{"id":"cdaa596b-3f2f-4ef0-afaa-83ef12cb1f1f","arxiv_id":"2607.08025","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Precomputed hierarchical geometry tokens plus a decoupled decoder enable competitive neural PDE accuracy on industrial 3D meshes exceeding 10M nodes with linear memory scaling.","lead":"PGD-NO is a neural PDE solver that precomputes geometry tokens with a deterministic mesh-decomposition algorithm, then queries solutions with a lightweight attention decoder. This decoupling lets it train on meshes of tens of millions of nodes on single-node hardware where prior neural operators run out of memory.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reader's already-flagged weakest assumption.","rationale":"The strongest claim is architectural: pre-computed geometry tokens + decoupled querying remove the single-node VRAM wall while remaining competitive on five industrial 3D PDE tasks. That claim is supported by the memory-scaling plot (Fig. 1), the complexity analysis (Appendix A), the 12 M / 60 M CFD-VOL numbers (Table 2), and the head-to-head accuracy table (Table 1). The only premise that must hold for both accuracy and the practical utility of the scaling is that the classical sharp-edge decomposition produces usable tokens; the reader already named this premise and correctly graded the paper CONDITIONAL. No additional load-bearing flaw (e.g., an unstated assumption that would invalidate the linear-memory argument, a circular evaluation, or an irreproducible result) is present. The concrete test above simply quantifies how much of the observed accuracy is attributable to the particular classical tokens versus any sparse token set of similar cardinality—an experiment that would tighten the conditional acceptance without overturning it. Verdict therefore remains CONDITIONAL at high confidence.","tokens_in":18097,"tokens_out":577,"duration_ms":6700,"concrete_test":"Re-run the JEB and Aircraft columns of Table 1 after replacing Algorithm 1 tokens with an equal number of tokens obtained from a pure spectral (Laplace–Beltrami) clustering baseline (or from a learned segmentation network) while keeping the identical PGD-NO decoder; if relative L2 error rises by more than ~15 % relative to the reported surface-token numbers, the hand-crafted decomposition is confirmed as the accuracy bottleneck.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest assumption (that the fixed dihedral-angle hierarchical decomposition of §3.1 / Algorithm 1 yields sufficiently expressive geometry tokens for free-form industrial shapes) is correctly identified as load-bearing for both accuracy and the claimed scalability advantage. The paper itself supplies supporting evidence that the assumption is imperfect: limited gains on the organic JEB set (Table 1), clear sensitivity to token count (Table 4: 65.4 % \to 35.4 % as tokens rise from 32 to 256), and the explicit limitation note that smooth topologies remain harder. No stronger internal inconsistency or hidden failure mode appears. Memory-complexity claims (Appendix A) follow directly from the surface/volume decoupling once tokens exist; the 60 M-node results (Table 2) rest on spatial interpolation of the original CFD-VOL mesh, which is a secondary experimental caveat already noted by the reader rather than a contradiction of the architectural claim. Thus the central argument holds under the same condition the reader already isolated.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes PGD-NO, a neural operator that extracts a fixed set of “geometry tokens” from a surface mesh via a deterministic iterative decomposition (dihedral-angle sharp-edge detection, hierarchical threshold decay, size filtering and hop-based merge; Algorithm 1 / §3.1) and then uses those tokens as a compact latent context for multi-head attention layers that query PDE solutions at arbitrary volume or surface points. By moving geometric encoding off the GPU and decoupling it from solution querying, the architecture claims linear memory scaling that supports meshes of 10–60 M (and theoretically ~100 M) nodes—scales at which Transolver++ and related operators exhaust single-node VRAM—while remaining competitive in L2 field error and engineering QoIs on five industrial 3D benchmarks (Heat Sink, JEB, DrivAerNet++, Aircraft, and a new CFD-VOL set). Three decoder variants, depth/token ablations, memory/time tables, error maps and attention-based interpretability visualizations are provided.","tokens_in":18359,"tokens_out":1157,"duration_ms":12483,"significance":"If the central claim holds, the work removes a practical single-node memory ceiling that currently limits neural PDE solvers on industrial free-form geometries, enabling high-fidelity surrogate modeling on meshes that previously required multi-node communication or aggressive down-sampling. The deterministic token construction, public code/datasets, direct comparison against strong recent baselines (Transolver++, GINO, GNOT, AB-UPT, etc.), and attention-based nodal importance maps constitute concrete engineering and scientific contributions. The approach is therefore of clear interest to the neural-operator and computational-engineering communities, provided the expressivity of the fixed geometric tokens is adequately characterized.","major_comments":[{"comment":"§3.1 / Algorithm 1 and the free-parameter list (Θ, α, ρ, h, M=128): the load-bearing premise that a non-learnable dihedral-angle hierarchical partition yields sufficiently expressive tokens for arbitrary free-form industrial shapes is only partially supported. Table 1 shows only modest gains on the organic JEB set; Table 4 shows large sensitivity to token count (65.4 % → 35.4 % as M rises from 32 to 256); the Limitations paragraph