{"id":"b7473d6b-8976-400b-832c-fff54fa4d669","arxiv_id":"2501.02932","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A composition-only model that applies ReLU to a weighted average of per-element learned parameters predicts band gaps with 0.575 eV MAE and yields chemically interpretable element weights.","lead":"A simple machine learning model predicts a crystal's band gap from its chemical formula alone by averaging one learned number per element and clipping the result at zero. It reaches a mean absolute error of 0.575 eV on a standard dataset, offering a fast, interpretable alternative to structure-based and black-box models.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The ReLU model's MAE advantage over the linear baseline may stem simply from clamping negative predictions to zero, a trivial non-negativity effect, so the paper's central claim that the ReLU form captures the band gap's atypical statistics is not yet established.","rationale":"The reader's verdict is already CONDITIONAL, citing missing baselines and fitting details. My concern sharpens this: the headline numerical comparison (ReLU vs linear) is not a fair test of the paper's modeling thesis, because any clamping of the linear predictor improves MAE on zero-inflated non-negative data. The paper even states the ReLU was designed to predict a substantial fraction of zeros; thus the comparison with the linear model is essentially comparing a model with a non-negativity constraint to one without. The missing clamped-OLS baseline and standard zero-inflated baselines make the empirical support for the 'mixed random variable' motivation weak. This does not overturn the paper; the model may still be a useful heuristic, and the paper limits its claim to 2-4 element materials. But the central contribution—that the ReLU form is 'suited towards the atypical statistics'—would need the additional baselines to be convincing. Hence the verdict stays CONDITIONAL (unchanged), with the condition being these baseline comparisons. The reader's weakest assumption focused on structure-dependence; my concern is adjacent but distinct, focusing on the baseline against which the model's success is measured, so agreement is partial.","tokens_in":13198,"tokens_out":6273,"duration_ms":140739,"concrete_test":"Run the same 10-fold cross-validation and compute the test MAE of (i) the post-hoc clamped OLS linear model, max(0, w_OLS·f(M)), and (ii) a two-stage model: logistic regression on element fractions to predict zero vs positive gap, followed by a linear (or ridge) regression on element fractions for the positive-gap subset. If either baseline achieves a test MAE within 0.05 eV of the ReLU model's 0.575 eV, the ReLU model's advantage is not specific to its mixed-variable design. Additionally, refit the ReLU model to minimize MAE directly; if that materially changes the weights or MAE, the reported results are sensitive to the loss mismatch.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim (Table I) is that the ReLU model (Eq. 8) outperforms the linear model (Eq. 9) by 0.249 eV MAE, attributed to the ReLU transform being suited to the band gap's mixed distribution. But because ~53% of labels are exactly 0 eV and the target is non-negative, any model that maps negative predictions to zero will improve MAE relative to an unconstrained linear model. The paper does not report the MAE of the simplest such baseline: the positive part of the OLS linear predictor, max(0, w_OLS · f(M)). It also does not compare with standard zero-inflated approaches (e.g., Tweedie regression, a two-stage classifier plus regressor, or a linear model with a non-negativity constraint on predictions). Without these baselines, the observed improvement does not specifically validate the 'mixed random variable' design; it may be a trivial consequence of enforcing non-negativity. The eCDF jump at 0 eV (Fig. 3) is also partly by construction, since the ReLU outputs exactly 0 for the halfspace of compositions with w·f <= 0. Thus the load-bearing claim that the model form captures the atypical statistics of the band gap is under-supported, and the proposed chemical interpretability of the weights is a heuristic that could be sensitive to this misspecification.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a composition-only model for the electronic band gap, epsilon_relu(M) = ReLU(sum_E w_E f_E(M)), with one learned weight per element. After documenting that the band gap distribution has a mass at 0 eV, the authors argue that a ReLU transform is suited to this mixed random variable. They fit the weights on a processed version of the Zhuo et al. dataset and report a 10-fold cross-validated MAE of 0.575 +/- 0.036 eV versus 0.824 +/- 0.035 eV for an ordinary least squares linear model. They also plot the empirical CDFs of predictions and present the learned weights on a periodic table as heuristic chemical interpretability.","tokens_in":13487,"tokens_out":5988,"duration_ms":59764,"significance":"If the central claim is established, the model is an attractive interpretable baseline and the mixed-random-variable framing is a useful corrective to treating the band gap as an ordinary continuous target. The paper's strengths are its simplicity, the use of cross-validation, the explicit empirical-CDF analysis, and the fact that the model has only one parameter per element. However, the specific claim that the ReLU form is responsible for capturing