{"id":"283076df-0667-4ef5-af9a-0383d7e896fb","arxiv_id":"2508.17792","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Universal machine learning interatomic potentials systematically lose accuracy under pressure up to 150 GPa, and fine-tuning on high-pressure DFT data recovers most of the lost performance.","lead":"This paper builds a large database of DFT calculations at pressures up to 150 GPa and shows that eight popular machine learning interatomic potentials become less accurate as pressure rises. Fine-tuning two of the models on high-pressure data restores most of their accuracy, pointing to missing training data as the main culprit.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pressure trend may be confounded by comparing MAEs on different material subsets: 0 GPa has 162k materials, elevated pressures have 187-190k.","rationale":"The reader's weakest assumption concerns non-converged high-pressure exclusions, but Table I shows a larger and potentially more serious composition mismatch at 0 GPa: 28k older Alexandria materials lack 0 GPa relaxation paths yet are included at high pressures. This directly threatens the central claim that accuracy deteriorates with increasing pressure, because the 0 GPa baseline and the high-pressure points are not computed on the same materials. The causal conclusion that the decline comes from training data rather than algorithmic constraints inherits this problem. A matched-set reanalysis is a simple, decisive check. I agree with the reader's conditional verdict; this concern adds a sharper technical condition that should be addressed before the benchmark is treated as a community standard. The paper is otherwise a valuable benchmark, the dataset is large and consistently computed, and the fine-tuning demonstration is reasonable in-distribution, but the composition confound must be ruled out.","tokens_in":11140,"tokens_out":7188,"duration_ms":88386,"concrete_test":"Restrict the benchmark to the intersection of materials with converged data at every pressure (0, 25, 50, 75, 100, 125, 150 GPa), preserving the same material-level split, and recompute Tables II and III. If the energy/volume MAEs still increase monotonically with pressure on this matched set, the pressure effect is confirmed; if the trend flattens or the 150 GPa MAE drops substantially, the reported deterioration is at least partly an artifact of comparing different material subsets.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central empirical trend is computed on non-identical material pools across pressures. Table I lists 162,000 materials at 0 GPa but 187,000-190,000 at every elevated pressure; the 0 GPa deficit is explicitly due to missing relaxation paths in older Alexandria entries, not only to non-convergence. Because the 90/5/5 split is applied at the level of all 190k materials, a material absent at 0 GPa can still appear in the test set at 25-150 GPa. If the ~28k older entries are systematically more difficult or less standard, the monotonic MAE increases in Tables II and III could partly reflect a change in test composition rather than pressure. The paper provides no matched-set analysis (e.g., intersection of materials with converged data at all pressures), so the headline claim that accuracy deteriorates with pressure, and the causal attribution to training data, are not yet cleanly established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a large DFT dataset of roughly 190k materials relaxed with PBE at pressures from 0 to 150 GPa, and uses it to benchmark eight universal machine learning interatomic potentials (uMLIPs) by comparing PBE reference volumes and energies against model relaxations. The authors report that most uMLIPs become substantially less accurate with increasing pressure, and show that fine-tuning two representative models (MatterSim-v1 and eSEN-30M-OAM) on high-pressure configurations largely recovers accuracy. The central claim is that the pressure-induced degradation stems from limitations in the training data rather than algorithmic constraints, and that targeted fine-tuning is an effective remedy.","tokens_in":11380,"tokens_out":3680,"duration_ms":49900,"significance":"If the central claim holds, this work provides a valuable resource: a consistent, large-scale high-pressure DFT dataset, a multi-model benchmark of practical relevance, and a concrete demonstration that fine-tuning on high-pressure configurations improves transferability. The study uses consistent DFT settings, trajectory-level split to avoid leakage, and a broad set of contemporary uMLIPs. The dataset and code are promised to be released. However, the quantitative headline is currently supported mainly by aggregate MAEs from a single split, with no matched-set control across pressures and no analysis of the discarded non-converged calculations, so several load-bearing components of the claim need additional evidence before the conclusions can be accepted at face value.","major_comments":[{"comment":"The pressure trend is computed on non-identical material pools. Table I shows 162k materials