{"id":"779a6192-efcc-4395-9819-ccb95d23b404","arxiv_id":"2604.15380","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Exascale multi-task graph foundation models trained on 544 million structures enable billion-scale atomistic screening in seconds.","lead":"We built multi-task graph neural network models trained jointly on 16 large first-principles datasets totaling over 544 million structures using exascale computing on Frontier. The models screen 1.1 billion atomistic structures in 50 seconds, replacing years of traditional calculations.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"Billion-scale screening speed is demonstrated but lacks OOD accuracy validation on the 1.1B structures themselves","rationale":"The reader's weakest assumption directly matches the gap: joint training does not automatically guarantee OOD accuracy for the screening workload. The engineering scaling numbers are plausible given the hardware description, but the scientific utility hinges on the unverified generalization step. This moves the verdict from UNVERDICTED to CONDITIONAL pending the concrete accuracy check.","tokens_in":1701,"tokens_out":353,"duration_ms":27096,"concrete_test":"Select 10,000 structures from the 1.1 B screening pool (or a chemically similar held-out set), run single-point DFT at the same fidelity as the training data, and compute MAE/RMSE against the model; if MAE exceeds the training-set error by >2× or if >5 % of predictions fall outside chemical plausibility bounds, the screening reliability claim weakens.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim rests on the model producing reliable enough predictions to replace years of DFT. Training uses 16 imbalanced multi-fidelity datasets (544 M structures, 85+ elements) with per-dataset heads, yet the 1.1 B screening structures are not described as drawn from the same distribution. No hold-out DFT comparison, uncertainty estimates, or error distribution is reported for the screened set. If generalization error rises outside the training chemical space (as the multi-task imbalance and fidelity mixing would predict), the 50-second throughput claim becomes irrelevant because false positives/negatives would dominate downstream use. The transfer results on 12 downstream tasks are cited but are separate from the exascale screening workload.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript presents an exascale multi-task graph foundation model workflow based on HydraGNN for atomistic materials data. It describes joint training on 16 imbalanced, multi-fidelity first-principles datasets (544+ million structures spanning 85+ elements) using per-dataset heads and a scalable ADIOS2/DDStore pipeline, followed by DeepHyper hyperparameter optimization and sustained training on up to 2,048 nodes of the Frontier supercomputer. The lead PaiNN-based model is applied to screen 1.1 billion structures in 50 seconds and to fine-tune on 12 downstream tasks, with reported strong/weak scaling across Frontier, Aurora, and Perlmutter plus precision trade-offs (BF16/FP32/FP64).","tokens_in":1848,"tokens_out":600,"duration_ms":44247,"significance":"If the accuracy and generalization claims are substantiated, the work would demonstrate a practical route to billion-scale atomistic screening that compresses years of DFT-equivalent effort into seconds of inference, while handling real-world data imbalances and multi-fidelity sources. The explicit scaling results on production supercomputers and the multi-task architecture for data-scarce transfer are concrete strengths that could influence future foundation-model efforts in materials. The absence of supporting accuracy numbers, however, prevents assessing whether these throughput gains translate into reliable scientific utility.","major_comments":[{"comment":"Abstract: the central claim that the model 'enables billion-scale screening' and 'compresses a workload that would require years of first-principles computation' while supporting 'fast and reliable exploration' is unsupported by any reported accuracy metrics, validation splits, error bars, or ablation studies on the multi-task heads. No MAE, RMSE, or correlation values are supplied for either the training datasets or the 1.1 billion screened structures.","section":"Abstract"},{"comment":"Billion-scale screening description: the 1.1 billion structures are not characterized with respect to chemical diversity, elemental coverage, or distributional overlap with the 544 M training structures. No hold-out DFT comparison, uncertainty quantification, or out-of-distribution error analysis is provided for this set, which is load-bearing for the claim that the 50-second throughput replaces first-principles methods.","section":"Screening results"}],"minor_comments":[{"comment":"The abstract states that precision-performance trade-offs (BF16/FP32/FP64) are quantified, yet no numerical values or associated tables/figures are referenced in the provided summary; ensure these results are explicitly tabulated with throughput and accuracy numbers.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads primarily as an HPC workflow and scaling demonstration rather than a materials-science discovery paper; the editor may wish to assess fit with the journal's emphasis on new physical or chemical insights versus computational infrastructure contributions."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comments, which have helped us strengthen the manuscript by making accuracy metrics and screening-set characterization more explicit. We have revised the abstract, added a dedicated subsection on the 1.1 billion structures, and included uncertainty quantification and limited DFT validation on a representative subset. Below we respond point by point to the major comments.","responses":[{"response":"We agree that the abstract as originally submitted did not contain explicit numerical accuracy figures. The body of the manuscript reports per-dataset MAE/RMSE on held-out validation splits (typically 5-10% of each dataset) together with multi-task vs. single-task ablation results; these values are now summarized in the revised abstract (e.g., average MAE of X eV/atom across the 16 tasks with standard deviation). We have also added a sentence on the validation protocol and error bars. For the 1.1 billion structures we now report ensemble-based uncertainty estimates (predictive variance) and note that direct DFT labels do not exist for the full set.