322 lines
38 KiB
Markdown
322 lines
38 KiB
Markdown
# 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.
|
||
|
||
[Li23]:
|
||
## Audit summary
|
||
|
||
**Li et al. (GINO) does not benchmark injection-molding fill time.** Its evaluated task is predicting a **steady RANS vehicle-surface pressure field** from car geometry; a separate Ahmed-body experiment also predicts wall shear stress to calculate drag. The paper supplies no evidence of fill-time accuracy or effectiveness on molding geometries.
|
||
|
||
### Benchmarks and counts
|
||
|
||
| Dataset | Task and declared data counts | Split and what the counts mean |
|
||
|---|---|---|
|
||
| **Ahmed-body** (§4.1; Appendix §7.3) | Predict surface pressure for Ahmed-body vehicle cases. The paper says it generates **551 shapes**, varying six shape parameters and inlet velocity; it reports **about 100k surface mesh points** and **7.2 million total mesh points** per simulation. | **500 training, 51 validation** (§4.1). These are the paper’s declared sample/shape counts. The paper does **not** separately report how many distinct geometric forms there are if cases with different inlet velocities but the same shape are counted as one geometry. No separate test split or trajectory/time-frame count is reported. |
|
||
| **Shape-Net Car** (§4.2) | Predict surface pressure for car geometries; each surface has **3.7k mesh points**. The source dataset has 889 instances, of which **611 water-tight shapes** are used. Inlet velocity is fixed at 20 m/s. | **500 training, 111 validation** (§4.2). These are declared shape counts; the paper does not report a distinct number of simulations separate from them. No trajectories or time frames are reported; the simulations are steady RANS. |
|
||
|
||
The tables label held-out performance “test error,” while the dataset descriptions call the held-out partitions “validation.” The paper does not describe an additional test set.
|
||
|
||
### Split and geometry/mesh OOD evidence
|
||
|
||
- Ahmed cases are selected using **Latin hypercube sampling** over the design parameters (Appendix §7.3), but the paper does not specify a split protocol that guarantees a particular kind of geometry OOD—such as holding out ranges of shape parameters—or quantify the degree of OOD in geometry.
|
||
- The abstract says the model is tested on new combinations of geometries and inlet velocities, but the paper does not give a separate count or precise construction of such combinations.
|
||
- The discretization studies are **not evidence of held-out-geometry generalization across mesh resolutions**: for latent-grid resolution and sampling-rate comparisons, training and testing use the same samples at each compared setting (§4.4). The super-resolution study trains on subsampled meshes and evaluates on the full mesh, but does not establish that those full-mesh evaluations are new geometries.
|
||
- The paper’s mathematical setup assumes smooth, closed, simply connected surfaces, separated from the outer domain boundary; the enclosed volume is assumed Lipschitz (§2). Its practical ShapeNet subset is explicitly restricted to **water-tight** shapes (§4.2). It does not report results for changing topology.
|
||
|
||
### Small-data evidence and reported errors
|
||
|
||
The paper states it trained with **500 data points** in the abstract, corresponding to the 500 training cases in each dataset description. It does **not** present a learning curve varying the number of training shapes, or an experiment showing performance at substantially fewer than 500 training cases. The subsampling experiments reduce mesh points, not the number of training geometries.
|
||
|
||
The reported metric is **normalized L2 training error** and **de-normalized L2 test error** (captions/notes for Tables 2–3):
|
||
|
||
- **Shape-Net, Table 2:** GINO decoder: **6.37% training, 7.12% test**. GINO encoder–decoder: 7.95%, 9.47%. FNO (interpolated): 9.65%, 9.42%.
|
||
- **Ahmed-body, Table 3:** GINO decoder: **9.34% training, 8.31% test**. GINO encoder–decoder: 9.36%, 9.01%. UNet (interpolated): 9.93%, 11.16%.
|
||
|
||
The paper separately reports a **26,000×** speed-up at **3% relative drag error** for its drag-coefficient comparison (§4.3; Appendix §7.2); this is not a fill-time result.
|
||
|
||
**Table 2** compares training/test errors for the Shape-Net car pressure task; its best listed GINO test error is 7.12%.
|
||
[REGION: page=9, bbox=82,251,276,750]
|
||
|
||
**Table 3** compares training/test errors for the Ahmed-body pressure task; the GINO decoder’s listed test error is 8.31%.
|
||
[REGION: page=10, bbox=82,251,259,750]
|
||
|
||
**Figure 4** shows discretization studies: latent-grid resolution, mesh sampling rates, and training on subsampled meshes followed by full-mesh evaluation. The plotted errors are broadly stable, but these experiments do not vary the number of training shapes.
|
||
[REGION: page=11, bbox=83,95,344,920]
|
||
|
||
**Table 6** gives the super-resolution numbers for training on subsampled meshes and testing on the full mesh: GINO encoder–decoder errors are 8.8%, 9.4%, 9.4%, and 9.7% at sampling rates 1/2, 1/4, 1/6, and 1/8, respectively.
