection; this information is important for judging the validity of the reported test-set R².","section":"Section II"},{"comment":"The first row of Table I appears garbled (\"Mean value 5 3 9 5\"), and the units of the log-σ_K residuals are ambiguous (percent of log₁₀? of the cross section?). Define the residual definitions explicitly so that the reported standard deviations are interpretable.","section":"Table I"},{"comment":"The caption says \"Curves as in Figure 5,\" but Figure 11 shows no experimental circles; please state explicitly that no experimental data exist for Li, F, P, and Re.","section":"Figure 11"},{"comment":"The statement \"no data cleansing was implemented\" is somewhat at odds with the outlier analysis in Table I. Clarify that the 1% trimming is applied only to the statistical summary, not to the training or test sets.","section":"Section II"},{"comment":"The code is distributed through Google Colab only; consider archiving a stable version in a permanent repository (e.g., Zenodo) to ensure long-term access and reproducibility.","section":"Data Availability"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal and the code and data availability are commendable. The key technical problem is that the advertised generalization to unmeasured elements and energy regions is not validated by the current random-split test. A leave-one-element-out analysis (or at least a Z-split) is essential; if that analysis shows moderate-to-good performance, the paper could be acceptable after revision. If the element-wise errors are large, the conclusions must be softened to claims of smooth interpolation within the measured Z range. No citation or novelty concerns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The useful part of this paper is real: they take a straightforward fully connected network with two inputs, Z and log(U), train it on a few thousand experimental K-shell ionization points, and get tight residuals on a held-out 20% of the data. The code is open, the data are open, and the write-up is honest about the messiness of the experimental compilation. They also did sensible architecture exploration, showing why ReLU gives jagged curves and why tanh fixes it. That part deserves credit.\n\nThe soft spot is exactly where the stress-test note points. The test set is a random split of individual measurements, so the same elements appear in both training and testing. The R^2 = 0.997 on log sigma_K measures how well the network interpolates in energy for familiar elements. It does almost nothing to support the claim that IKEBANA delivers reliable predictions for the 27 elements with no measurements at all. Those predictions are pure interpolation in Z, with no independent check. The comparison with Bote et al. in Figure 11 is qualitative, and Bote itself deviates from data for the light measured elements near threshold, so that comparison is not a strong anchor either. There are no uncertainty bounds on the extrapolated values.\n\nThat said, I would not call the paper flawed at its core. For elements with data, the model is a good empirical fit, and for unmeasured elements the curves are plausible. The problem is the advertised reliability. The authors could fix this fairly easily: run a leave-one-element-out cross-validation, or at least a split by atomic number, and show that predictions for held-out elements land within the scatter of nearby measured elements. They could also add a simple uncertainty estimate, even a crude heteroscedastic one from the residual distribution. Without those, the right language is 'moderate-confidence estimates' not 'reliable predictions.'\n\nI would send this to peer review. It is a useful tool for applied atomic physics, the repo is functional, and the methodological gap is fixable rather than fatal. The referee should push on the extrapolation claim and ask for element-wise validation, but the paper is worth the referee time.\n\nBring it to reading group if you want to discuss how far a random-split test can be stretched toward extrapolation claims; it is a clean teaching example.","headline":"A practical NN surrogate for K-shell ionization that measures well on elements with data, but the extrapolation to unmeasured elements is asserted rather than demonstrated.","tokens_in":9733,"tokens_out":1567,"would_cite":false,"duration_ms":21172,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["34.80.Dp"],"model":"deepseek-v4-flash","headline":"A fully connected neural network taking only atomic number and overvoltage as inputs reproduces experimental K-shell ionization cross sections from hydrogen to uranium and, the paper argues, extends reliably to elements and energies no…","keywords":["K-shell ionization","electron-impact ionization","cross sections","neural network","machine learning","overvoltage scaling","atomic number","ikebana"],"falsifier":"Measure the K-shell ionization cross section of one of the showcased unmeasured elements, lithium, fluorine, phosphorus, or rhenium, across the threshold-to-peak energy range; if the experimental curve disagrees with the ikebana prediction by much more than the roughly 20 percent scatter that characterizes the existing database, the universal-prediction claim for unmeasured targets fails.","tokens_in":8783,"feed_emoji":"⚛️","tokens_out":7047,"duration_ms":75867,"temperature":0.7,"pith_summary":"This paper attempts to establish that a small fully connected neural network can serve as a universal empirical estimator of K-shell ionization cross sections for electron impact. Trained on a compilation of about 2,500 experimental points covering hydrogen through uranium and energies from threshold to the GeV range, the network uses just two inputs, the atomic number $Z$ and the overvoltage $U = E_i/I_K$, and reaches $R^2 = 0.997$ on a held-out test set. The paper argues the model also produces continuous, physically smooth predictions for the 27 elements with no measurements at all and for energy regions no experiment has reached. If correct, this gives applied fields a fast, parameter-free route to cross sections that previously had to come from full theoretical calculations or from extrapolation of sparse data.","feed_headline":"Two inputs, whole periodic table: K-shell ionization predicted","feed_subtitle":"From H to U, one neural network reproduces measured cross sections and reaches 27 elements with