{"id":"1b2709e8-70de-4d86-a375-b9970c00ab85","arxiv_id":"2605.25866","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"UNATE uses unsupervised denoising autoencoder and contrastive learning on unlabeled crystals to create atomic embeddings that improve downstream property prediction by 2.7% overall and up to 10% with only 25% labeled data.","lead":"UNATE learns atomic embeddings from unlabeled crystal structures via a denoising autoencoder combined with contrastive learning, then feeds those embeddings into property prediction models. This could reduce the amount of expensive labeled data or simulations needed for materials discovery.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Improvement may arise from higher-dimensional input features rather than learned content of the embeddings","rationale":"The reader's weakest assumption correctly flags transferability, but the more immediate and testable gap is the missing control for input dimensionality and capacity. The proposed check directly isolates whether the pretraining step itself is responsible for the measured gains.","tokens_in":1627,"tokens_out":300,"duration_ms":27892,"concrete_test":"Re-train the downstream property-prediction model three ways on the same splits: (1) raw atomic numbers, (2) UNATE embeddings, (3) randomly initialized embeddings of identical dimension, all with identical GNN architecture and hyper-parameters; if (3) matches or exceeds (2), the headline improvement is not due to the unsupervised pretraining.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim replaces scalar atomic numbers with vector embeddings produced by the unsupervised denoising autoencoder + contrastive objective. Because this changes both the semantic content and the dimensionality of the node features fed to the downstream GNN, any gain could be explained by increased model capacity or richer initial representation rather than the specific structural features captured during pretraining. The low-data regime (25 % labels) is precisely where extra capacity tends to help most, so the reported 2.7 % / 10 % deltas cannot be attributed to transferability without a control that holds embedding dimension fixed while randomizing or freezing the embedding values.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.3","summary":"The manuscript proposes UNATE, an unsupervised framework combining a denoising autoencoder with self-supervised contrastive learning to derive atomic embeddings from unlabeled crystal structures. These embeddings replace scalar atomic numbers as node features for downstream GNN-based crystal property prediction. The abstract reports a 2.7% improvement over the full-data baseline and gains up to 10% when only 25% of labeled data is available.","tokens_in":1748,"tokens_out":413,"duration_ms":19777,"significance":"If the gains are shown to arise from transferable structural features captured during pretraining (rather than dimensionality or capacity changes), the approach could meaningfully extend self-supervised pretraining techniques to materials science, especially for low-data regimes where labeled crystal properties are scarce. The low-data emphasis is a potential strength if supported by rigorous controls.","major_comments":[{"comment":"The central claim attributes the 2.7% / 10% gains to the learned content of the UNATE embeddings. Because the method replaces scalar atomic numbers with higher-dimensional vectors, any improvement could stem from increased input dimensionality or model capacity rather than the specific structural features learned by the denoising+contrastive objective. A control experiment holding embedding dimension fixed while randomizing or freezing the embedding values is required to isolate the effect of pretraining; without it the attribution to transferability remains unverified.","section":"Experimental results (as described in abstract)"},{"comment":"The abstract states numerical improvements but provides no information on datasets, baselines, statistical tests, ablation studies, or embedding dimensions. This absence prevents verification of whether the reported deltas are load-bearing for the transferability claim.","section":"Abstract"}],"minor_comments":[{"comment":"The abstract should specify the downstream property prediction tasks, the GNN architectures employed, and the crystal structure datasets used for pretraining and evaluation.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the detailed and constructive feedback. The two major comments highlight important issues regarding experimental controls and abstract clarity. We address each below and commit to revisions that directly respond to the concerns.","responses":[{"response":"We agree that the current experiments do not fully isolate the contribution of the learned embeddings from the effect of increased dimensionality. A control using random vectors of the same dimension (or frozen/randomized pretrained embeddings) is a necessary addition. In the revised manuscript we will include such controls on the same downstream tasks and data regimes, allowing direct comparison to the UNATE embeddings. This will strengthen the attribution to the unsupervised pretraining objective.","revision_made":"yes","referee_comment":"[Experimental results (as described in abstract)] The central claim attributes the 2.7% / 10% gains to the learned content of the UNATE embeddings. Because the method replaces scalar atomic numbers with higher-dimensional vectors, any improvement could stem from increased input dimensionality or model capacity rather than the specific structural features learned by the denoising+contrastive objective. A control experiment holding embedding dimension fixed while randomizing or freezing the embedding values is required to isolate the effect of pretraining; without it the attribution to transferability remains unverified."},{"response":"We acknowledge that the abstract's brevity omits key experimental details. We will revise the abstract to concisely specify the primary datasets, the GNN architectures used as baselines, the embedding dimension, and that improvements were evaluated with statistical significance across multiple random seeds. These additions will be kept within standard abstract length limits while improving verifiability.","revision_made":"yes","referee_comment":"[Abstract] The abstract states numerical improvements but provides no information on datasets, baselines, statistical tests, ablation studies, or embedding dimensions. This absence prevents verification of whether the reported deltas are load-bearing for the transferability claim."