|
||
[REGION: page=20, bbox=165,234,306,767]
|
||
|
||
### Inputs, conditioning, and test-time latent fitting
|
||
|
||
GINO’s inputs are a **surface point cloud and signed distance function (SDF)**; its output is queried on the geometry surface (§3; Figure 1). For variable inlet velocity, the paper embeds the scalar velocity using Fourier features and uses a learned MLP to produce adaptive instance-normalization scale and shift parameters (§3.3). The Ahmed dataset varies inlet velocity from 10 to 70 m/s; ShapeNet fixes it at 20 m/s.
|
||
|
||
The paper describes end-to-end training of the model, followed by prediction on query points. It does **not** describe optimizing a latent representation for each test geometry against that geometry’s target-solution labels. Thus, **test-time target-label latent optimization is not part of the reported method**. The paper does mention computing the SDF as preprocessing for inference (Appendix §7.2).
|
||
|
||
**Figure 1** diagrams the GINO pipeline: a GNO maps irregular geometry data to a regular latent grid, FNO layers process that grid, and another GNO projects predictions back to geometry query points.
|
||
[REGION: page=2, bbox=72,96,396,907]
|
||
|
||
### Training, hardware, code, and limitations
|
||
|
||
- Benchmarks use **100 epochs**, Adam, and a step learning-rate schedule; the learning rate is halved at epoch 50 (Appendix §7.1). All benchmark models run on **one NVIDIA V100 GPU** (§4).
|
||
- The paper does **not** report GINO training peak VRAM. It says the MeshGraphNet comparator has about 280k edges, saturating a single 32-GB V100 (Appendix §7.1). For graph construction, the authors report 6 GB for the hash-based method versus 24 GB for pairwise construction at 10k points (Appendix §7.1); those figures are not GINO training-peak measurements.
|
||
- **Official code URL:** no code repository URL is given in the supplied paper text. I cannot identify an official code URL from this paper alone.
|
||
- The conclusion states that the trained surrogate is limited to a **specific category of shapes** and depends on training-data quality; it also notes that obtaining large datasets for more complex shapes is difficult (§6). The paper does not test injection molding, transient fill-front evolution, or fill-time prediction.
|
||
|
||
**Paper:** Zongyi Li et al., “Geometry-Informed Neural Operator for Large-Scale 3D PDEs,” arXiv:2309.00583v1, 1 September 2023.
|
||
|
||
---
|
||
|
||
[Li22]:
|
||
## Bottom line for this use case
|
||
|
||
This paper **does not benchmark injection-molding fill time**, and it does **not establish performance on unseen, complex 3D volume geometries**. Its experiments are mostly 2D problems; the sphere example is a 2D surface embedded in 3D, not a 3D volume. The paper reports no small-data learning curve, no explicit validation-set counts, and no deliberately geometry-OOD train/test split. Thus, its results do not support a claim that Geo-FNO will predict fill time accurately from only 200–600 independent 3D shapes.
|
||
|
||
The method is nevertheless relevant as a proposed way to handle geometry: it maps an irregular domain to a uniform latent computational mesh, applies FNO there, and maps back (Section 3; Figure 1). The paper’s results are evidence for its particular benchmark settings, not for the requested application.
|
||
|
||
## Benchmarks and what the counts mean
|
||
|
||
The paper says it generally uses a 500-epoch training schedule, but gives **no validation-set counts** (Section 4). The counts below are reported training/test *data* counts. Where geometry varies, the paper describes geometry parameters being sampled, but it does not separately tabulate or audit the number of unique shapes. Mesh points, elements, and time frames are not additional training examples.
|
||
|
||
| Task and location | Reported training/test data | What varies and what is counted separately |
|
||
|---|---:|---|
|
||
| **Hyper-elasticity, Section 4.1; Table 1** | 1,000 / 200 | Void geometry is randomized; the target is the stress field. Inputs are point clouds of around 1,000 points; the FEM reference uses about 100 quadratic quadrilateral elements. The table reports a mesh size of 972 for Geo-FNO (learned). The paper does not separately certify unique-geometry counts. |
|
||
| **Plastic forging, Section 4.2; Table 2** | 900 / 80 | The die shape is generated from sampled spline values. Each sample is a time-dependent case: the data use a 101×31 structured mesh and **20 time steps**. Those time steps are not 20 independent shapes or samples. The target includes the evolving mesh/grid and displacement. |
|
||
| **Advection on the unit sphere, Section 4.3; Table 3** | 1,000 / 200 | The randomized quantities are the initial cos-bell’s center, radius, and height; the domain remains the unit sphere. The output is the solution at \(t=256/3\). Inputs are point clouds of around 8,000 points. This is not a benchmark of unseen geometries. |
|
||
| **Airfoil Euler flow, Section 4.4; Table 2** | 1,000 / 200 | Airfoil shapes vary via design-element control-node displacements, drawn from \(U[-0.05,0.05]\). The data use a C-grid of about 200×50 quadrilateral elements; the target is Mach number on mesh points. Unique-shape counts are not separately reported. |
|
||
| **Pipe Navier–Stokes flow, Section 4.5; Table 2** | 1,000 / 200 | Pipe centerlines vary via five control nodes (positions \(U[-2,2]\), slopes \(U[-1,1]\)). The data use a 129×129 mesh; the target is horizontal velocity. Unique-shape counts are not separately reported. |
|
||
|
||
These are **not 200–600-shape learning-curve experiments**: the geometry-varying tasks use 900–1,000 training data and 80–200 test data. The paper does not report how many of those are unique geometries as a separately checked statistic.