the atypical statistics is not yet supported, because the paper does not test simpler non-negativity baselines or compare with existing composition-only models. The reported MAE values are internally consistent with the described training procedure, but the load-bearing comparison is incomplete.","major_comments":[{"comment":"The reported difference between Eq. (8) and Eq. (9) does not establish that the ReLU nonlinearity, rather than mere non-negativity, drives the improvement. Since 53.2% of the labels are exactly 0 eV and the ReLU clamps every negative pre-activation to zero, the MAE gain can be reproduced in large part by the positive part of the OLS predictor, max(0, w_OLS dot f(M)); this baseline is not reported. The authors should also report a zero-inflated baseline (e.g., Tweedie regression or a two-stage classifier plus regressor). If the positive-part baseline achieves MAE close to 0.575 eV, then the claim that the ReLU form is specifically suited to the band gap's mixed statistics is unsupported.","section":"Section IV, Table I"},{"comment":"The jump in the ReLU model's prediction eCDF at 0 eV is partly by construction: for any composition with w dot f(M) <= 0, Eq. (8) returns exactly zero, so the model is guaranteed to produce a point mass at zero whenever a non-negligible fraction of the test set falls in that halfspace. Using this jump as evidence that the model captures the mixed nature of the target is therefore circular unless compared against a non-negativity-enforcing baseline. The manuscript should quantify the similarity of predicted and label eCDFs (e.g., Kolmogorov-Smirnov distance) for the ReLU model and for max(0, linear), and report the fraction of test predictions exactly equal to zero for each.","section":"Section IV, Fig. 3"},{"comment":"The paper does not compare its claimed MAE with published composition-only band gap models, despite citing several such works in Section I (e.g., Zhuo et al. 2018 [39], Venkatraman 2021 [38], Goodall and Lee 2020 [40], CrabNet [42,44]). Without at least one quantitative comparison on the same data and splits, the practical significance of the 0.575 eV MAE is unclear; the model could be either a strong simple baseline or far worse than existing composition-based methods. A table with published MAE values on the same (or closely matched) dataset would make the claim concrete.","section":"Section I and Section IV"}],"minor_comments":[{"comment":"There is a typo in the sentence 'the band gap is actually not a continous random variable'; 'continous' should be 'continuous'.","section":"Section II"},{"comment":"The dataset is described only in Supplementary Section S1; the main text should also state the number of materials (4,603), the 53.2% fraction of zero-gap labels, the restriction to 2-4 element materials, and the provenance of labels (experimental versus Materials Project DFT) so that the reported MAE is interpretable without consulting the supplement.","section":"Section III, last paragraph"},{"comment":"The green shaded region is not defined in the caption beyond 'standard deviation'; please specify whether it is the standard deviation of the eCDF across cross-validation folds at each point or a pointwise interval of another kind.","section":"Section IV, Fig. 3"},{"comment":"The Code Availability section says the code 'will be made available in a public repository' but provides no repository link or DOI; for a reproducible machine learning study, a URL or accession code should be provided at submission time.","section":"Code Availability"},{"comment":"The abstract and title claim 'predicting band gap from chemical composition' without mentioning the 2-4 element restriction and the exclusion of pure-element materials; a brief caveat would prevent overgeneralization of the model's scope.","section":"Abstract and Section III"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is well written and the model is simple, but the central claim that the ReLU form captures the band gap's atypical statistics is not yet established because the comparison with the linear baseline conflates non-negativity with the mixed-distribution design. This is fixable by adding the positive-part-of-OLS baseline, zero-inflated baselines, and a quantitative comparison with published composition-only models. I would also encourage the editor to require the code repository before final acceptance, since the current 'will be made available' statement is insufficient for reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is the one thing to know: this is a lightweight, interpretable composition-only model for band gaps, and the ReLU twist is a sensible idea. But the paper's central claim—that ReLU specifically captures the atypical mixed distribution—is under-supported because the only baseline is an unconstrained linear model. Clamping negative predictions alone would improve MAE on a dataset with 53% zeros, so the reported advantage does not isolate the statistical design.\n\nWhat is good: the model is genuinely simple (one weight per element), the eCDF analysis nicely motivates the mixed random variable framing, and the cross-validated MAE of 0.575 eV is plausible for a composition-only heuristic. The periodic-table visualization is a nice way to show chemical interpretability. The authors are candid that the interpretation is heuristic and that the scope is 2-4 element solids.