Files
MeltMind/artifacts/phase1-ml-survey-2026-11-17/physics-architecture-fulltext-extractions.md
T

22 KiB
Raw Blame History

Undermind full-text extraction receipt

These are tool-produced paper readings, not independent experimental reproductions. Exact-version caveats and corrections in the lane map take precedence over overbroad summary sentences here.

[Liu23]:

Audit conclusion

This paper does not evaluate injection-molding fill-time prediction, 3D domains, or a benchmark split defined by unique geometries. It supports a narrower claim: DAFNO can encode irregular 2D domains and, in one brittle-fracture example, handle an evolving crack topology. The reported benchmark counts are counts of samples; the paper does not separately report unique-geometry counts or establish a systematic held-out-geometry split. Its results therefore do not establish effectiveness for fill-time prediction on 200–600 unseen complex 3D shapes.

The manuscript is arXiv:2305.00478v2, dated 28 Oct 2023 (also identified as a NeurIPS 2023 paper).

Tasks, counts, and what is—and is not—held out

Paper task Reported data counts Geometry / OOD interpretation
Hyperelasticity (§4.1; Appendix B.1) Dataset split: 1,000 training, 200 validation, 200 test samples. Small-data experiments use 10, 100, or 1,000 training samples. The domain is a unit square with a randomly shaped central void; the output is stress. The paper does not state the number of unique void geometries in each split or say that the split is by geometry. Thus the 10/100/1,000 learning-curve points are sample counts, not documented unique-shape counts.
Airfoil design (§4.2; Appendix B.2) 1,000 training, 100 validation, 100 test samples. The data are variations of NACA-0012 airfoils, with Mach number as output. The number of unique airfoil geometries and whether test geometries are disjoint from training geometries are not reported. The additional irregular-grid experiment tests a mesh treatment, not a stated geometry-held-out split.
Crack propagation (§4.3; Appendix B.3) Training includes 450 consecutive snapshots from a simulation at traction 4 MPa, plus sinusoidal displacement/force examples. Main text says 4,096 sinusoidal instances, whereas Appendix B.3 says 2,048. The 450 are snapshots from a trajectory, not 450 independent geometries. The physical setup is one 2D plate with a pre-crack; the crack evolves and can branch. Tests use different tractions (2 and 6 MPa) and produce different crack patterns, but the paper gives no number of independent test shapes or test trajectories.
Pipe flow (Appendix B.4) 1,000 training samples are stated. No validation/test sample counts or unique-geometry counts are stated there. The paper characterizes the pipes as having simple geometry.

The paper’s OOD evidence is therefore not a controlled unseen-shape benchmark. Hyperelasticity and airfoil have train/validation/test sample splits but no documented unique-shape split. Crack propagation tests extrapolation to unseen loading magnitudes and resulting crack patterns within the same plate setup; it is not a set of independent initial 3D geometries.

Learning-curve and benchmark errors

Errors below are relative L2 errors, reported as mean ± standard deviation over five random-seed trials (§4). The paper does not report a fill-time metric.

Hyperelasticity, test error (Table 2, §4.1):

Training samples eDAFNO iDAFNO
10 16.446% ± 0.472% 16.669% ± 0.523%
100 4.247% ± 0.066% 4.214% ± 0.058%
1,000 1.094% ± 0.012% 1.207% ± 0.006%

For context at 1,000 samples, Table 2 reports test errors of 2.316% ± 0.283% for Geo-FNO, 3.193% ± 1.622% for F-FNO, 4.550% ± 0.062% for FNO with a sharp mask, and 1.415% ± 0.025% for FNO with smoothed χ. These are sample-size results, not a documented learning curve over independently counted unique geometries.

