✓✓ Beats tuned baseline 2026

Exact Elasticity-Complex Message Passing

Implementation & benchmark of arXiv:2608.23829 — A Vector-Valued Co-Chain/Chain Complex Associated to the Elasticity Complex

Usefulness6/10

Difficulty6/10

Novelty6/10

Architecture Stability Accuracy Generalization Graph nn Geometry World model Differential geometry Geometry Linear algebra Topology

Source paper: A Vector-Valued Co-Chain/Chain Complex Associated to the Elasticity Complex arXiv:2608.23829 ⓘ · analyzed Aug 29, 2026

AI-generated research hypothesis, automatically tested. Not peer-reviewed.

Idea description

Construct a mesh neural network with node, edge, face, and cell feature spaces modeled on the four spaces of the discrete elasticity complex. Replace unconstrained cross-order message passing by fixed incidence and geometric operators whose compositions vanish exactly, so gradient-like, incompatibility-like, and divergence-like features cannot contain algebraically spurious components.

Formulas

$$C^{\infty}(\overline{\Omega})\otimes\mathbb{V}\xrightarrow{\varepsilon}C^{\infty}(\overline{\Omega})\otimes\mathbb{S}\xrightarrow{\operatorname{inc}}C^{\infty}(\overline{\Omega})\otimes\mathbb{S}\xrightarrow{\operatorname{div}}C^{\infty}(\overline{\Omega})\otimes\mathbb{V}$$

$$\varepsilon=\operatorname{sym}\operatorname{grad},\qquad \operatorname{inc}u=\operatorname{curl}(\operatorname{curl}u)^{T}$$

$$I_{h}^{1}\varepsilon=\varepsilon I_{h}^{0},\qquad I_{h}^{2}\operatorname{inc}=\operatorname{inc}I_{h}^{1},\qquad I_{h}^{3}\operatorname{div}=\operatorname{div}I_{h}^{2}$$

$$D_{1}D_{0}=0,\qquad D_{2}D_{1}=0,\qquad h_{k+1}=\phi_{k+1}(D_{k}h_{k})+\psi_{k+1}(h_{k+1})$$

Mathematical statement

The paper starts from the elasticity complex C-infinity of Omega tensor V to C-infinity of Omega tensor S to C-infinity of Omega tensor S to C-infinity of Omega tensor V, with operators epsilon, inc, and div. Here V is R^3, S is the space of symmetric 3 by 3 matrices, epsilon u equals sym grad u, and inc u equals curl of the transpose of curl u. The discrete construction is algebraic and uses simplicial geometry. Its key guarantee is the commuting set I_h^1 epsilon equals epsilon I_h^0, I_h^2 inc equals inc I_h^1, and I_h^3 div equals div I_h^2, where I_h^k maps a smooth field to discrete features at order k. Implement a discrete analogue with feature spaces X_0, X_1, X_2, and X_3 and linear maps D_0 from X_0 to X_1, D_1 from X_1 to X_2, and D_2 from X_2 to X_3. Enforce D_1 D_0 equals zero and D_2 D_1 equals zero. D_0 is a geometry-aware symmetric-gradient stencil, D_1 an incompatibility stencil, and D_2 a divergence stencil. The zero compositions ensure that features generated as discrete gradients are annihilated by the next operator, preventing exact compatible fields from being interpreted as defects.

Implementation notes

1. Integration point: use a tetrahedral or triangular mesh graph network for physical simulation, geometric segmentation, or mesh completion. Maintain four tensors: h_0 on vertices, h_1 on directed edges, h_2 on oriented faces, and h_3 on cells. Store edge tangent vectors, face normals, cell volumes, and orientations. Build sparse geometry-aware matrices D_0, D_1, and D_2. D_0 maps vertex vector features to edge symmetric-gradient features using finite differences projected onto edge tangents. D_1 maps edge symmetric-tensor features to oriented face incompatibility features using signed edge-face incidences and local curl-curl contractions. D_2 maps face tensor fluxes to cell vector divergence using signed face-cell incidences and face normals. 2. Pseudocode: initialize h_0 from node inputs; at each layer compute g_1 = D_0 h_0, g_2 = D_1 h_1, and g_3 = D_2 h_2; update h_0 = h_0 + MLP_0(concat(h_0, D_0_transpose g_1)), h_1 = h_1 + MLP_1(concat(h_1, g_1, D_1_transpose g_2)), h_2 = h_2 + MLP_2(concat(h_2, g_2, D_2_transpose g_3)), and h_3 = h_3 + MLP_3(concat(h_3, g_3)). Use normalized sparse operators and residual connections. 3. Compute D_k deterministically from mesh incidence and geometry; do not learn them. Numerically verify the Frobenius residuals ||D_1 D_0||_F divided by ||D_0||_F and ||D_2 D_1||_F divided by ||D_1||_F before training. 4. First experiment: train on a small tetrahedral-mesh elasticity benchmark such as synthetic displacement-to-stress prediction, comparing a standard MeshGraphNet against this complex network on randomly remeshed versions of the same domain. Measure one-step error, 50 to 100 step rollout error, and cross-mesh generalization. Success is lower error and reduced rollout drift at equal parameter count, particularly when training on one mesh resolution and testing on another. Also verify that complex violations remain at numerical precision.