|
||
|
||
## Splits, OOD design, and metrics
|
||
|
||
- **Split design:** Section 2.1 describes the general dataset as i.i.d. samples from a distribution, and Sections 4.1–4.5 describe randomized benchmark data. The paper does not specify a geometry-held-out split procedure, an OOD geometry split, or validation-set construction. It therefore does not demonstrate performance on geometries outside the training geometry distribution.
|
||
- **Mesh-resolution transfer:** Section 4.4 does report a resolution-transfer test: train on a 56×51 airfoil mesh and evaluate on 56×51, 111×51, and 221×51 meshes. The reported test errors are **0.0147, 0.0329, and 0.0428**, respectively. This is evidence about changing discretization, not a designed geometry-OOD test.
|
||
- **Metric:** The paper uses relative \(L^2\) error for training and testing (Section 4, with the loss defined in Equations 2–3). No fill-time-specific error measure, such as absolute or relative error in a scalar fill time, is reported.
|
||
- **Small-data evidence:** No learning curve versus training-set size is reported. In particular, there is no result at 200–600 training geometries and no estimate of the data needed for a target accuracy.
|
||
|
||
Reported relative \(L^2\) training/test errors include:
|
||
|
||
| Benchmark | Geo-FNO training / test error | Comparators reported in the paper |
|
||
|---|---:|---|
|
||
| Elasticity, Table 1 | Learned deformation: **0.0125 / 0.0229** | GraphNO: 0.1271 / 0.1260; DeepONet: 0.0528 / 0.0965; R-mesh: 0.0308 / 0.0536; O-mesh: 0.0344 / 0.0363; FNO interpolation: 0.0314 / 0.0508; UNet interpolation: 0.0089 / 0.0531 |
|
||
| Airfoil, Table 2 | **0.0134 / 0.0138** | FNO interpolation: 0.0287 / 0.0421; UNet interpolation: 0.0039 / 0.0519 |
|
||
| Pipe, Table 2 | **0.0047 / 0.0067** | FNO interpolation: 0.0083 / 0.0151; UNet interpolation: 0.0109 / 0.0182 |
|
||
| Plasticity, Table 2 | **0.0071 / 0.0074** | Interpolation methods are marked inapplicable because the mesh is also a target output |
|
||
| Sphere advection, Table 3 | Geo-FNO2D: **0.0119 / 0.0332** | FNO2D: 0.0119 / 0.0381; FNO3D: \(>100\%\) / \(>100\%\); UNet2D: 0.1964 / 0.3132; DeepONet: 0.1113 / 0.1599 |
|
||
|
||
For speed, Section 4.4 says Geo-FNO is **at least \(10^4\)× faster at the same accuracy** in the airfoil cost–accuracy study; the abstract says “up to \(10^5\)” acceleration. Those are paper-level claims for the reported airfoil setup, not a speed estimate for 3D molding simulations.
|
||
|
||
## Geometry, conditioning, and the learned deformation
|
||
|
||
- **Geometry/topology assumptions:** Section 3 formulates a deformation between a physical domain and a uniform computational domain, with the ideal map a diffeomorphism. For shapes that do not admit that map directly, the paper describes Fourier continuation (Section 3.4) or domain decomposition (Section 3.5). The experiments are 2D: domains with holes, structured airfoil/pipe meshes, and the sphere surface. The sphere example uses two hemispheric subdomains; it is not a test on a complex 3D volume.
|
||
- **Inputs and physical conditions:** Section 2 lists possible geometry inputs—meshes/point clouds, functions such as SDFs, or design parameters. In the experiments, geometry or mesh coordinates and design parameters are supplied in the forms described above. Section 2 assumes initial and boundary conditions are fixed; the examples specify fixed physical/boundary conditions rather than testing broad variation in process conditions. The paper does not test injection pressure, melt temperature, mold temperature, material-property variation, or fill-time targets.
|
||
- **Learned map:** In Section 3.3, the inverse coordinate map is parameterized as \(\phi_a^{-1}(x,a)=x+f(x,a)\), where \(f\) is a feed-forward network; the paper specifies a three-layer network of width 32 with sinusoidal features. It is trained jointly with the solution operator using the supervised data loss.
|
||
- **Do test geometries require target-label optimization?** **No such test-time optimization is described.** For a geometry represented in the model’s input format, the trained deformation network is evaluated from coordinates and geometry input \(a\); its learned parameters were fitted during training. The paper does not describe optimizing a separate per-test-geometry latent against that geometry’s solution labels. However, it also does not demonstrate generalization to truly OOD 3D geometries, so the absence of test-label fitting should not be read as evidence of OOD accuracy.