at 0 GPa versus 187-190k at every elevated pressure, with the 0 GPa deficit explicitly attributed to missing relaxation paths, not only to non-convergence. Because the 90/5/5 split is applied at the level of the full 190k set, a material absent at 0 GPa can nevertheless be in the test set at 25-150 GPa. The monotonic MAE increases in Tables II and III could therefore be partly caused by a change in test composition rather than pressure. Please provide a matched-set analysis, e.g., restricting to materials with converged data at all pressures, and report the test-set compositions; if the trend persists, this concern is fully resolved.","section":"Section II.A, Table I vs Tables II/III"},{"comment":"The manuscript drops calculations that did not converge under pressure without further analysis. Table I states that missing materials under pressure 'concern calculations that did not converge,' but there is no discussion of how many materials are affected per pressure, why they failed, or whether they are systematically the hardest high-pressure systems. If non-converged cases are preferentially difficult, the reported errors are optimistic and fine-tuning gains may not transfer to the most challenging materials. Please quantify the convergence failures and perform a sensitivity analysis, e.g., comparing models on the converged subset against bounds that include the non-converged systems.","section":"Table I caption and Section II.A"},{"comment":"All quantitative claims rest on a single 90/5/5 split with no uncertainty estimates. The differences between some models and the fine-tuning improvements are small relative to the spread of values (e.g., eSEN at 150 GPa changes from 41.7 to 32.4 meV/atom; several models differ by only a few meV/atom at low pressure). Without repeated splits, bootstrap confidence intervals, or multiple seeds, the reported ranking and the magnitude of the fine-tuning benefit are not statistically grounded. Please provide uncertainty estimates for the MAEs in Tables II and III, at least for the test-set metric.","section":"Tables II/III and Section II.B"},{"comment":"The attribution 'originates from fundamental limitations in the training data rather than algorithmic constraints' is stronger than the evidence supports. Fine-tuning on high-pressure data improves the two tested models, which is consistent with a data-coverage explanation, but it does not rule out algorithmic constraints: eSEN-30M-OAM, without explicit high-pressure fine-tuning, is already the best model under pressure, suggesting that training strategy or architecture has a large effect. A cleaner test would be to train or fine-tune the same architecture on datasets with and without high-pressure configurations, or at least to soften the causal claim and limit it to 'the models' original training distributions lack sufficient high-pressure coverage.'","section":"Abstract and Section III"}],"minor_comments":[{"comment":"Typo: 'dataset develop in this work' should be 'dataset developed in this work.'","section":"Table I caption"},{"comment":"Duplicate phrase: 'materials materials science' near the end of the introductory paragraph.","section":"Section I"},{"comment":"The dataset size is given as 32 million single-point calculations in Section II.A and 30 million in Section III. Please make this consistent.","section":"Section II.A vs Section III"},{"comment":"The code is referred to as 'v asp' with missing spacing; should be 'VASP.'","section":"Section IV.A"},{"comment":"Axis labels use 'A3' and 'A' without proper superscripts or math formatting; this should be fixed for clarity.","section":"Figures 3/4"},{"comment":"The reasons for MACE-MPA-0's 'density renormalization' and its effect on high-pressure performance are mentioned but not explained or quantitatively assessed. A brief explanation or citation would help.","section":"Section II.B"},{"comment":"The exact model checkpoints and versions (e.g., MatterSim-v1.0.0-5M) are listed, but for full reproducibility the commit hashes or release IDs of the pretrained checkpoints and libraries should be provided.","section":"Section IV.B"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth knowing about: this paper builds a 32-million-configuration PBE relaxation dataset from 0 to 150 GPa and benchmarks eight uMLIPs plus two fine-tuned variants on it. The dataset is the real contribution. A shared benchmark of this kind, with consistent DFT settings and relaxation paths, has been missing, and the community will likely use it. The side-by-side comparison is also useful, and the authors are honest about model-specific quirks (e.g., ORB's over-pressurization, M3GNet's poor compressibility). The fine-tuning demonstration is a reasonable proof of concept, though it is in-distribution by design.