Verification

Beats tuned baseline

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Confirmed ✓

Stage 2 · Practical benchmark result: Significant win ✓

Methodology: 8 paired seeds, permutation test (p<0.05) against a learning-rate-tuned baseline. How verification works

Stage 1 — Mechanism check agent confidence 8/10

Built a reproducible simplicial-complex MVP with deterministic incidence operators D0, D1, and D2, plus a toy defect-regression comparison. The algebraic guarantee was observed exactly: relative residuals for D1D0 and D2D1 were both 0.0, and compatible-field leakage was 9.4e-17 versus 8.36 for an unconstrained cross-order map. The exact operator did not improve the toy regression MSE (0.00226 versus 0.000282), so this demonstrates the claimed structural cancellation but not an end-to-end accuracy win.

Agent confidence: 8/10

Baseline: Unconstrained random edge-to-face map: compatible leakage 8.3618; toy MSE 0.0002822; accuracy 1.0
Idea: Incidence complex: relative ||D1D0||/||D0|| = 0.0, relative ||D2D1||/||D1|| = 0.0; compatible leakage 9.399e-17; toy MSE 0.0022564; accuracy 1.0

Limitations: This is an algebraic simplicial incidence prototype, not a full 3D elasticity complex with symmetric-tensor geometry-aware finite-element operators, physical elasticity data, remeshing, rollout evaluation, sparse GPU execution, or matched-parameter MeshGraphNet. The learned toy task is small and synthetic, and the unconstrained baseline is a random linear cross-order map rather than a full standard MeshGraphNet.

How to run: python3 complex_mvp.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.23829) Sep 3, 2026

Beats the tuned baseline by 99.5% (p=0.0081, wins 8 / 8 paired seeds; metric: lower is better).

Benchmark: Tetra_elasticity_complex
Model: mlp_tiny
Paired seeds: 8
Baseline mean: 0.0329
Idea mean: 0.0002
Effect (Δ): -0.0327 (−99.5%; negative = idea better)
Wins: 8 / 8 paired seeds
p-value: 0.0081 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±41.4%
Mechanism: Confirmed ✓
Practical effect: Helps
Baseline tuning: swept over 3 configs

Limitations:

The benchmark is a small synthetic two-tetrahedron PDE-style regression task, not a full 3-D elasticity finite-element discretization, physical elasticity dataset, MeshGraphNet, remeshing, cross-resolution generalization, sparse-operator scaling, or long-horizon rollout study.

How to run: python3 run_bench.py

Verdict computed by deterministic test code from paired-seed statistics — not by the language model.

Stage 2 — Benchmark Sep 3, 2026 latest Worked ✓

Agent confidence: 8/10

The registered tetrahedral PDE track was benchmarked with identical MLP systems, matched training budgets, a shared three-point learning-rate grid, and 8 paired seeds. The exact-complex projection achieved lower test MSE on every seed: 0.000175605 versus 0.0329221 for the tuned baseline, paired delta -0.0327465 and permutation p=0.0081. The mechanism signature also confirmed exact algebraic cancellation and near-zero trained-model response to compatible fields, though this remains a small synthetic two-tetrahedron demonstration rather than a full finite-element elasticity benchmark.

Baseline: Tuned lr=0.01; 8-seed test MSE mean 0.0329221145, std 0.0152480952.
Idea: Tuned lr=0.01; 8-seed test MSE mean 0.0001756050, std 0.0000313208; idea wins 8/8; paired delta -0.0327465095; permutation p=0.0081.

Benchmark result