Mesh-Agent MLP Suite — Model Card
Model Summary
This work presents a suite of four lightweight Multilayer Perceptron (MLP) models designed to support intelligent mesh-refinement assistance in Finite Element Analysis (FEA). Each model targets a specific simulation preprocessing task:
| Model | Task | Level | Output |
|---|---|---|---|
| Model 1 | Local refinement flag | Feature | Binary probability |
| Model 2 | Convergence prediction | Part | Binary probability |
| Model 3 | Local mesh sizing | Feature | Regression |
| Model 4 | Global mesh sizing | Part | Regression |
Independent modeling avoids multi-loss interference and improves stability on the limited available dataset.
Dataset Reference
All models were trained using the publicly available dataset:
MeshRefine-FEA Corner Bracket Dataset
Hugging Face: https://huggingface.co/datasets/Iris314/Mesh-gen-train
Dataset card includes simulation metadata, adaptive refinement history, convergence labels, and geometric descriptors.
Architecture and Rationale
Graph Neural Networks (GNNs) were evaluated initially due to their alignment with geometric adjacency. However, the limited number of unique CAD parts led to severe overfitting and unstable training. Support Vector Machines were also considered but were rejected because they do not provide probabilistic outputs, which are required for confidence-aware threshold control. A unified multi-task MLP was tested but exhibited degraded performance due to conflicting optimization objectives.
Four independent MLPs were therefore selected for stronger performance-to-complexity trade-offs and for greater explainability, which is essential in engineering education contexts. To propagate local stress-driven insights into global decisions, feature-level models are trained first and their top-three predicted refinement regions are appended to the input of the part-level models. This ensures dimensionally consistent message passing while preserving local importance.
Training Details
| Setting | Value |
|---|---|
| Optimizer | Adam |
| Learning rate | 1e-4 |
| Epochs | 500 |
| Loss functions | BCE / MSE |
| Thresholding | Youden’s J statistic |
| Best classification threshold | 0.1588 |
| Cross-validation | 5-fold (part-level models) |
| Framework | PyTorch |
Low learning rate and extended training duration reduce overfitting under limited data conditions. A high-recall bias ensures that refinement-critical zones are always captured.
Evaluation
| Model | Metric | Score | Notes |
|---|---|---|---|
| Model 1 | Accuracy | 0.82 | Refinement classification |
| Model 2 | Accuracy | 0.78 | CV: 0.58 ± 0.06 |
| Model 3 | MSE | 0.253 | Local mesh sizing |
| Model 4 | MSE | 0.364 | CV: 0.5135 ± 0.2575 |
Confidence scores are presented in the GUI to support user interpretation of predictions.
Intended Use
Suitable for:
- Beginner FEA instruction and mesh setup feedback
- Early detection of non-convergent setups
- Estimating balanced local and global mesh sizing to reduce runtime
Not intended for certified structural verification in safety-critical systems.
Limitations
- Dataset consists only of corner bracket geometries
- Assemblies and nonlinear load modes not represented
- Human verification of results remains required
Ethical Considerations
This tool is designed for educational scaffolding and not for automated deployment in high-risk engineering design without expert review.
Citation
Model citation
@model{mesh_agent_mlp_2025,
title={Mesh-Agent MLP Suite for Intelligent FEA Refinement},
author={X. Tang et al.},
year={2025},
note={Feature-aware mesh refinement prediction models for educational CAE assistance},
}