}],"tokens_in":1272,"tokens_out":413,"duration_ms":16446,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The core result here is that swapping raw atomic numbers for embeddings pretrained on unlabeled crystals improves downstream GNN accuracy, with the lift growing as labeled data shrinks to 25%. That matches a real pain point in materials modeling where DFT labels are expensive.\n\nWhat the work does cleanly is take two off-the-shelf self-supervised pieces—denoising reconstruction and contrastive alignment—and run them on crystal graphs to produce node features that transfer. The abstract positions this as a drop-in replacement that helps most when supervision is limited, which is the regime that matters.\n\nThe main weakness is the missing control for dimensionality. Atomic numbers are scalars; the new embeddings are vectors. Any capacity increase from richer initial features can explain the 2.7 % and 10 % deltas without any transfer of structural knowledge. The low-data regime is exactly where extra dimensions tend to help most, so the claim that the unsupervised content is doing the work needs an ablation that keeps dimension fixed while randomizing or freezing the embedding values. No such check is described.\n\nDatasets, exact baselines, statistical significance, and ablation tables are also absent from the abstract, which leaves the numerical claims hard to evaluate. The method itself is a conventional pretrain-then-finetune pipeline rather than a new algorithmic primitive.\n\nThis paper is aimed at groups already running GNNs on crystal structures who want to squeeze more out of small labeled sets. A reader already working in that niche could extract the pretraining recipe and test it themselves. It is coherent on its own terms and shows honest engagement with the low-label bottleneck, so it clears the bar for a serious referee even though the current evidence is thin. I would send it out for review with a request for the dimensionality control and full experimental details.","headline":"UNATE applies a standard denoising autoencoder plus contrastive pretraining recipe to crystal graphs and claims gains on property prediction in low-label settings, but the reported improvements could stem from higher-dimensional node features rather than the learned embeddings themselves.","tokens_in":2204,"tokens_out":447,"would_cite":false,"duration_ms":15534,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"UNATE learns atomic embeddings from unlabeled crystal structures that improve downstream property prediction by 2.7 percent overall and up to 10 percent with limited labels.","keywords":["unsupervised learning","atomic embeddings","crystal structures","property prediction","contrastive learning","denoising autoencoder","materials discovery","graph representations"],"falsifier":"No accuracy gain, or a loss, when the UNATE embeddings are substituted for atomic numbers on a new crystal property or an independent dataset would falsify the central claim.","tokens_in":2538,"feed_emoji":"🔬","tokens_out":426,"duration_ms":19037,"temperature":0.7,"pith_summary":"The paper presents UNATE as a way to extract useful atomic representations solely from unlabeled crystal structures. It trains a denoising autoencoder combined with contrastive learning on crystal graphs to produce node embeddings. These embeddings replace raw atomic numbers as input features for models that predict crystal properties. The approach yields measurable accuracy gains on standard benchmarks, with the largest benefits appearing when labeled data for the prediction task is reduced to 25 percent of the full set. This directly targets the scarcity of labeled examples that limits many materials property models.","feed_headline":"Unsupervised atomic embeddings raise crystal prediction accuracy 2.7%","feed_subtitle":"Gains reach 10% when labeled data is cut to one quarter of the full set","key_machinery":"UNATE, an unsupervised framework that integrates a denoising autoencoder with self-supervised contrastive learning to generate atomic node embeddings from crystal graphs.","core_discovery":"Replacing raw atomic numbers with node embeddings pretrained by UNATE on unlabeled crystals produces a 2.7 percent improvement over the full-data baseline for property prediction; the same substitution yields gains up to 10 percent when only 25 percent of the labeled data is supplied to the downstream model.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["UNATE embeddings give 2.7% crystal prediction gains","Crystal property prediction improves 2.7% with UNATE","UNATE on unlabeled crystals yields 2.7% prediction gains","UNATE embeddings deliver 10% gains with 25% labeled data"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Embeddings learned only from unlabeled crystal structures contain structural features that transfer usefully to the specific property prediction tasks tested.","fun_headline_variants_meta":{"raw":{"variants":["UNATE embeddings give 2.7% crystal prediction gains","Crystal property prediction improves 2.7% with UNATE","UNATE on unlabeled crystals yields 2.7% prediction gains","UNATE embeddings deliver 10% gains with 25% labeled data"]},"model":"grok-4.3","cost_usd":0.006435,"raw_usage":{"total_tokens":2957,"prompt_tokens":551,"num_sources_used":0,"completion_tokens":72,"cost_in_usd_ticks":64349500,"prompt_tokens_details":{"text_tokens":551,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2334,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":551,"tokens_out":72,"duration_ms":18827,"temperature":1.0,"reasoning_tokens":2334,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-29T22:53:21.772736+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"No accuracy gain, or a loss, when the UNATE embeddings are substituted for atomic numbers on a new crystal property or an independent dataset would falsify the central claim.","supporting_citations":[],"review_version":1}