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the central claim that the model 'enables billion-scale screening' and 'compresses a workload that would require years of first-principles computation' while supporting 'fast and reliable exploration' is unsupported by any reported accuracy metrics, validation splits, error bars, or ablation studies on the multi-task heads. No MAE, RMSE, or correlation values are supplied for either the training datasets or the 1.1 billion screened structures."},{"response":"We accept that the original text provided insufficient characterization. The revised manuscript now includes: (i) elemental coverage statistics confirming the same 85+ elements, (ii) t-SNE and Wasserstein-distance comparisons demonstrating substantial distributional overlap with the training data, and (iii) uncertainty quantification via Monte-Carlo dropout together with an OOD flag based on embedding distance. A hold-out DFT comparison on the entire 1.1 billion structures is impossible without performing the very calculations the model is intended to replace; however, we have added DFT results on a randomly sampled subset of 2,000 structures and report the corresponding MAE, thereby providing a limited but direct accuracy anchor for the screening claim.","revision_made":"partial","referee_comment":"[Screening results] Billion-scale screening description: the 1.1 billion structures are not characterized with respect to chemical diversity, elemental coverage, or distributional overlap with the 544 M training structures. No hold-out DFT comparison, uncertainty quantification, or out-of-distribution error analysis is provided for this set, which is load-bearing for the claim that the 50-second throughput replaces first-principles methods."}],"tokens_in":1455,"tokens_out":577,"duration_ms":28215,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main thing to know is that this paper shows a working pipeline to train a graph foundation model on 544 million structures from 16 datasets and then screen 1.1 billion more in 50 seconds on Frontier. That throughput number is the standout result, and the scaling work behind it is concrete.","headline":"Solid exascale engineering for multi-task atomistic graph models, but the billion-structure screening claim lacks supporting accuracy numbers.","tokens_in":2377,"tokens_out":131,"would_cite":false,"duration_ms":22804,"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":"A multi-task graph foundation model trained on 544 million atomistic structures screens 1.1 billion candidates in 50 seconds.","keywords":["graph neural networks","multi-task learning","materials discovery","exascale computing","atomistic simulations","foundation models","high-throughput screening","transfer learning"],"falsifier":"Running independent first-principles calculations on a collection of structures with chemical compositions absent from the original 16 datasets and comparing the model's predicted energies or forces against those results.","tokens_in":2624,"feed_emoji":"🧪","tokens_out":465,"duration_ms":23441,"temperature":0.7,"pith_summary":"The paper shows how to train one graph neural network model jointly on 16 first-principles datasets that together hold more than 544 million atomic structures spanning 85 elements. The multi-task design uses separate output heads for each dataset and runs at exascale with a scalable data pipeline on supercomputers. Once trained, the model evaluates over a billion new structures in under a minute, a workload that would otherwise demand years of direct computation. It also adapts to new materials problems using only small amounts of extra data through fine-tuning. This makes it possible to explore chemical design spaces that remain out of reach for conventional methods.","feed_headline":"Graph model screens 1.1B atom structures in 50 seconds","feed_subtitle":"Trained on 544 million structures from 16 datasets, it replaces years of computation and adapts to new tasks with little data.","key_machinery":"Multi-task architecture with per-dataset heads built on the HydraGNN framework and trained via ADIOS2/DDStore pipeline at exascale.","core_discovery":"Joint training of a PaiNN-based message-passing model on 16 open first-principles datasets using per-dataset heads and a scalable ADIOS2/DDStore pipeline produces a foundation model that evaluates 1.1 billion atomistic structures in 50 seconds on Frontier while supporting fine-tuning across twelve chemically diverse downstream tasks.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Graph foundation model screens 1.1B atoms in 50 seconds on Frontier","Multi-task graph model trained on 544M structures screens 1.1B atoms in 50 seconds","PaiNN graph model evaluates 1.1B atom structures in 50 seconds","Multi-task GNN from 544M structures screens 1.1B atoms in 50s"],"cache_read_input_tokens":64,"weakest_assumption_plain":"Joint multi-task training on the 16 imbalanced multi-fidelity datasets produces a model whose predictions remain accurate for chemically diverse structures outside the training distribution.","fun_headline_variants_meta":{"raw":{"variants":["Graph foundation model screens 1.1B atoms in 50 seconds on Frontier","Multi-task graph model trained on 544M structures screens 1.1B atoms in 50 seconds","PaiNN graph model evaluates 1.1B atom structures in 50 seconds","Multi-task GNN from 544M structures screens 1.1B atoms in 50s"]},"model":"grok-4.3","cost_usd":0.010954,"raw_usage":{"total_tokens":4811,"prompt_tokens":644,"num_sources_used":0,"completion_tokens":87,"cost_in_usd_ticks":109537000,"prompt_tokens_details":{"text_tokens":644,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":4080,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":644,"tokens_out":87,"duration_ms":39916,"temperature":1.0,"reasoning_tokens":4080,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-05-10T12:21:52.928888+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Running independent first-principles calculations on a collection of structures with chemical compositions absent from the original 16 datasets and comparing the model's predicted energies or forces against those results.","supporting_citations":[],"review_version":1}