\n\nThe soft spots. First, the ablation logic. Comparing ReLU to linear OLS confounds the non-linearity with the non-negativity constraint. The right control is max(0, OLS prediction), or a two-stage classifier plus regressor, or zero-inflated regression. Without that, the claim that the model is \"suited for the atypical statistics\" is not established. Second, no comparison to existing composition-only baselines like Goodall-Lee or CrabNet. Even a table with their published numbers on the same dataset would position this properly. Third, the code is not available yet, and fitting details are in supplementary sections not present in the preprint.\n\nNone of these are fatal. The model works as a fast heuristic, and the distribution analysis is worth reading. But the paper needs revision before acceptance.\n\nRecommendation: send it to peer review. A serious referee will ask for the missing baselines and the code. The core idea is small but legitimate, and the paper is clearly written.","headline":"Simple ReLU composition model for band gaps is a useful heuristic, but the paper overclaims its statistical motivation without comparing the right baselines.","tokens_in":14002,"tokens_out":3113,"would_cite":true,"duration_ms":29036,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A model with one learned parameter per chemical element predicts a crystal's electronic band gap from its composition alone, reaching a cross-validated mean absolute error of 0.575 eV by treating the gap as a mixed random variable with a…","keywords":["band gap prediction","chemical composition","machine learning","interpretable model","mixed random variable","ReLU","element parameters","materials informatics"],"falsifier":"Find one chemical formula with two known polymorphs whose measured or high-quality computed band gaps differ by more than about 0.575 eV (for example, carbon as diamond versus graphite); the model is forced to output the same prediction for both, so the larger of the two errors would exceed the claimed average and falsify the composition-only assumption.","tokens_in":13002,"feed_emoji":"⚛️","tokens_out":12207,"duration_ms":103293,"temperature":0.7,"pith_summary":"This paper tries to establish that the electronic band gap of a crystalline solid can be predicted from its chemical formula alone, using a model simple enough to write on a periodic table. The authors first show that the gap is statistically atypical: its empirical distribution has a large discrete mass at exactly 0 eV (metals) plus a continuous positive part, so it is a mixed random variable rather than a standard continuous target. They therefore propose $\\hat{\\varepsilon}_{\\rm relu}(M)=\\mathrm{ReLU}\\big(\\sum_E w_E f_E(M)\\big)$, one learned parameter $w_E$ per element, with a rectified linear unit clamping negative composition-weighted averages to zero. On 4,603 two-, three-, and four-element crystalline compounds, the model reaches a cross-validated test MAE of $0.575 \\pm 0.036$ eV, beating the same model without the clamp ($0.824 \\pm 0.035$ eV), and its learned element weights form a chemically sensible pattern. If this holds, a small lookup table gives instant, interpretable band gap estimates for materials whose crystal structure may not even be known.","feed_headline":"One learned weight per element predicts band gap from formula alone","feed_subtitle":"A ReLU clamp on a composition average beats a linear model: 0.575 vs 0.824 eV MAE, with interpretable element weights.","key_machinery":"The load-bearing object is the ReLU composition model $\\hat{\\varepsilon}_{\\rm relu}(M)=\\max\\big(0,\\sum_E w_E f_E(M)\\big)$, where $f_E(M)$ is the fraction of element $E$ in the chemical formula and $w_E$ is the single learned parameter for that element. The rectified linear unit (the maximum of the input and zero) is the mechanism that matches the target statistics: it enforces non-negativity and concentrates all negative pre-activations into a point mass at exactly 0 eV, mirroring the empirical distribution of the labels. The element-fraction representation is the only input — no crystal structure, bond geometry, or coordination enters — so the fitted model is entirely summarized by the vector of element weights.","core_discovery":"The paper's central claim is that band gap prediction should be treated as modeling a mixed random variable, and that the right simple model for this task is a composition-weighted average of per-element parameters followed by a ReLU: $\\hat{\\varepsilon}_{\\rm relu}(M)=\\mathrm{ReLU}\\big(w\\cdot f(M)\\big)$. The discrete mass of metals at 0 eV is reproduced exactly by the clamp, and the continuous positive part is reproduced by the linear pre-activation for compounds whose weighted element score is positive. Trained by gradient descent on mean squared error over 4,603 materials and evaluated with 10-fold cross-validation, this model attains a test MAE of $0.575 \\pm 0.036$ eV, outperforming the linear baseline $\\hat{\\varepsilon}_{\\rm linear}(M)=w\\cdot f(M)$ at $0.824 \\pm 0.035$ eV. The learned weights, visualized on the periodic table, put positive values on the ionic, large-gap elements of the left and right columns and negative values on metallic transition elements, which the authors present as heuristic chemical interpretability: elements with larger parameters tend to form materials with larger band gaps.","pith_inferences":["A direct test of the chemical interpretation would be to regress the learned element