no data.","key_machinery":"The central object is IKEBANA, a fully connected feedforward neural network with two inputs $Z$ and $\\log U$, three tanh hidden layers of widths 32, 16, and 32, L2 regularization ($\\lambda = 10^{-5}$), batch normalization, and a dropout rate of $10^{-4}$, ending in a linear output that regresses $\\log \\sigma_K$. The logarithmic scaling of cross section and overvoltage compresses the many orders of magnitude spanned by the data, while the atomic number stays linear. The tanh activation is the load-bearing design choice: it makes the predicted energy dependence smooth and continuous, whereas the ReLU version produced jagged, unphysical curves that had to be artificially smoothed.","core_discovery":"The central claim is that K-shell electron-impact ionization cross sections, expressed as a function of atomic number and scaled incident energy, form a smooth pattern that a 2–32–16–32–1 network with tanh activations can learn directly from raw experimental data. The trained model reproduces the experimental compilation with $R^2 = 0.997$ on the held-out test set and with log-scale residuals whose standard deviation is about 0.5 percent after the extreme 1 percent tails are removed. The paper further claims that the model agrees with measurements at least as well as the distorted-wave and plane-wave Born approximations near thresholds and peak regions, and it exhibits predictions for lithium, fluorine, phosphorus, and rhenium as examples of targets for which no experimental data exist. The smoothness of tanh, rather than any external smoothing, is what gives the predicted energy curves their physical shape.","pith_inferences":["A natural stress test would be to hold out an entire well-measured element, such as silver, retrain on all other elements, and compare the predicted curve with the full experimental set; that would measure genuine cross-element generalization rather than interpolation within a dense cloud of points.","Because the network uses only the overvoltage scaling and atomic number, the same training recipe could plausibly be applied to L-shell or M-shell ionization, or to positron impact, whenever a comparable experimental compilation becomes available; the paper does not claim this extension.","Predictions for unmeasured elements and extreme energies are best read as interpolation hypotheses; an extension that would strengthen that reading is a calibrated predictive uncertainty for every element and energy the model is asked to cover."],"forward_implications":["Reliable cross-section estimates become available immediately for the 27 elements with no K-shell ionization measurements and for energy intervals where no experiment exists, at least to the extent that interpolation in $Z$ and $U$ is physical.","Applications such as X-ray emission analysis, plasma modeling, and radiation transport can query the ikebana code with any atomic number and energy window and obtain continuous cross-section curves without running a full theoretical calculation.","The residual statistics give an empirical target for future theory: a calculation that wants to improve on current practice should beat the roughly 20 percent scatter that characterizes the experimental database itself once extreme outliers are removed.","The model supplies an independent, data-driven benchmark for identifying discrepant measurements, as in the cases of chlorine and titanium where reported points deviate systematically from both the network and the theoretical curves."],"supporting_citations":[{"why":"Supplies the updated experimental database of over 2,500 measured K-shell ionization cross sections that form the training data and all experimental comparisons.","marker":"[1]"},{"why":"Provides the neural-network architecture, scaling, and training procedure on which the ikebana model is built.","marker":"[4]"},{"why":"Provides the theoretical DWBA/PWBA cross-section curves used throughout as the main independent benchmark for the network's predictions.","marker":"[6]"}],"fun_headline_variants":["Neural network predicts K-shell ionization from H to U","Two inputs, full periodic table: IKEBANA maps K-shell ionization","IKEBANA net matches experiments and predicts missing elements","Smooth tanh network gives K-shell cross sections without data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that K-shell ionization cross sections vary smoothly with atomic number and with incident energy scaled by the K-shell ionization energy, so predictions for the 27 elements with no measurements and for energy ranges no experiment has covered inherit their accuracy from neighboring measured data.","fun_headline_variants_meta":{"raw":{"variants":["Neural network predicts K-shell ionization from H to U","Two inputs, full periodic table: IKEBANA maps K-shell ionization","IKEBANA net matches experiments and predicts missing elements","Smooth tanh network gives K-shell cross sections without data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000243,"raw_usage":{"total_tokens":1466,"prompt_tokens":819,"completion_tokens":647,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":435,"completion_tokens_details":{"reasoning_tokens":575}},"tokens_in":435,"tokens_out":647,"duration_ms":6858,"temperature":1.0,"reasoning_tokens":575,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:44:30.240015+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the K-shell ionization cross section of one of the showcased unmeasured elements, lithium, fluorine, phosphorus, or rhenium, across the threshold-to-peak energy range; if the experimental curve disagrees with the ikebana prediction by much more than the roughly 20 percent scatter that characterizes the existing database, the universal-prediction claim for unmeasured targets fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the updated experimental database of over 2,500 measured K-shell ionization cross sections that form the training data and all experimental comparisons."},{"cited_title":"Bivort Haiek, A","cited_arxiv_id":null,"evidence_quote":"Provides the neural-network architecture, scaling, and training procedure on which the ikebana model is built."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the theoretical DWBA/PWBA cross-section curves used throughout as the main independent benchmark for the network's predictions."}],"review_version":1}