itself notes that smooth topologies remain harder. A systematic sensitivity study (or a simple adaptive/learnable refinement of the tokens) is needed before the accuracy and scalability claims can be regarded as robust across the full range of industrial geometries the abstract targets.","section":null},{"comment":"Table 2 (60 M-node CFD-VOL results): the larger meshes are obtained by random spatial sampling and 3D interpolation of the original ~12 M mesh. While this is a legitimate stress test of memory scaling, it does not constitute an independent high-fidelity ground-truth dataset; residual interpolation error could inflate or mask true model error. The manuscript should either (a) quantify the interpolation error floor or (b) clearly label the 60 M experiment as a memory-scaling demonstration rather than a full accuracy claim at that resolution.","section":null},{"comment":"Experimental reporting (Tables 1–5, §4): all accuracy numbers are single-run point estimates with no error bars, multiple random seeds, or statistical tests. Given that several of the reported margins over Transolver++ are modest (e.g., Aircraft field 4.77 vs 4.98, CFD-VOL 12.2 vs 15.8), the absence of variability estimates weakens the claim of consistent superiority. At minimum, standard deviations over 3–5 seeds (or bootstrap intervals) for the primary L2 and QoI metrics should be supplied.","section":null}],"minor_comments":[{"comment":"Appendix A complexity analysis: the claimed O(4Nv) vs O(8Nv) reduction is plausible but would be clearer with an explicit operation count table matching the three decoder variants of Figure 2.","section":null},{"comment":"Figure 1 (right) and Table 5: peak-memory numbers are given for a 4-way GH200; stating the per-GPU footprint and whether model-parallel or data-parallel partitioning was used would improve reproducibility.","section":null},{"comment":"Notation: the segmentation matrix is denoted both S and M in Algorithm 1 and §3.2; a single consistent symbol would avoid confusion.","section":null},{"comment":"Related-work §2.2: classical mesh-segmentation literature is cited, yet the concrete relationship between the proposed dihedral-angle procedure and existing approximate-convex-decomposition or spectral methods is left implicit; a short paragraph situating Algorithm 1 would help.","section":null},{"comment":"Typos / polish: “processable mesh resolution”, “single node bottleneck” (hyphenation), and a few missing articles appear in the abstract and introduction; a light copy-edit pass is warranted.","section":null}],"recommendation":"major_revision","confidential_remarks":"The central architectural idea (pre-computed surface tokens + decoupled volume querying) is sound and the empirical package is already stronger than many neural-operator submissions. The three major points above are fixable within a revision cycle; I would expect a revised manuscript that addresses token expressivity, clarifies the 60 M experiment, and adds basic statistical reporting to be acceptable. No concerns about novelty disclosure or citation patterns."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The practical punchline is simple: by moving geometry encoding offline into a deterministic sharp-edge hierarchical decomposition and then decoding with attention, they get linear memory scaling that actually runs 10–60 M node industrial meshes on ordinary hardware while staying competitive with Transolver++ and the usual baselines.\n\nWhat is new is the clean decoupling. Geometry tokens come from an iterative dihedral-angle partition + merge (Algorithm 1), not from learned physics tokens or high-res grids. That lets volume queries be partitioned freely, so the memory bottleneck collapses to surface operations. They ship five industrial 3D benchmarks (including their own 12 M-node CFD-VOL), direct head-to-heads, depth/token ablations, memory/time tables, error maps, and attention-based interpretability. Code and data are promised. The scaling numbers and the surface/volume complexity argument in Appendix A hold up once the tokens exist.\n\nThe soft spot is exactly the one the reader flagged and the stress-test confirmed: the tokens are classical and hand-tuned (angle schedule, min-size, hop merge, keep top-128). On clean faceted shapes (Heat Sink) the gains are large; on organic free-form (JEB) they shrink and token-count sensitivity is obvious (65 % → 35 % error). The paper itself notes the limitation for smooth topologies. The 60 M results use spatial interpolation of the original mesh—secondary caveat, not a contradiction. Missing error bars and uneven baseline compute budgets are real but do not sink the central claim.\n\nThis is for people who actually need surrogates on industrial free-form meshes larger than a single GPU can hold. The math is straightforward, the data are real, the citations are appropriate. I would send it to referees; the architecture is concrete enough and the evidence sharp enough to deserve the time. Engage if you care about large-scale 3D operator learning; the decomposition is the piece to watch or improve.","headline":"Solid engineering fix for the single-node VRAM wall on million-scale 3D neural PDE solvers; competitive accuracy, real scaling, one load-bearing classical decomposition.","tokens_in":18954,"tokens_out":500,"would_cite":true,"duration_ms":5750,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Precomputing geometry tokens lets a neural PDE solver train on meshes of tens of millions of nodes without exhausting single-GPU memory.","keywords":["neural operator","geometry decomposition","million-scale meshes","PDE surrogate","physics attention","industrial CFD","memory scalability"],"falsifier":"Train PGD-NO and a strong