weights against electronegativity values; a strong correlation would confirm the model has rediscovered a known heuristic rather than a dataset artifact.","Because the model outputs one gap per formula, its reported error averages over the training distribution of structures; for same-formula polymorphs with very different gaps, the error on at least one member of each pair is bounded below by half the gap difference, so the MAE should not be read as a per-material guarantee.","The same ReLU-of-weighted-average construction could be applied to other material properties with a physical floor and a large atom of observations at that floor; the accuracy gain from the clamp should scale with the size of the zero mass in the target distribution."],"forward_implications":["A lookup table of element weights gives an instant band gap estimate for any 2-4 element crystalline solid, including materials with unknown crystal structure, at the reported cross-validated MAE of $0.575 \\pm 0.036$ eV.","The ReLU clamp is doing real work: the same model without it has a test MAE of $0.824 \\pm 0.035$ eV, makes negative predictions, and fails to reproduce the jump at 0 eV in the label distribution.","The learned element weights reproduce familiar chemistry — positive weights for ionic left- and right-side elements, negative weights for central transition metals — so the model doubles as a heuristic chemical rule for gap size.","Modeling the gap as a mixed random variable with a single model removes the need for the two-stage classify-then-regress pipeline sometimes used for band gap prediction."],"supporting_citations":[{"why":"Supplies the band gap dataset (experimental gaps plus 0 eV labels) used for training and evaluation.","marker":"[39]"},{"why":"Provides the benchmark data processing and cross-validation setup used in the paper.","marker":"[48]"},{"why":"Provides the DFT-calculated 0 eV band gap labels for metals in the dataset.","marker":"[10]"},{"why":"Introduces the one-parameter-per-element weighted-average modeling style that this paper adapts from classification to band gap regression.","marker":"[46]"},{"why":"Extends the same modeling style to metallicity classification, motivating the connection between gap and metal/nonmetal labels.","marker":"[47]"},{"why":"Shows prior composition-only ML for band gap prediction, the context that the ReLU model is designed to improve upon in interpretability and statistical match.","marker":"[38]"}],"fun_headline_variants":["Simple ReLU model predicts band gaps from composition alone","One weight per element: a simple ReLU model beats linear for band gaps","ReLU clamp on element weights gives interpretable band gap predictions","Band gap from formula: a learned model with per-element weights and a twist","Mixed statistics inspire simple model: ReLU on composition average"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that a material's band gap is fully determined by its chemical formula through a fixed piecewise-linear rule, so that all structure-dependent physics such as bonding geometry, coordination, and polymorph identity can be ignored.","fun_headline_variants_meta":{"raw":{"variants":["Simple ReLU model predicts band gaps from composition alone","One weight per element: a simple ReLU model beats linear for band gaps","ReLU clamp on element weights gives interpretable band gap predictions","Band gap from formula: a learned model with per-element weights and a twist","Mixed statistics inspire simple model: ReLU on composition average"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000447,"raw_usage":{"total_tokens":2289,"prompt_tokens":1011,"completion_tokens":1278,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":1189}},"tokens_in":627,"tokens_out":1278,"duration_ms":8570,"temperature":1.0,"reasoning_tokens":1189,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:00:22.768366+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Find one chemical formula with two known polymorphs whose measured or high-quality computed band gaps differ by more than about 0.575 eV (for example, carbon as diamond versus graphite); the model is forced to output the same prediction for both, so the larger of the two errors would exceed the claimed average and falsify the composition-only assumption.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the band gap dataset (experimental gaps plus 0 eV labels) used for training and evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the benchmark data processing and cross-validation setup used in the paper."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the DFT-calculated 0 eV band gap labels for metals in the dataset."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the one-parameter-per-element weighted-average modeling style that this paper adapts from classification to band gap regression."},{"cited_title":"Ma and M","cited_arxiv_id":null,"evidence_quote":"Extends the same modeling style to metallicity classification, motivating the connection between gap and metal/nonmetal labels."},{"cited_title":"Venkatraman, The utility of composition-based machine learning models for band gap prediction, Computational Materi- als Science 197, 110637 (2021)","cited_arxiv_id":null,"evidence_quote":"Shows prior composition-only ML for band gap prediction, the context that the ReLU model is designed to improve upon in interpretability and statistical match."}],"review_version":1}