Table 2 reports the hyperelasticity test-error learning curve: the three columns are training-sample counts, and the rows compare DAFNO, baselines, and mask/smoothing ablations. The unique geometry counts behind those samples are not given.
[REGION: page=6, bbox=645,155,881,841]

Table 5 gives the fuller hyperelasticity results, including train and test errors for the same 10/100/1,000-sample regimes. It is useful for checking whether the learning-curve trend is accompanied by a train–test gap.
[REGION: page=16, bbox=69,155,446,846]

Airfoil design (Table 3, §4.2): eDAFNO test relative L2 error is 0.596% ± 0.005%; iDAFNO is 0.642% ± 0.020%; eDAFNO on irregular grids is 0.659% ± 0.007%. Comparators include Geo-FNO at 1.650% ± 0.175%, F-FNO at 0.794% ± 0.025%, and FNO with mask at 3.725% ± 0.108%. These scores are for the stated airfoil dataset split, whose geometry-level split is unspecified.

Table 3 reports airfoil train/test errors, including the irregular-grid eDAFNO result (0.659% ± 0.007% test error); it does not document a held-out-geometry protocol.
[REGION: page=8, bbox=69,170,249,829]

The irregular-grid figure illustrates mapping physical irregular meshes to computational uniform grids and compares absolute-error distributions. It supports mesh-handling evidence, not a geometry-OOD split.
[REGION: page=8, bbox=479,156,879,846]

Crack propagation: the paper gives visual crack-pattern comparisons and time-dependent relative-error plots, not a single aggregate test error for the 2 and 6 MPa scenarios (§4.3). Figure 7 compares relative error in χ and displacement against time for training at 4 MPa and tests at 2 and 6 MPa. No test trajectory count is reported.

Figure 6 compares ground-truth and eDAFNO crack patterns at 2, 4, and 6 MPa; Figure 7 plots time-wise geometry-mask and displacement errors for those loading cases. They show the loading/topology generalization claim, but not counts of independent test geometries or a scalar fill-time metric.
[REGION: page=10, bbox=72,155,344,849]
[REGION: page=10, bbox=324,156,614,842]

Geometry, mesh, conditioning, and boundary assumptions

  • Dimensionality: The examples are 2D. The brittle-fracture setup is a 40 mm × 40 mm plate represented in a periodic box. There is no 3D benchmark or 3D result.
  • Geometry representation: DAFNO embeds a characteristic function χ in the iterative Fourier layer; the paper proposes smoothing it near the boundary to reduce Gibbs artifacts (§3, Eqs. 4–7). For evolving cracks, χ is time-dependent (§4.3, Eq. 10). This permits topology change in the demonstrated 2D setting.
  • Mesh: Standard FFT requires a rectangular periodic box with uniformly meshed grids. The paper’s conclusion explicitly lists this as a limitation. The airfoil experiment uses a grid mapping to handle irregular grids; nonuniform FFT or grid deformation are proposed as ways to extend the approach.
  • Physical conditioning in the examples: Hyperelasticity uses a fixed bottom edge and a fixed tensile traction on the top, (t=[0,100]); airfoil flow uses fixed far-field conditions (including (M_\infty=0.8)) and no penetration on the airfoil. The paper does not demonstrate broad generalization across varying boundary/loading conditions for those benchmarks. In the crack example, the model takes displacement components (u_1,u_2) and χ as inputs and predicts internal-force components; the time integrator then advances the displacement.
  • Boundary-condition limitation: The conclusion says the applications focus on applying the same types of boundary conditions to changing domains (for example, all Dirichlet or all Neumann), and identifies transfer across different boundary-condition types as future work. This matters for fill-time problems where geometry, inlet/outlet, wall, and process conditions may vary together.
  • Geometry OOD: A changed crack pattern is shown, but the paper does not supply a geometry-disjoint train/test protocol or unique-shape counts for the benchmark splits. It does not establish performance on unseen complex 3D shapes.

Architecture versus physics loss; test-time optimization

DAFNO’s architectural contribution is geometry-aware gating of the convolutional integral with χ while retaining FFT computation (§3, Eq. 5 and Eq. 6). This is not the same as imposing the governing PDE as a soft residual loss.

The paper defines a supervised data objective in Eq. 2 and reports relative L2 errors. Its PDEs describe the simulated data-generating problems; the paper does not describe training DAFNO with a soft PDE-residual penalty. In the fracture rollout, the learned internal-force operator is inserted into the dynamic equation and a Velocity–Verlet integrator advances the state. That use of a physical solver at rollout is distinct from residual-based training.