baseline on a large library of deliberately smooth, organic geometries that contain almost no sharp dihedral edges; if the accuracy gap reverses or PGD-NO collapses while the baseline does not, the claim that precomputed tokens are universally sufficient is false.","tokens_in":18974,"feed_emoji":"🔧","tokens_out":926,"duration_ms":10266,"temperature":0.7,"pith_summary":"Neural PDE solvers promise to replace expensive engineering simulations, yet they hit a hard wall: the memory needed to encode a free-form 3-D mesh grows so fast that even the latest multi-GPU nodes cannot handle meshes beyond roughly ten million nodes. This paper claims the wall is artificial. Geometric features can be extracted once, offline, by a deterministic hierarchical decomposition that carves a mesh into a small set of \"geometry tokens\" based on sharp edges and connectivity. With those tokens in hand, a lightweight attention decoder only needs to query solution values at individual points; volume nodes never participate in the expensive encoding step. The resulting operator, PGD-NO, therefore scales linearly in memory and can train on meshes of 60–100 million nodes while matching or beating current state-of-the-art accuracy on five industrial benchmarks that range from heat sinks to full-vehicle aerodynamics. A sympathetic reader cares because the same architecture also yields node-wise attention maps that reveal which geometric parts drive the prediction, turning an otherwise black-box surrogate into an interpretable design tool.","feed_headline":"Precomputed tokens free neural PDE solvers from the 10M-node wall","feed_subtitle":"A deterministic geometry split lets models train on 60–100 million nodes while matching industrial accuracy","key_machinery":"Precomputed Geometry Decomposition (Algorithm 1): a hierarchical graph partition driven by successive dihedral-angle thresholds that yields a sparse segmentation matrix S mapping mesh nodes onto a few dozen geometry tokens; those tokens are then cross-attended with query-point features inside stacked Geometric Token Layers.","core_discovery":"A non-learnable, iterative geometry decomposition that produces a sparse set of geometry tokens is already rich enough to let a simple multi-head attention decoder learn high-fidelity PDE solutions on free-form industrial meshes. Because token extraction is moved entirely offline, solution querying can be partitioned arbitrarily across devices, removing the single-node VRAM ceiling that has limited every previous neural operator.","pith_inferences":["If the same decomposition is applied once to a multi-resolution CAD assembly, the resulting tokens could act as a shared latent vocabulary for transfer learning across entirely different physics (thermal, structural, fluid).","Replacing the fixed angle thresholds with a cheap, once-per-geometry learned refinement of the merge radii would test whether the current accuracy plateau on organic shapes is an artifact of the non-learnable front-end.","The linear memory profile suggests that real-time interactive design loops—where an engineer deforms a surface and immediately sees updated fields—become feasible even for full-vehicle meshes."],"forward_implications":["Meshes of 50–100 million nodes become routine training targets on a single high-memory GPU or a modest multi-GPU node.","Surface-derived tokens alone can drive accurate external-flow volume predictions, simplifying the pipeline for CFD surrogates.","Node-wise attention maps extracted from the Geometric Token Layers supply free, multi-scale design heuristics that highlight load paths and boundary-condition influence zones.","Because encoding is deterministic and mesh-format agnostic, the same token set can serve as a common interface for foundation models trained across heterogeneous CAD and meshing sources."],"fun_headline_variants":["Offline geometry tokens break 10M-node VRAM wall for PDE solvers","Iterative split frees neural operators from single-GPU mesh limits","Precomputed tokens enable high-fidelity PDE learning on 10M+ meshes","Geometry tokens decouple encoding so queries scale past one VRAM","Sparse offline tokens let attention decode industrial PDE solutions"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The fixed, hand-tuned sharp-edge hierarchy produces tokens that already capture every multi-scale geometric feature a free-form industrial shape needs; if it misses something, the decoder has no way to invent the missing information.","fun_headline_variants_meta":{"raw":{"variants":["Offline geometry tokens break 10M-node VRAM wall for PDE solvers","Iterative split frees neural operators from single-GPU mesh limits","Precomputed tokens enable high-fidelity PDE learning on 10M+ meshes","Geometry tokens decouple encoding so queries scale past one VRAM","Sparse offline tokens let attention decode industrial PDE solutions"]},"model":"grok-4.5","effort":"low","cost_usd":0.005776,"raw_usage":{"total_tokens":1512,"prompt_tokens":728,"num_sources_used":0,"completion_tokens":73,"cost_in_usd_ticks":57760000,"prompt_tokens_details":{"text_tokens":728,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":711,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":728,"tokens_out":73,"duration_ms":6637,"temperature":1.0,"reasoning_tokens":711,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T13:30:26.197004+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train PGD-NO and a strong baseline on a large library of deliberately smooth, organic geometries that contain almost no sharp dihedral edges; if the accuracy gap reverses or PGD-NO collapses while the baseline does not, the claim that precomputed tokens are universally sufficient is false.","supporting_citations":[],"review_version":1}