\n\nThe central empirical trend, rising energy MAE with pressure, is visible in the tables and is consistent across all eight models. That part is solid. But the causal claim that the decline comes from training data rather than algorithmic constraints is an overreach. Nothing in the data cleanly separates those two factors. The stress-test concern is real: the 0 GPa test set contains 162k materials while elevated pressures use 187-190k, and the 0 GPa deficit includes older Alexandria entries missing relaxation paths, not just non-converged calculations. Since the split is on all 190k materials, a material absent at 0 GPa can appear in the test set at higher pressures. If those older entries are systematically harder, some of the monotonic MAE increase could come from test composition rather than pressure. A matched-set analysis, using only materials converged at every pressure, would settle this. The paper should also report error bars or repeated splits; the differences between many models at a given pressure look smaller than the noise you would expect from a single split of 9,000 test materials.\n\nThe exclusion of non-converged calculations is a minor concern but should be quantified. If non-convergence correlates with difficult high-pressure structures, the reported errors are optimistic. The authors mention the exclusion only in a table caption.\n\nI would not block this paper on these issues. The dataset and the benchmark are a step forward, and the main trend will likely survive matched-set analysis. But the causal framing should be toned down, and the dataset and code should be public for the benchmark to become a community standard. This deserves serious peer review, with attention to the matched-set point and the missing error bars.\n\nFor a reading group, I would say yes: it is a compact, useful paper that invites discussion about how to benchmark extrapolation.","headline":"Useful high-pressure benchmark and dataset, but the headline causal claim outruns the evidence; the pressure trend would be cleaner with matched-set analysis.","tokens_in":742,"tokens_out":753,"would_cite":true,"duration_ms":22742,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Universal machine-learning potentials lose accuracy under pressure; fine-tuning restores it","keywords":["machine learning interatomic potentials","universal MLIPs","high pressure","DFT dataset","fine-tuning","extrapolation","materials discovery"],"falsifier":"Take the materials whose high-pressure relaxations did not converge, re-relax them with a more robust DFT protocol, and measure the energy and volume errors of the fine-tuned models on that held-out set; if the energy error is substantially larger than 31 meV/atom, the reported pressure recovery is incomplete for the hardest systems.","tokens_in":11052,"feed_emoji":"📉","tokens_out":8311,"duration_ms":83905,"temperature":0.7,"pith_summary":"Universal machine-learning interatomic potentials (uMLIPs) aim to give density-functional-theory accuracy at a tiny fraction of the cost, but this paper argues their reliability stops at ambient pressure. Benchmarking eight leading uMLIPs on a new dataset of about 190,000 materials and 32 million density-functional calculations, the authors find energy errors that grow from a few meV per atom at 0 GPa to tens or hundreds of meV per atom at 150 GPa, with the worst degradation in models trained only on near-equilibrium ambient structures. The same data, used to fine-tune two of the models, brings the 150 GPa error back down to about 31 meV per atom. The paper concludes that the pressure blind spot is a training-data problem, not an algorithmic one, and that compressed configurations belong in the next generation of universal potentials.","feed_headline":"Machine-learning force fields err up to 30x more at 150 GPa","feed_subtitle":"A new 32-million-configuration dataset shows training data, not algorithms, are the bottleneck.","key_machinery":"The load-bearing object is a new pressure dataset built by taking about 190,000 structures from the base DFT database, re-relaxing them with the same PBE settings at 0, 25, 50, 75, 100, 125, and 150 GPa, and recording equilibrium structures, total energies, forces, stress tensors, and relaxation-path configurations—32 million single-point calculations in all. The benchmark protocol then runs each uMLIP through the same relaxations and compares final relaxed energies and volumes to the PBE reference, while the 90–5–5 material-level split prevents leakage between training, validation, and test sets.","core_discovery":"The central claim is that uMLIP accuracy under pressure is governed by whether compressed atomic environments appear in the training distribution. Across pressures from 0 to 150 GPa, the benchmark shows a systematic decline: the least-affected model goes from 4.1 to 41.7 meV/atom mean absolute energy error, while an earlier ambient-trained model rises from 33.8 to 346.8 meV/atom. Models that saw high-pressure or non-equilibrium structures during training degrade less, and targeted fine-tuning on the pressure dataset reduces the best 150 GPa error to about 31 meV/atom at the cost of a modest loss at ambient pressure. The authors read this as evidence that the bottleneck is data coverage, not","pith_inferences":["The reported fine-tuned errors are computed on materials whose high-pressure relaxations converged; if