DAFNO itself does not require optimizing a learned geometry latent against target solution labels: its geometry input is explicitly encoded by χ. The paper says Geo-FNO uses a deformation network trained with FNO, but it does not describe per-test-geometry latent optimization or say that test solution labels are used for such optimization. Accordingly, the paper provides no evidence that test geometry latents require target-solution-label optimization; it also does not give a detailed test-time adaptation protocol for Geo-FNO.

As a general methodological distinction—not a claim that this paper did it—optimizing against a PDE residual need not require target solution labels, but it does require access to the correct test-case PDE (including coefficients/forcing as applicable), domain geometry, and boundary/initial conditions, plus a way to evaluate the residual and boundary conditions on the target case. If those assumed equations or conditions are wrong or unavailable, the residual objective does not supply reliable supervision by itself.

Training, hardware, code, and limitations

  • The numerical-experiment section says models were trained for up to 500 epochs, with hyperparameters tuned on validation data; experiments were repeated over five random seeds (§4). Detailed batch size and actual memory use are not stated.
  • Appendix B.1 reports hyperelasticity per-epoch runtimes on an NVIDIA RTX A6000 GPU with 48 GB memory (Table 4): eDAFNO 2.00 s/epoch, iDAFNO 1.70 s/epoch. The paper states the card’s capacity, not the VRAM allocated or peak usage.
  • The authors’ code and data URL, given in the abstract, is https://github.com/ningliu-iga/DAFNO.
  • Stated limitations include standard FFT’s reliance on uniform grids, the experiments’ restriction to common boundary-condition types across changing domains, and future work on transfer across different boundary-condition types (Conclusion, §5). For the requested application, the further evidence gaps are that all presented tasks are 2D and there is no injection-molding task, fill-time target, 3D benchmark, or unique-geometry count/split.

Table 4 reports per-epoch hyperelasticity runtimes; the paper identifies the RTX A6000 48 GB card but does not report actual VRAM consumption.
[REGION: page=15, bbox=426,157,532,844]

Brief evidence quotes: the abstract says “with only one training crack simulation sample”; §4.3 specifies “450 snapshots” from the 4 MPa simulation; Appendix B.3 gives a different sinusoidal count from the main text; and the conclusion says the method focuses on “uniformly meshed grids.” These statements should not be interpreted as counts of independent geometries or as evidence of fill-time accuracy.


[Li21]:

Audit verdict

This paper does not benchmark injection-molding fill-time prediction, complex 3D geometries, or generalization to unseen geometries. Its experiments use 1D/2D PDEs on prescribed domains. The reported sample counts are counts of PDE inputs/instances (such as initial conditions or coefficient fields), not counts of distinct geometries. It reports no validation-set counts, mesh/topology OOD split, or fill-time metric. Thus it provides no empirical basis for claiming effectiveness on 200–600 independent molding shapes.

The manuscript is arXiv:2111.03794v4, dated 29 July 2023 (first page). It describes PINO as a Fourier neural operator (FNO) trained with data and/or PDE constraints.