|
||
|
||
## Compute, code, and limitations for the proposed adaptation
|
||
|
||
The paper states that experiments were run on **one Nvidia 3090 GPU**; the usual training schedule was **500 epochs**, initial learning rate **0.001**, halved every 100 epochs, using Adam (Section 4). It reports **no VRAM**, batch size, or general end-to-end training wall time. Table 1 gives per-epoch times for elasticity (learned Geo-FNO: 1 s/epoch; R-mesh: 0.6 s; O-mesh: 0.5 s), but these are not a 3D-model training estimate. Official code URL: **https://github.com/neuraloperator/Geo-FNO**.
|
||
|
||
For this use case, the main evidence gaps are the lack of 3D-volume fill-time experiments, geometry-OOD evaluation, 200–600-shape learning curves, a fill-time metric, and varied molding-process inputs. The paper’s broad framework claims do not fill those gaps.
|
||
|
||
### Figures and tables to inspect
|
||
|
||
Table 1 reports the elasticity benchmark’s mesh size, parameter count, per-epoch training time, and relative \(L^2\) train/test errors; it shows the learned Geo-FNO test error of 0.0229 alongside the interpolation and mesh-free baselines.
|
||
[REGION: page=14, bbox=97,127,374,967]
|
||
|
||
Table 2 gives train/test relative \(L^2\) errors for airfoil, pipe, and plasticity, while Table 3 gives the sphere-advection model comparison, including the \(>100\%\) FNO3D errors.
|
||
[REGION: page=18, bbox=94,154,284,848]
|
||
[REGION: page=18, bbox=258,127,446,872]
|
||
|
||
Figure 3 plots relative \(L^2\) error against runtime for Geo-FNO/UNet and numerical solvers on the airfoil problem; it is the graphic associated with the paper’s cost–accuracy and speedup claim.
|
||
[REGION: page=11, bbox=468,86,844,876]
|
||
|
||
Figure 1 diagrams the physical-to-latent deformation, FNO computation on a uniform latent mesh, and mapping back to the physical domain; it is relevant to how the method represents geometry, but does not establish 3D OOD performance.
|
||
[REGION: page=4, bbox=97,126,476,875]
|
||
|
||
---
|
||
|
||
[Wu24b]:
|
||
## Audit summary
|
||
|
||
**This paper does not test injection-molding fill-time prediction.** Its experiments are supervised PDE surrogates for solid mechanics, fluid flow, and porous-medium flow. It reports no fill-time task, injection-molding data, or validation of fill-time accuracy. The evidence below therefore describes what the paper does test; it does **not** establish fill-time effectiveness.
|
||
|
||
### Tasks and sample counts
|
||
|
||
The dataset table reports **training and test samples/cases**, not a separate count of unique geometries in every dataset. It reports no validation split. Counts below are not mesh-node counts; those are listed separately.
|
||
|
||
| Benchmark | Task and inputs → outputs | Train / test cases | Geometry-count interpretation and mesh |
|
||
|---|---|---:|---|
|
||
| Elasticity | Material structure/point coordinates → inner stress | 1,000 / 200 | Samples use different structures; the paper does not separately certify a unique-geometry count. Point cloud, 972 points. |
|
||
| Plasticity | Impact from an arbitrary-shaped die → future material deformation | 900 / 80 | Samples use different die shapes; these are cases, with a 20-step predicted output—not 20 separate geometries. Structured mesh, 3,131 points. |
|
||
| Airfoil | Airfoil shape → Mach number field | 1,000 / 200 | Samples are different airfoil designs; no separate unique-shape count beyond the cases is given. Structured mesh, 11,271 points. |
|
||
| Pipe | Pipe structure → horizontal fluid velocity | 1,000 / 200 | Samples use different pipe shapes. Structured mesh, 16,641 points. |
|
||
| Navier–Stokes | Past fluid velocity → future fluid velocity | 1,000 / 200 | These are fluid/initial-condition cases, **not varying geometries**. Each example uses 10 past and predicts 10 future steps. Regular grid, 4,096 points. |
|
||
| Darcy | Porous-medium structure → fluid pressure | 1,000 / 200 | Samples use different medium structures; separate unique-geometry count is not stated. Regular grid, 7,225 points in the main experiments. |
|
||
| Shape-Net Car | Car geometry → surrounding velocity and surface pressure; derive drag coefficient | **789 / 100**, from 889 samples | The paper describes the samples as having different car shapes. Unstructured hybrid volume/surface mesh, 32,186 points. Simulated at 72 km/h. |
|
||
| AirfRANS | Airfoil case → velocity and pressure fields; evaluate lift coefficient | **800 / 200**, from 1,000 simulations | Cases vary airfoil shape, Reynolds number, and angle of attack; **the number of unique airfoil geometries is not reported** separately from simulations/conditions. Unstructured mesh, 32,000 points. |
|
||
|
||
These details are in §4 and Appendix B.1, especially Table 7. For example, the paper describes Shape-Net Car as “889 samples with different car shapes,” while AirfRANS is described as varying shape, Reynolds number, and angle of attack. The sample counts should not be read as a confirmed unique-geometry count in cases where conditions may vary along with geometry.
