AgentFEM Pipe ROM
v1.3 — Power-piping research sections
Four pipe services, 24 new solid-FE components, 108 assembly records.
Main steam, hot reheat, cold reheat and high-pressure feedwater now have independently assembled component operators. The Space switches section geometry, stiffness, mass and effective thermal expansion together. It includes thermal movement, support reactions, modes, free decay and forced-response comparison.
The main-steam section uses a published calculated bore/wall pair; other sections and all effective material constants, layouts and loads are benchmark assumptions. See case definitions and provenance. This is still a dry-pipe, constant-property research model, not plant design approval.
New validation: 108 independent Python/browser comparisons, 12 harmonic sweeps, and four refined-bend mesh checks. Original records are preserved. The complete library now contains 72 components. Data and operators use CC BY 4.0; solver/example code retains its existing Apache-2.0 license.
中文：新增主蒸汽、再热热段、再热冷段和高压给水四类研究截面。可比较管路、 支撑与阻尼器设置对热位移和振动的影响。尺寸来源、研究假设与验证报告一并提供。
v1.2 — steady harmonic response
The browser solver also accepts harmonic: {node: 3, axis: 2, force: 100, frequency: 10, zeta: 0.02, damper: {node: 3, axis: 2, c: 1000}}.
Force is peak N, frequency Hz, damping Ns/m; axis 0/1/2 means global X/Y/Z.
It solves complex dynamic stiffness in 12 mass-normalized system modes,
including full non-proportional damper coupling. harmonic.field contains real
and imaginary centreline amplitudes for exp(iωt), about linear equilibrium.
The reference has no added damper but retains the same assumed modal damping.
Sweeps cover 0.2–2 f1 at 161 discrete frequencies; they do not identify exact peaks.
Run node example_harmonic.cjs with Node.js; no Python server is needed.
Nineteen independent SciPy cross-checks and 12/24-mode comparisons are provided.
Generic component dimensions and temperature-independent steel are retained;
this is not yet a plant-specific main-steam or reheater model.
Connect components. Solve a new pipe system in milliseconds.
Interactive lab → · Full FE dataset → · AgentFEM →
A downloadable library of 48 physics-based component reduced-order models for straight pipes and 90° elbows. This is a matrix-based model, not a pretrained neural network. It needs no training framework or server for browser inference.
Inputs → outputs
Choose component IDs, connection rotations, support DOFs, uniform temperature rise, and gravity. The solver returns port motion, recovered centreline displacement, support reactions, and natural frequencies/mode shapes of the connected system.
Each component stores a 12 × 12 static stiffness matrix, 24 × 24 Craig–Bampton stiffness and mass matrices, thermal and gravity load maps, and centreline recovery operators. Units are SI; rotational DOFs use radians. Elbows include real straight tangents of 2D on each side. There is no nearest-neighbour matching between routes.
Run in Python
pip install numpy scipy huggingface_hub
hf download HaomingLuo/AgentFEM-Pipe-ROM --local-dir pipe-rom
cd pipe-rom
python example.py
The included pipe_rom.py loads compact JSON matrices and uses SciPy to solve an L-shaped route. You can modify the component sequence and supports directly. No FEniCS installation is needed to use the exported model; reproducing the original solid matrices requires AgentFEM/DOLFINx.
Run in JavaScript
solver.js is the same kernel used by the Space. With Node.js:
const fs = require('node:fs');
const {solve} = require('./solver.js');
const load = id => JSON.parse(fs.readFileSync(`data/${id}.json`));
const result = solve({
components: [load('straight-d200-t06-r5'),
load('elbow-d200-t06-r2'), load('straight-d200-t06-r5')],
deltaT: 100, gravity: [0, 0, -9.81],
supports: [{node: 0, fixed: [0, 1, 2, 3, 4, 5]}]
});
console.log(result.frequency, result.maxDisplacement);
The default route gives approximately 3.27 mm maximum centreline displacement, 17.89 Hz first natural frequency and 188.3 kg dry pipe mass. The exact values and comparisons are in the dataset validation records.
Verification and range
v1.1: time-domain damper response
Add dynamic: {node: 3, axis: 2, c: 1000} to the JavaScript request (c in N·s/m; axis 0/1/2 = global X/Y/Z). result.dynamics contains physical times, modal coordinates, probe displacement, undamped reference, sampled damper force and energy. The browser integrates 12 system modes with the full non-proportional damping matrix. Initial condition: first undamped mode, largest coordinate component 1 mm, zero velocity. This is linear free decay about equilibrium, not earthquake response or transient thermal analysis.
For an independent Python reference, run python example_dynamic.py; dynamics.free_decay uses SciPy's matrix exponential rather than Newmark. All 49 published histories were checked against this formulation. Maximum signal discrepancy / initial peak was 1.17×10⁻⁵; six 12-versus-24-mode checks were below 1.66×10⁻⁴. See validation/dynamic_validation.json. No additional solid-component training or neural weights are involved.
The Space's agentfem.pipe-project/1 JSON records route, supported discrete dimensions, supports, temperature, gravity and grounded damper. Project files carry the model version and are validated on import. CSV and printable comparison reports contain unamplified numerical results. Browser draft storage is local to the device.
Three continuous-solid assembly comparisons: gravity tip-displacement discrepancies under 0.24%; first-three-frequency discrepancies under 0.25%. Two mesh-refinement checks: tip-displacement changes under 0.29%; first-three-frequency changes under 0.38%. Reported only for those tests.
All 512 released assembly cases were cross-checked between SciPy and JavaScript, with relative frequency differences below 5 × 10⁻⁹. Runtime measurements are in validation/browser_campaign.json; timing depends on hardware and route size.
Linear elastic, small-displacement, homogeneous steel, dry pipe. Interface sections are mechanically rigid with prescribed radial thermal dilation. Spring supports are bilateral linear springs without initial preload. The displayed thermal animation interpolates a static result; mode amplitudes are normalized. Not a pressure-code or fatigue assessment. Full methodology and reproducible generators are in the linked dataset.
Licenses
Numerical operators: CC BY 4.0. Original assembly code: Apache-2.0. The bundled ml-matrix library is MIT licensed; its license is in vendor/. Data provenance is recorded in the dataset.
中文
AgentFEM 管道组件降阶模型。将实体有限元计算得到的直管、弯管算子组合起来，在浏览器或 Python 中实时计算新的管路。可修改拼接方向、支撑、温升与自重，查看位移、反力及振型。
本版采用组件降阶，而非神经网络预测；这些算子与配套数据也为后续物理 AI 模型提供可复用的基线。