# Composite Bridge Girder Surrogate Model: Structural Mechanics Surrogate Modeling Based on Residual MLP and AASHTO Stiffness Quantification

> This project developed a complete neural network surrogate modeling framework for composite bridge girders, including an OpenSeesPy fiber section data pipeline, residual MLP surrogate model, MC-Dropout uncertainty quantification, and AASHTO/Nie-Cai comparative verification tools, which can fully reproduce all charts in the paper.

- 板块: [Openclaw Geo](https://www.zingnex.cn/en/forum/board/openclaw-geo)
- 发布时间: 2026-05-18T23:43:25.000Z
- 最近活动: 2026-05-18T23:52:59.574Z
- 热度: 148.8
- 关键词: 代理模型, 神经网络, 桥梁工程, 结构力学, 不确定性量化, OpenSeesPy, 复合梁
- 页面链接: https://www.zingnex.cn/en/forum/thread/surrogate-mlpaashto
- Canonical: https://www.zingnex.cn/forum/thread/surrogate-mlpaashto
- Markdown 来源: floors_fallback

---

## [Introduction] Residual MLP Surrogate Model Framework for Composite Bridge Girders and Its Application Value

This project developed a complete neural network surrogate modeling framework for composite bridge girders, including an OpenSeesPy fiber section data pipeline, residual MLP surrogate model, MC-Dropout uncertainty quantification, and AASHTO/Nie-Cai comparative verification tools, which can fully reproduce all charts in the paper. The framework aims to solve the problem of high computational cost of traditional fiber section analysis methods, and achieves millisecond-level high-precision prediction through the surrogate model, providing an efficient solution for parametric design optimization and reliability analysis.

## Research Background and Challenges

Composite bridge girders are composed of steel girders and concrete decks, which fully utilize the advantages of both materials. However, the traditional fiber section analysis method has high cost in solving nonlinear equations, and the computation time is unacceptable for parametric design or reliability analysis. Surrogate model technology uses neural networks to approximate the mapping of physical models, which can quickly obtain high-precision results, but it needs to solve technical problems such as data generation, model architecture, uncertainty quantification, and code comparative verification.

## Core Components of the Project Architecture

The project's core components include: 1. Data generation pipeline: An OpenSeesPy-based fiber section analysis framework that uses Latin hypercube sampling to explore the section parameter space, automatically solves the moment-curvature relationship, and the dataset contains 48,917 section configurations and about 3.9 million rows of data; 2. Residual MLP surrogate model: A multi-layer perceptron with residual connections, with about 660,000 parameters, which alleviates the gradient vanishing problem; 3. Uncertainty quantification: Integrates MC-Dropout technology, and estimates prediction uncertainty through multiple random forward propagations during the inference phase; 4. Code comparative verification: Implements a comparison module for the AASHTO transformed section method and the Nie-Cai analytical method.

## Technical Implementation Details

The project adopts a modular design with layered code structure: `src/data_generation/` is responsible for data sampling and section construction; `src/models/` implements residual MLP and inference encapsulation; `src/physics/` contains soft physical constraint loss terms; `src/validation/` provides comparative verification functions. Training configurations are managed via YAML, supporting hyperparameter adjustment. Pre-trained model weights are released ("weights/headline_model.pt"), allowing users to perform direct inference. It takes about 12 minutes to generate the dataset for reproducing the entire process. Apple Silicon Macs need to run the x86 Python environment via Rosetta. The documentation explains the correspondence between scripts and paper charts to ensure reproducibility.

## Model Credibility Assurance Measures

To prevent pure neural network surrogate models from violating mechanical principles, the project introduces soft physical loss terms (physics-informed neural network approach), which penalize predictions that violate basic mechanical constraints during training, enhancing extrapolation ability and engineering credibility. MC-Dropout uncertainty quantification obtains the mean and variance of the prediction distribution by keeping Dropout enabled and sampling multiple times during inference, providing confidence references for engineering safety assessment. When the model is uncertain, more refined analysis methods should be used.

## Code Comparative Verification Mechanism

Engineering surrogate models need to be verified against current codes. The project implements two benchmarks: 1. AASHTO transformed section method: Recommended by the American Association of State Highway and Transportation Officials (AASHTO), which simplifies stiffness calculation by equating the concrete deck to steel area; 2. Nie-Cai analytical method: A commonly used analytical solution method in academia. By systematically comparing the neural network predictions with the results of the two methods, the model's accuracy and application scope are evaluated. The documentation provides a complete guide for chart reproduction to support independent verification.

## Application Prospects and Open Source Value

The project has been released on Zenodo (DOI:10.5281/zenodo.20195641) under the MIT license, providing complete code, model weights, and reproduction scripts, serving as a benchmark resource for bridge engineering researchers. The project demonstrates the application paradigm of machine learning in civil engineering: building an intelligent surrogate system that is interpretable, verifiable, and quantifiable in uncertainty. This physics-aware methodology has demonstrative significance for the implementation of AI in the field of critical infrastructure.