the roughly one to three thousand non-converged materials per pressure are the hardest cases, real-world errors on those systems could be larger—a direct test would be to relax them with a more robust protocol and remeasure.","The same fine-tuning recipe was demonstrated on only two model families; applying it to the other benchmarked uMLIPs would likely shrink their high-pressure errors too, since their degradation follows the same data-coverage pattern.","The result suggests a design rule for the next generation: sample the compressed structural manifold explicitly rather than relying on ambient datasets augmented by rattling or volume scaling.","Nothing in the data establishes behavior beyond 150 GPa; extending to terapascal conditions, relevant for planetary interiors, would require new training data in that regime."],"forward_implications":["At pressures above 25 GPa, untuned uMLIP predictions of relaxed energies and volumes should be treated as unvalidated; errors can exceed 100 meV/atom.","Fine-tuning on high-pressure configurations is a cheap and effective fix; two models dropped their 150 GPa energy error to roughly 31 meV/atom.","Ambient-pressure rankings do not predict pressure behavior; the most accurate ambient model degrades fastest under compression.","Training strategy matters: models exposed to high-pressure molecular dynamics or non-equilibrium denoising extrapolate markedly better.","The dataset doubles as a reusable benchmark for measuring progress in out-of-distribution generalization of future universal potentials."],"supporting_citations":[{"why":"The base DFT database whose structures, parameters, and PBE settings the pressure extension reuses, making the new dataset consistent with existing training data.","marker":"[29]"},{"why":"Introduces one of the earliest universal graph potentials, benchmarked here as the ambient-trained baseline with the strongest pressure degradation.","marker":"[31]"},{"why":"The OMat24 dataset that several benchmarked models were trained on; defines the ambient-dominated training distribution whose coverage is being tested.","marker":"[38]"},{"why":"Introduces MatterSim, one of the two models fine-tuned here; its inclusion of high-pressure molecular dynamics data explains its milder degradation.","marker":"[37]"},{"why":"Introduces eSEN-30M-OAM, the best untuned high-pressure performer and the other fine-tuned model; its denoising-based training is the proposed reason for its robustness.","marker":"[48]"},{"why":"Documents ORB-v3's removal of rattled and volume-scaled structures, the design choice the paper links to that model's fast pressure degradation.","marker":"[47]"},{"why":"Supplies the non-equilibrium denoising method used by eSEN, the training strategy the paper identifies as a factor in pressure robustness.","marker":"[49]"}],"fun_headline_variants":["Pressure stress reveals ML potentials' training blind spot","ML interatomic potentials wobble under extreme pressure—fix: data","High-pressure data gap degrades universal ML potentials","Why ML force fields fail at 150 GPa: it's the data","Targeted fine-tuning rescues ML potentials under pressure"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The benchmark and fine-tuning results exclude materials whose pressure relaxations did not converge in DFT (roughly one to three thousand of 190,000 per pressure), and the paper assumes that missing those cases does not bias the measured errors.","fun_headline_variants_meta":{"raw":{"variants":["Pressure stress reveals ML potentials' training blind spot","ML interatomic potentials wobble under extreme pressure—fix: data","High-pressure data gap degrades universal ML potentials","Why ML force fields fail at 150 GPa: it's the data","Targeted fine-tuning rescues ML potentials under pressure"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000329,"raw_usage":{"total_tokens":1636,"prompt_tokens":669,"completion_tokens":967,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":413,"completion_tokens_details":{"reasoning_tokens":885}},"tokens_in":413,"tokens_out":967,"duration_ms":10354,"temperature":1.0,"reasoning_tokens":885,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T16:43:34.864760+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the materials whose high-pressure relaxations did not converge, re-relax them with a more robust DFT protocol, and measure the energy and volume errors of the fine-tuned models on that held-out set; if the energy error is substantially larger than 31 meV/atom, the reported pressure recovery is incomplete for the hardest systems.","supporting_citations":[{"cited_title":"Schmidt, T","cited_arxiv_id":null,"evidence_quote":"The base DFT database whose structures, parameters, and PBE settings the pressure extension reuses, making the new dataset consistent with existing training data."},{"cited_title":"Chen and S","cited_arxiv_id":null,"evidence_quote":"Introduces one of the earliest universal graph potentials, benchmarked here as the ambient-trained baseline with the strongest pressure degradation."}],"review_version":1}