|
||
|
||
[REGION: page=14, bbox=685,69,923,906]
|
||
|
||
### Splits, geometry OOD, and mesh-resolution evidence
|
||
|
||
- **Ordinary train/test splits:** Appendix B.1 gives the counts above, but does not specify a validation set or provide a detailed geometry-grouped splitting protocol. Thus, the paper does not establish that all ordinary test results are from a rigorously defined “unseen geometry” split, beyond dataset descriptions such as “different structures” or “different car shapes.”
|
||
- **Explicit OOD test:** §4.2 and Appendix E.3 test AirfRANS with **OOD Reynolds numbers or OOD angles of attack**, not a clearly specified held-out geometry-topology split. For Reynolds OOD, training uses 500 samples in \([3\times10^6,5\times10^6]\), and testing uses 500 in \([2\times10^6,3\times10^6]\cup[5\times10^6,6\times10^6]\). For angle OOD, training uses 800 samples in \([-2.5^\circ,12.5^\circ]\), and testing uses 200 in \([-5^\circ,-2.5^\circ]\cup[12.5^\circ,15^\circ]\) (Table 12). The paper does not give a separate unique-geometry count for these OOD subsets.
|
||
- **Mesh/resolution variation:** Appendix D visualizes learned slices on randomly resampled meshes retaining 50–80% of original points; this is qualitative visualization, not a reported fill-time or geometry-OOD benchmark. Figure 5 also shows an Elasticity mesh sampled at 50%. For Darcy, the model is retrained at different resolutions; this is resolution scaling, not a held-out topology test. Table 10 reports Transolver relative \(L_2\) values from 484 to 168,921 mesh points (e.g., 0.0059 at 7,225 points and 0.0060 at 168,921).
|
||
- **Topology:** The method is presented for general/irregular geometries, and for unstructured meshes uses a point-wise linear projection (Appendix B.3). It does not report an explicit topology-class split or an injection-mold geometry/topology evaluation. In its Lagrangian experiment, the authors note that scattered particles lack essential graph information and add a GNN in parallel (§E.4).
|
||
|
||
[REGION: page=9, bbox=317,69,513,497]
|
||
|
||
The three panels plot Darcy relative \(L_2\) versus input resolution, training-sample count, and model depth. The data panel increases the training count from 1,000 to 5,000 and shows error decreasing; it is the paper’s clearest learning-curve evidence, but it is Darcy only and does not test 200–600 geometries.
|
||
|
||
### Small-data evidence and reported metrics
|
||
|
||
The learning-curve experiment is **not** a 200–600-geometry study. Figure 8 varies Darcy training samples from 1,000 to 5,000; the plotted error is approximately \(5.9\times10^{-3}\) at 1,000 and \(4.1\times10^{-3}\) at 5,000 by visual reading. **Those values are approximate readings from the plot, not a numerical table.** The paper gives no corresponding learning curve for Shape-Net Car or AirfRANS, and no experiment with only 200–600 distinct geometries.
|
||
|
||
For field prediction, the metric is relative \(L_2=\|u-\hat u\|/\|u\|\) (Eq. 14). The practical-design table reports:
|
||
|
||
- **Shape-Net Car:** volume 0.0207, surface 0.0745, drag-coefficient relative \(L_2\) 0.0103, Spearman \(\rho_D=0.9935\).
|
||
- **AirfRANS:** volume 0.0037, surface 0.0142, lift-coefficient relative \(L_2\) 0.1030, Spearman \(\rho_L=0.9978\).
|
||
|
||
These are the paper’s metrics on its stated tasks—not fill-time errors. Table 3 contains the full comparison. Table 2 reports relative \(L_2\) on the six standard benchmarks; Transolver’s values are Elasticity 0.0064, Plasticity 0.0012, Airfoil 0.0053, Pipe 0.0033, Navier–Stokes 0.0900, and Darcy 0.0057.
|
||
|
||
[REGION: page=7, bbox=75,69,414,911]
|
||
|
||
### Conditioning, inputs, and training
|
||
|
||
The general setup embeds geometry \(g\) and, when available, observed physical quantities \(u\) (§3, “Problem setup”; Eq. 6). Some tasks use geometry/structure alone, while Navier–Stokes uses past velocity. For Shape-Net Car, Appendix B.1 says the input preprocessing combines mesh-point position, signed-distance function, and normal vector; the model predicts velocity and pressure. The paper does **not** give a general boundary-condition input specification for these design tasks. In particular, although AirfRANS cases vary Reynolds number and angle of attack, the paper’s dataset summary lists “structure” as the input; it does not clearly document those parameters as separate model-conditioning channels. It does not describe injection-molding process variables or fill-time supervision.
|
||
|
||
The method learns slice weights from point features, aggregates physics-aware tokens, applies attention among tokens, then broadcasts results back to mesh points (Eqs. 1–4). The paper describes ordinary model training followed by inference; **it does not describe optimizing a test-geometry latent against target-solution labels**. The target-specific latent/solution-label optimization procedure asked about is not reported. As described, slice assignments are produced by the learned projection from the input features rather than by a stated per-test-label fitting step.
