HaomingLuo/AgentFEM-Pipe-Components

AgentFEM Pipe Components

v1.2 — thermal movement and forced vibration

Adds 19 harmonic-response sweeps, including a blocked-force-direction control. Each sweep has 161 frequencies from 0.2 to 2 times the first natural frequency, with/without an added grounded viscous damper. Both curves use an explicitly assumed 2% background modal damping ratio and 100 N peak harmonic force. Displacements are peak amplitudes at the force joint/direction, not RMS. These are component-ROM labels, not plant measurements or new full-solid solves.

Independent SciPy assembly/complex solves agree within 1.09e-8 relative error. For these cases, 12 versus 24 modes changes the curve by at most 0.078% of the 24-mode peak. See validation/harmonic_validation.json and source/validate_harmonic.py. This tests numerical implementation and retention, not physical accuracy against an actual power-station piping system.

Reusable solid finite-element components for interactive piping mechanics.

Play with a pipe → · Download the ROM → · AgentFEM →

48 hollow-pipe components. 512 static/modal assembly examples and 49 free-decay histories with grounded viscous dampers. Full stiffness and mass matrices, reduction bases, thermal and gravity loads, and reproducible validation.

What is in the release?

Data	Count	How the labels are obtained
Straight tubes	24	3D solid finite elements, Q2 displacement
90° elbows with straight tangents	24	3D solid finite elements, Q2 displacement
Assembled pipe systems	512	Component ROM solves; not 512 independent full-solid simulations
Continuous-mesh assembly checks	3, plus the development pilot	Independent continuous solid assemblies without rigid internal ports
Mesh-refinement checks	2	21,168 → 70,080 displacement DOFs per component

Four outer diameters (120, 160, 200, 240 mm), two thickness ratios (t/D = 0.04, 0.06), three straight lengths (L/D = 3, 5, 8), and three elbow bend radii (R/D = 1.5, 2, 3). Every elbow includes a straight tangent of 2D at each end. These tangents are part of the geometry and mass.

Reference steel: E = 200 GPa, ν = 0.3, ρ = 7,850 kg/m³, α = 12 × 10⁻⁶/K. Temperature-independent, small-displacement linear elasticity. Hexahedral meshes use Q1 coordinates and Q2 displacement interpolation, 24 axial stations, 24 circumferential divisions and one radial cell. Each component has 576 cells and 21,168 displacement DOFs.

Download and use

from datasets import load_dataset
components = load_dataset("HaomingLuo/AgentFEM-Pipe-Components",
                          "components", split="components")
examples = load_dataset("HaomingLuo/AgentFEM-Pipe-Components",
                        "assemblies", split="heldout_spatial")
print(components[0]["id"], examples[0]["frequencies_Hz"])

Download one component's arrays without the entire dataset:

from huggingface_hub import hf_hub_download
import numpy as np
path = hf_hub_download("HaomingLuo/AgentFEM-Pipe-Components",
    "components/elbow-d200-t06-r2/component.npz", repo_type="dataset")
c = np.load(path)                       # no pickle required
print(c["S"].shape, c["Kr"].shape)       # (12, 12), (24, 24)

See SCHEMA.md for DOF ordering and dimensions. source/ contains generation, assembly and validation code. SHA256SUMS lists the published artifact hashes.

Component mechanics

Each port has three translations and three small rotations. Constraint modes statically condense the solid interior onto two rigid mechanical port sections. Twelve mass-normalized fixed-interface modes form a Craig–Bampton dynamic reduction. Prescribed radial thermal dilation is retained at ports, so free uniform thermal expansion is not spuriously blocked.

The browser assembles these operators for the actual route and supports, then solves equilibrium and a generalized eigenproblem. It does not retrieve the closest example.

Validation

All 48 components pass rigid-motion, free uniform-expansion, mass and positive-definite reduced-mass checks. The three continuous-mesh assembly checks have tip-displacement errors of approximately 0.11%, 0.23%, and 0.15% under gravity, with first-three-frequency errors below 0.25%. The development pilot additionally checks uniform heating.

The two mesh-refinement checks change tip displacement by less than 0.29%, and the first three frequencies by less than 0.38%. These are measured discrepancies on the named tests, not a universal error bound for every route or support placement. Detailed values, retained-mode checks, and the initial, superseded interface experiment are preserved in validation/.