Benchmarks and what their counts mean

Benchmark and task Reported data / test counts Resolution and reported results Geometry interpretation
Burgers, Section 4 / §4.1; Appendix A.1: learn initial condition → solution trajectory for 1D periodic Burgers, with viscosity (\nu=0.01). Appendix A.1 reports 1,000 initial conditions for training and error averaged over 200 test instances. The separate multiresolution experiment in §4.1 uses 800 low-resolution data instances and 800 PDE instances, with 200 test instances. In §4.1, data are (32\times25), PDE loss is imposed at (128\times100), and testing uses (32\times25,\ 64\times50,\ 128\times100). Table 1 reports data-only relative (L_2) errors of 0.32%, 3.32%, 3.76%, versus data+PDE 0.17%, 0.28%, 0.38% at those three resolutions. Appendix A.1 reports 0.38% test relative (L_2) error for its setup. One fixed 1D periodic domain; these are initial-condition/solution examples, not geometries.
Darcy flow, §4 / §4.1; Appendix A.2: learn coefficient field (a) → steady solution (u) on the unit square, with (f=1) and zero Dirichlet boundary. The coefficient is piecewise constant, representing two media. §4.1 reports 1,000 low-resolution data instances and 1,000 PDE instances, with 500 test instances for the multiresolution experiment. Appendix A.2 says 1,000 coefficient conditions for operator training; it does not give a separate test count for the Table 2 setup. Multiresolution data are (11\times11), PDE loss (61\times61), testing (11\times11,\ 61\times61,\ 211\times211). Table 1: data-only errors 5.41%, 9.01%, 9.46%; data+PDE 5.23%, 1.56%, 1.58%. Table 2 reports 6.97% for DeepONet with data, 1.22% for PINO with data, and 1.50% for PINO without data. One fixed 2D unit-square domain; coefficient fields vary, not domain geometry.
Kolmogorov flow, (Re=500,\ T=0.125), §4.1: learn initial vorticity → vorticity trajectory. §4.1 reports 800 low-resolution data instances and 2,200 PDE instances, with 200 test instances. Table 3 varies data count: 0, 800, 2,200, with 2,200 PDE instances in each case. Data (64\times64\times33); PDE loss (256\times256\times65); evaluation at those and (128\times128\times33). Table 3’s (256\times256\times65) relative (L_2) errors are 6.22% (0 data), 6.01% (800), and 5.04% (2,200). Table 1’s multiresolution data-only errors are 8.28%, 8.27%, 8.30%; data+PDE errors are 6.04%, 6.02%, 6.01% at the three listed test resolutions. Fixed periodic 2D torus; the varying instances are initial conditions, not shapes.
Kolmogorov flow, (Re=500,\ T=0.5), §4.1 / Appendix A.4: operator generalization across initial conditions. Appendix A.4 says 4,000 data points and a test set of 300 instances; Table 7 varies data counts (400, 4,000, or 0) and additional PDE instances. These are PDE samples/initial conditions, not geometries. Table 7 reports, at (128\times128\times65): with 400 data, errors are 33.32% (0 additional PDE), 31.74% (40k), 31.32% (160k); with 4k data, 25.15% (0), 24.15% (100k), 24.22% (400k); with 0 data and 100k PDE instances, 74.36%. These are relative solution errors; Table 7 also reports equation errors. Fixed periodic 2D domain; no geometry variation.
Long-time transient flow, §4.2 / Appendix A.3: Navier–Stokes flow build-up, (Re\sim20,\ T=50). 4,800 training data; Appendix A.3 reports 5 test instances. Counts refer to flow examples, not geometries. Last-time-step relative (L_2) errors: FNO 3.04%, PINO without instance-wise fine-tuning 2.87%, PINO with fine-tuning 1.84%. Fixed domain and forcing; no geometry split.
Instance-wise Kolmogorov fine-tuning, §4.2, Table 4. Table 4 compares PINN with PINO using 0 data/0 PDE instances, 0.4k data/0 PDE instances, and 0.4k data/160k PDE instances; comparison is averaged over 20 instances (Fig. 4 / Appendix A.4). Table 4 reports PINN 18.7%, 4577 s; PINO 0.9%, 608 s; 0.9%, 536 s; and 0.9%, 473 s, respectively. Instances of a fixed 2D flow problem, not geometry examples.
Lid-driven cavity, §4.2 and Appendix B.2: solve a single 2D cavity-flow problem with velocity–pressure formulation. No train/validation/test split or number of independent geometries is reported; this is an instance-wise solve, not an operator-learning geometry benchmark. Reports relative error 14.52% after 2 minutes at resolution (65\times65\times50). Fixed square cavity with specified wall velocities.

Learning-curve evidence: Table 3 is the clearest small-data comparison: the reported error changes from 6.22% with no data to 6.01% with 800 data instances and 5.04% with 2,200, with 2,200 PDE instances. Table 7 provides further data/PDE-instance sweeps, including the exact values above. These are learning curves over PDE instances, not over independent geometries. The manuscript does not state a validation-set size or give a train/validation/test split protocol sufficient to audit shape-level independence.