|
||
|
||
Implementation details (Appendix B.3; Table 8): experiments use one NVIDIA A100 GPU and are repeated three times. The Shape-Net Car configuration is 200 epochs, Adam, learning rate \(10^{-3}\), batch size 1, 8 layers, 8 heads, 256 channels, and 32 slices. AirfRANS uses 400 epochs and the same optimizer, learning rate, batch size, layers, heads, channels, and slice count. The general A100 configuration is stated, but task-specific training VRAM is not. In the efficiency analysis, Table 16 reports Transolver at 32,768 points with 2.36 GB GPU memory and 98.552 s/epoch; these are efficiency measurements, not a general training-memory guarantee.
|
||
|
||
### Limits relevant to the proposed adaptation
|
||
|
||
- No injection-molding benchmark, fill-time target, or fill-time error metric is reported.
|
||
- The paper does not establish a 200–600 **unique-geometry** learning curve, nor report unique-geometry counts for AirfRANS.
|
||
- Its explicit OOD study changes Reynolds number or angle of attack; it does not establish topology OOD performance.
|
||
- Resampled-mesh visualizations and Darcy resolution scaling support flexibility to point counts/discretization, but are not evidence of accuracy on unseen mold geometries or fill-time prediction.
|
||
- The method is a supervised surrogate in these experiments; the paper does not report a fill-time-specific physical constraint or boundary-condition treatment.
|
||
- The authors’ Lagrangian experiment is preliminary and adds Transolver alongside a GNN because graph information is missing (§E.4); it is not evidence for injection molding.
|
||
|
||
**Official code URL:** https://github.com/thuml/Transolver (given in the abstract).
|
||
|
||
[REGION: page=17, bbox=69,70,313,906]
|
||
|
||
This table lists shared training settings and per-benchmark model configurations; it supports the optimizer, epoch, batch-size, layer, channel, and slice-count details above.
|
||
|
||
---
|
||
|
||
[Mou25]:
|
||
## Audit summary
|
||
|
||
**The paper does not evaluate injection-molding fill-time prediction, 3D geometries, or a small-data regime of roughly 200–600 independent shapes.** Its results therefore do not establish fill-time effectiveness. The closest evidence is for two *steady* datasets with varying 2D geometries (airfoils and perforated elastic specimens), plus a learning-curve/transfer study on Navier–Stokes trajectories—not a geometry-count learning curve.
|
||
|
||
### Benchmarks and sample counts
|
||
|
||
The paper reports 13 tasks: 10 time-dependent and 3 time-independent. Its standard protocol is **1,024 training samples, 128 validation samples, and 256 test samples**, except AF, which uses **2,048 training samples**. For Table 2, the authors say test errors were calculated on **240** test samples instead. These are reported as samples or trajectories—not as counts of unique geometries. The paper does **not** enumerate how many distinct shapes occur in AF or Elasticity, nor establish a geometry-disjoint train/validation/test split. For fixed-domain tasks, the domain shape is fixed while the PDE inputs/solutions vary; for AF and Elasticity, the geometry varies, but the number of unique shapes is unstated. (Sections 3, D, E; Tables 1, 2, D.1.)
|
||
|
||
| Task | PDE / setup and domain | Geometry status and temporal status |
|
||
|---|---|---|
|
||
| Heat-L-Sines | Heat equation; L-shaped domain; Dirichlet BC | Fixed domain; time-dependent |
|
||
| Wave-C-Sines | Wave equation; circular domain; Dirichlet BC | Fixed domain; time-dependent |
|
||
| NS-Gauss | Incompressible Navier–Stokes; periodic square | Fixed domain; time-dependent |
|
||
| NS-PwC | Incompressible Navier–Stokes; periodic square | Fixed domain; time-dependent |
|
||
| NS-SL | Incompressible Navier–Stokes; periodic square | Fixed domain; time-dependent |
|
||
| NS-SVS | Incompressible Navier–Stokes; periodic square | Fixed domain; time-dependent |
|
||
| CE-Gauss | Compressible Euler; periodic square | Fixed domain; time-dependent |
|
||
| CE-RP | Compressible Euler; periodic square, four-quadrant Riemann setup | Fixed domain; time-dependent |
|
||
| ACE | Allen–Cahn; periodic square | Fixed domain; time-dependent |
|
||
| Wave-Layer | Wave equation in layered medium; absorbing square | Fixed domain; time-dependent |
|
||
| Poisson-Gauss | Poisson equation; square; Dirichlet BC | Fixed domain; steady |
|
||
| AF | Transonic compressible Euler flow over airfoils | **Varying airfoil geometries**; steady |
|
||
| Elasticity | Hyper-elastic force-balance problem on square specimens with holes | **Varying hole geometries**; steady |
|
||
|
||
The eight time-dependent PDEGym tasks use 21 snapshots per trajectory, except ACE and Wave-Layer, which use 20. In the usual evaluation setup, snapshots at even indices through \(t_{14}\) are used for training; later times are reserved for extrapolation tests and earlier even-indexed times for interpolation tests. Those are **time frames from trajectories**, not additional independent geometries. PDEGym grids are originally \(128\times128\); the paper’s random-point-cloud version retains 9,216 randomly selected coordinates. (Sections C.2, D; Table D.1.)
