AgentFEM × GINO · Bracket-128

Predict a displacement field from geometry. Explore weight and stiffness.

Dataset · Interactive lab · AgentFEM-Learning

Try the direct browser demo for a visual introduction. The demo uses recorded candidates; the local inference example below runs the released hybrid model.

This checkpoint combines a GINO backbone with a symmetry-aware residual MLP and a fixed-face mask. It is trained on 128 AgentFEM bracket geometries; the released seed (2026) was selected using validation error, not test error.

On 32 fresh held-out geometries: mean surface displacement error 3.08%, mean top-seat deflection error 0.80%. These are hybrid-model results, not raw GINO accuracy.

Run locally

pip install -r requirements.txt
python inference.py --model . --sample example.npz

Download this repository first. No finite-element installation is needed for inference.

import numpy as np
from inference import BracketPredictor, PARAMS
d = np.load("example.npz", allow_pickle=False)
model = BracketPredictor(".")
u = model.predict(d["surface_points"], dict(zip(PARAMS,d["parameters"])), load_N=10000)
# u: displacement in metres at each supplied surface point

operator_state.pt contains the backbone weights, architecture and training normalization. correction.pt contains residual-network weights. Both are required for the hybrid result. mlp_only.pt is an independent coordinate-network baseline; choose kind="mlp_only" to inspect it. Use kind="gino" for the raw backbone. example.npz contains one fresh test geometry with its FEM reference.

Data-size comparison

Means across three training seeds. Every size is trained from scratch with the same architecture and fixed validation/test sets. Full metrics, per-case records and seed variation are in evaluation.json and learning_curve.csv.

Scope

Fixed symmetric double-arm bracket family, fixed foot undersides, downward uniform seat loading and E=70 GPa, nu=0.33. Parameters: depth and radius are reference half-dimensions in metres, waist is dimensionless, bow is metres. Ranges: depth 0.015–0.032 m; radius 0.010–0.023 m; waist 0.04–0.38; bow 0.005–0.030 m. Predictions are surface displacement, not stress or a strength certificate. Different materials, topology, contact or nonlinear response require another model. The existing Space is a lightweight recorded-candidate demo, not this checkpoint running on an HF inference endpoint.

Method and credit

GINO: Li et al., Geometry-Informed Neural Operator for Large-Scale 3D PDEs. Backbone implementation: NeuralOperator. FEM data: AgentFEM. Training integration: AgentFEM-Learning. Code and weights: Apache-2.0. Dataset: CC BY 4.0.