Table 1 compares data-only and data-plus-PDE errors across training and finer test resolutions; it supports the paper’s resolution-generalization claim, not generalization across geometries.
[REGION: page=12, bbox=660,96,869,907]

Table 3 shows the (Re=500,\ T=0.125) Kolmogorov-flow error versus number of data samples, with 2,200 PDE instances held fixed; it is relevant to small-data evidence but its counts are not shape counts.
[REGION: page=13, bbox=285,96,447,905]

Table 7 reports Kolmogorov-flow errors across data counts, additional PDE-instance counts, and resolutions; it provides the larger instance-count sweep, again on a fixed geometry.
[REGION: page=23, bbox=284,147,727,854]

Splits, geometry OOD, and physical inputs

  • Geometry/topology OOD: No experiment varies domain shape or topology between training and test. The reported problems use fixed domains: a 1D periodic interval, 2D unit square, 2D periodic torus, or a fixed square cavity. The paper does not report counts of unique geometries, mesh families, topology classes, or a geometry-held-out split. Resolution changes are tested in some experiments, but that is mesh/resolution extrapolation on the same domain, not geometry OOD.
  • Geometry representation: The FNO uses Fourier computation on grids. Appendix C describes “Fourier continuation” by zero-padding inputs and embedding the problem in a periodic domain; the paper demonstrates this for non-periodic settings. That is not a benchmark of arbitrary complex 3D geometry or varying mesh connectivity.
    Figure 10 illustrates zero-padding and smooth continuation of the FNO output; it explains the paper’s treatment of non-periodicity, not a test of unseen shapes.
    [REGION: page=26, bbox=82,96,303,906]
  • Physical conditioning: Examples condition on PDE inputs such as initial conditions (Burgers and Navier–Stokes), the Darcy coefficient field, and in some settings PDE parameters such as Reynolds number. For the cavity, the wall velocities are prescribed; for Darcy, (f=1) and zero Dirichlet conditions are specified. The paper does not study per-geometry molding boundary conditions, material/process inputs, or fill time.
  • BCs and PDE enforcement: PINO does not make the governing PDE or boundary conditions true by architecture. The PDE/equation loss is an optimization loss; §3.2 explicitly describes it as a “soft constraint.” The FNO architecture supplies an operator parameterization and efficient differentiation, while residuals and boundary/initial-condition terms are used in losses. Some boundary handling is problem-specific (for example, a Darcy boundary mollifier in Appendix A.2 or Fourier continuation in Appendix C).
  • What test-time optimization needs: In the instance-wise procedure (§3.2), PINO uses the learned operator output as an ansatz and optimizes against the PDE loss, optionally with an operator-anchor loss. For that residual-based optimization, the target instance must provide enough information to evaluate the assumed PDE residual and applicable boundary/initial constraints—such as the governing equation, coefficients/parameters, forcing, and conditions. A target solution label is not required by the PDE-only residual term, although data/anchor terms may also be used. This is not evidence that optimization works when the PDE or boundary specification is unknown or incorrect.

Training, hardware, and code

  • Appendix A says experiments use NVIDIA V100 and A100 GPUs. Burgers training is reported as about 20 minutes on one NVIDIA 3090 GPU for 500 epochs; Darcy training as about 1 hour on one V100; the Kolmogorov operator training as up to 2 days on one V100. The paper gives no VRAM figures.
  • Official code URL: No code repository URL is stated in the manuscript text supplied here, so an official URL cannot be verified from this paper alone.
  • Limitations acknowledged by the authors: The conclusion says the FNO/FFT implementation is “hard to extend to higher dimensional problems”; it also notes limitations in fine-tuning convergence and identifies transfer across geometries as future work. The paper’s demonstrated problems are 1D/2D, not 3D injection molding.

Bottom line: The paper provides evidence for combining PDE residuals with operator learning, for resolution extrapolation on fixed domains, and for instance-wise PDE-loss fine-tuning. It provides no fill-time results, no 3D complex-geometry benchmark, no 200–600-shape study, and no geometry-held-out evaluation. It therefore cannot establish fill-time effectiveness or unseen-shape generalization for the proposed use case.


[Mou26]: ❌ Error: Could not read PDF for [Mou26].

Region format: page is 1-indexed; bbox=top,left,bottom,right, normalized 0-1000 from top-left.