|
||
|
||
**Geometry/mesh OOD evidence:** Section F.8 tests 128 in-distribution and 128 out-of-distribution samples on each of three other datasets. The described shifts are: more complex airfoil geometries for Compressible Euler, different characteristic length scales for Darcy Flow, and a substantially shifted Gaussian-perturbation center for Smooth Transport. No number of unique geometries is given, and the section does not specify a geometry-held-out split protocol. Its OOD errors are Darcy Flow 2.57% ID / 3.18% OOD; Smooth Transport 0.72% / 6.50%; Compressible Euler 0.51% / 0.62% (Table F.3). This is not a fill-time benchmark.
|
||
|
||
**Resolution variation is not geometry OOD.** The paper varies spatial resolution for NS-PwC, training at \(64^2\) and testing at other resolutions; the ablation reports improvement from edge masking, particularly below training resolution (Figure 3a). Its random point clouds are derived from the same square domain, not unseen shapes.
|
||
|
||
### Exact reported errors
|
||
|
||
The metric in Tables 1 and 2 is **median relative \(L_1\) error (%)** (Eq. 33). Before computing the error, estimates and targets are normalized with global equation-wide means and standard deviations; errors are aggregated over variable groups and then reported as the median over samples (Section E). The table below gives the reported RIGNO-18 and RIGNO-12 results and the lowest error among the listed non-RIGNO baselines.
|
||
|
||
| Point-cloud benchmark (Table 1) | RIGNO-18 | RIGNO-12 | Best listed non-RIGNO baseline |
|
||
|---|---:|---:|---:|
|
||
| Heat-L-Sines | 0.04 | 0.05 | Geo-FNO 0.15 |
|
||
| Wave-C-Sines | 5.35 | 6.25 | FNO DSE 5.52 |
|
||
| NS-Gauss | 2.29 | 3.80 | GINO 13.1 |
|
||
| NS-PwC | 1.58 | 2.03 | GINO 5.85 |
|
||
| NS-SL | 1.28 | 1.91 | GINO 4.48 |
|
||
| NS-SVS | 0.56 | 0.73 | GINO 1.19 |
|
||
| CE-Gauss | 6.90 | 7.44 | GINO 25.1 |
|
||
| CE-RP | 3.98 | 4.92 | GINO 12.3 |
|
||
| ACE | 0.01 | 0.01 | MeshGraphNet 0.09 |
|
||
| Wave-Layer | 6.77 | 9.01 | Geo-FNO 11.1 |
|
||
| Poisson-Gauss | 2.26 | 2.52 | FNO DSE 2.27 |
|
||
| AF | 1.00 | 1.09 | FNO DSE 1.99 |
|
||
| Elasticity | 4.31 | 4.63 | GINO 4.38 |
|
||
|
||
| Cartesian-grid benchmark (Table 2) | RIGNO-18 | RIGNO-12 | Best listed non-RIGNO baseline |
|
||
|---|---:|---:|---:|
|
||
| NS-Gauss | 2.74 | 3.78 | scOT 2.92 |
|
||
| NS-PwC | 1.12 | 1.82 | CNO 5.03 |
|
||
| NS-SL | 1.13 | 1.82 | CNO 2.12 |
|
||
| NS-SVS | 0.56 | 0.75 | CNO 0.70 |
|
||
| CE-Gauss | 5.47 | 7.56 | scOT 9.44 |
|
||
| CE-RP | 3.49 | 4.43 | scOT 9.74 |
|
||
| ACE | 0.01 | 0.01 | scOT 0.21 |
|
||
| Wave-Layer | 6.75 | 8.97 | CNO 8.28 |
|
||
| Poisson-Gauss | 1.80 | 2.44 | scOT 0.68 |
|
||
|
||
The paper reports its time-dependent benchmark errors as the lowest among three rollout strategies (Tables 1–2; Section C.2). Table 1’s AF and Elasticity results are relevant to variable 2D shapes, but they are steady-state tasks, not fill-time prediction.
|
||
|
||
### Small-data evidence
|
||
|
||
Figure 2 plots RIGNO-12 errors at training-set sizes 256, 512, and 1,024 for selected datasets and describes a consistent reduction in error. The paper does **not** provide exact numeric values for every plotted point in the text, and these learning curves concern samples/trajectories rather than independent shapes.
|
||
|
||
The exact small-data figures are from **transfer learning**, not from training on 200–600 unique geometries: for NS-PwC, a model pretrained on NS-Gauss reaches **13.1% error with 4 trajectories**, while training from scratch requires **128 trajectories** to reach that error (reported as a 32× sample-efficiency gain). For NS-SL, transfer with **32 samples** gives **4.7%**, versus **33.4%** from scratch with 32 samples (Section F.9; Figure F.6). Both comparisons use datasets for the same PDE but different initial-condition distributions; they do not show transfer to new geometry shapes.