Twelve mixed-load/support checks compare recovered full displacement fields with unreduced rigid-port FE systems. Separately, all 512 assembly cases agree between the SciPy and JavaScript assemblers; this cross-language check verifies implementation consistency, not independent physical accuracy.

Assembly examples and splits

The cases contain L, U, spatial and offset routes; fixed inlet; optional free, spring, vertical-restraint or fixed end conditions; gravity; and ΔT ∈ {0, 25, 50, 100, 150, 200} K. Spring stiffness is sampled between 10⁴ and 10⁶ N/m. Seed: 20261010.

development contains 384 planar cases. heldout_spatial contains 128 spatial cases, separated by topology. Components are shared across splits: this is not a held-out-material or held-out-component benchmark. No neural model has been trained on these splits in this release.

Potential uses: physics-preserving component surrogates, graph assembly models, operator learning, uncertainty studies, and interactive mechanics education. Local stress peaks, code compliance, pressure effects, fluid mass, contact, spring preload, plasticity and fatigue are not labelled here.

Reproducibility and licensing

v1.1: viscous dampers and time histories

free_decay/validation contains 49 validation cases: L, U and spatial routes; free or vertically spring-supported ends; global Y/Z damper directions; c = 0, 1, 10 or 100 kN·s/m; plus a blocked-direction negative control. D = 200 mm, t = 12 mm, straight L/D = 5, elbow R/D = 2. All cases release the first undamped mode from a largest coordinate component of 1 mm, with zero velocity. Duration is six first-mode periods. No background damping or external forcing is added.

The single damper connects a joint to ground. Its full modal matrix is C = c b bᵀ, not a fitted exponential envelope. Twelve system modes are integrated with average-acceleration Newmark (β = 1/4, γ = 1/2). Each history includes 601 physical-time samples of measurement displacement, undamped reference, damper force, mechanical energy and dissipated energy. These labels reuse the existing solid-derived ROM; they are not new full-solid transient FE runs or field measurements.

All 49 histories were checked against an independent SciPy matrix-exponential implementation of the same modal equations: maximum displacement discrepancy / initial peak 1.17×10⁻⁵, force discrepancy normalized by max(1 N, peak reference force) below 1.98×10⁻⁴. Six strong-damping retention checks (12 vs 24 system modes) have maximum signal discrepancy / initial peak 1.66×10⁻⁴. These checks establish implementation and truncation evidence on the named cases, not general plant-level accuracy. See validation/dynamic_validation.json.

The full benchmark is a validation set, not a claimed independent ML train/test split. The Space also supports undo/redo, joint selection, endpoint preview, baseline comparisons, project JSON, CSV and printable reports. Individual force peaks are sampled at the published output times.

Time-integration reference: OpenSees Newmark documentation.

Generated using AgentFEM public solid operators, with project-local reduction and assembly code. AgentFEM source revision: a05b293104c15f8ca3916b98edb24801e0e9ab6c. Runtime: DOLFINx 0.11.0; Python, NumPy and SciPy; serial generation. Exact runtime versions are in provenance.json.

Data and numerical operators: CC BY 4.0. Original example/generation code: Apache-2.0. See LICENSE-DATA and LICENSE-CODE. No field measurements or proprietary plant geometry are included.

Citation

Luo, Haoming. AgentFEM Pipe Components: Solid-FE Components and Assembled Reduced-Order Examples. Hugging Face, 2026. Cite the repository revision used.

Method reference: Craig, R. R., & Bampton, M. C. C. (1968). Coupling of substructures for dynamic analyses. AIAA Journal, 6(7), 1313–1319. https://doi.org/10.2514/3.4741

中文简介

这是面向管道组件拼接与快速力学计算的数据集：48 个三维实体有限元组件，512 个组件降阶组合算例，以及可复现的验证记录。包含直管、带过渡直段的 90° 弯管、热膨胀、自重、支撑反力和低阶振动。

组件数据保留刚度矩阵、质量矩阵、降阶基与场恢复算子。组合算例由这些组件实时装配求解，不是各自重新进行完整实体仿真。可用于后续神经网络代理、组件图模型与交互式力学教学。

v1.1 新增 49 组黏性阻尼器自由衰减时程，记录位移、阻尼器作用力和能量，并提供独立时间积分核验。网页可保存管路方案、比较支撑调整前后的结果，并导出计算记录。