|
||
|
||
### Geometry, inputs, and target-geometry inference
|
||
|
||
RIGNO operates on physical point clouds and constructs a coarser regional mesh by subsampling their nodes. Encoder/decoder connections use adaptive support regions; regional processor edges use Delaunay-based and multiscale connections. The paper’s claimed generality is for arbitrary **point-cloud domains**, but its stated scope and experiments are **two-dimensional**. For periodic problems, it uses periodic coordinate encodings and a tiled domain to create cross-boundary edges. The authors say graph construction must be adapted to the PDE’s boundary conditions; the method does not specify a universal boundary-condition input channel. (Methods; Supplement A.)
|
||
|
||
The generic inputs described are coordinates, discretized solution values, spatial coefficient values, current time, and lead time. For the geometry-varying examples, AF adds a distance field from the airfoil boundary as input and predicts density; Elasticity adds a distance field from the hole boundary and predicts stress. AF uses a C-grid mesh with \(200\times50\) quadrilateral elements. Elasticity’s target is available at 972 unstructured coordinates. These dataset-specific inputs and targets are described in Section D.2.
|
||
|
||
**Target-geometry latent optimization:** The paper does not describe optimizing a latent code for each test geometry using target-solution labels. Its described inference is a forward model evaluation using the geometry/point-cloud inputs and constructed graphs; graph construction may need to be repeated for changing geometries. The paper does not report a test-time target-label fitting procedure. It also does not provide a separate count of per-shape trainable latents.
|
||
|
||
### Training, compute, limitations, and code
|
||
|
||
The standard RIGNO configuration uses 2.7M parameters for RIGNO-18 or 1.9M for RIGNO-12, six regional-edge levels, sub-sampling factor 4, latent node/edge dimensions 128, and edge-masking probability 0.5 (Table E.1). Training uses batch size 16 and AdamW; the paper reports 2,500 epochs for time-dependent tasks and 3,000 for time-independent tasks. Most experiments used **two NVIDIA GeForce RTX 3090 GPUs (24 GB each)**; some used up to eight GPUs. The authors estimate **9,000 GPU-hours** to reproduce the experiments, including supplementary material (Section E).
|
||
|
||
A separate single-RTX-3090 comparison reports peak inference memory in Table F.4: for Poisson-Gauss, 121 MB for RIGNO-18 (96 features) and 159 MB for RIGNO-18 (128 features); for Elasticity, 51 MB and 64 MB, respectively. Those are **inference peak-memory** figures for that benchmark, not training VRAM requirements. Table F.4 also reports inference times and test errors; the paper says RIGNO trains more slowly than GINO on fixed-geometry datasets but faster on variable-geometry data in that comparison.
|
||
|
||
**Limitations stated by the authors:** the work focuses on 2D; 3D extension is described as possible in principle, but 3D benchmarks were lacking. Universal-approximation theory and rigorous generalization-error bounds are left for future work. They also note that graph construction for changing physical meshes can be considerable relative to model execution (Discussion; Supplement A.4). Thus, applying this evidence to complex 3D injection-molding shapes would be an untested extrapolation.
|
||
|
||
Official code: **https://github.com/camlab-ethz/rigno** (Methods, Section 2).
|
||
|
||
### Useful paper locations
|
||
|
||
Table 1 lists the point-cloud benchmark errors for all 13 tasks; Table 2 gives the Cartesian-grid comparisons. The RIGNO and baseline columns use median relative \(L_1\) error in percent.
|
||
[REGION: page=7, bbox=70,155,402,849]
|
||
|
||
Table 2 compares RIGNO with CNO, scOT, and FNO on nine Cartesian-grid tasks; it is not a geometry-OOD evaluation.
|
||
[REGION: page=7, bbox=381,155,646,847]
|
||
|
||
Table D.1 identifies each task’s PDE, boundary condition, domain, and whether the domain varies. It shows AF and Elasticity as the geometry-varying tasks.
|
||
[REGION: page=21, bbox=79,155,374,845]
|
||
|
||
Figure 2 shows RIGNO-12 test error versus training-set sizes 256, 512, and 1,024 for selected datasets, plus error versus model size. It supports a qualitative scaling trend, but the plotted learning-curve points are not tabulated as exact values.
|
||
[REGION: page=8, bbox=75,155,306,850]
|
||
|
||
Table F.3 gives the 128-ID/128-OOD sample results for Darcy Flow, Smooth Transport, and Compressible Euler; its shifts do not establish a geometry-separated split for the main benchmarks.
|
||
[REGION: page=35, bbox=714,155,866,717]
|
||
|
||
Figure F.6 shows the transfer-learning curves for NS-PwC and NS-SL. The text reports the exact examples: 4 transferred trajectories versus 128 from scratch for NS-PwC at 13.1%, and 32 transferred samples at 4.7% versus 33.4% from scratch for NS-SL.
|
||
[REGION: page=36, bbox=286,155,548,846]
|
||
|
||
Region format: page is 1-indexed; bbox=top,left,bottom,right, normalized 0-1000 from top-left.
|