Accelerometer Flexure Surrogate Model

This repository contains a probabilistic surrogate model for a single-degree-of-freedom (SDOF) accelerometer flexure geometry. The model is designed to predict modal and structural responses based on geometric parameters, facilitating the analysis of cross-axis modal separation and structural integrity without requiring repeated high-cost Finite Element Analysis (FEA).

Project Objective

The goal is to replace computationally expensive FEA simulations with a fast, accurate surrogate model. By training an ensemble of neural networks, the system predicts the first four modal frequencies, maximum deformation, and maximum stress, while providing a measure of uncertainty to ensure the predictions are reliable for engineering decisions.

Geometry and Design

The surrogate model focuses on a flexure beam geometry intended for a MEMS accelerometer. The primary design objectives include:

• Single Degree of Freedom (SDOF): Ensuring the structure responds primarily along the intended axis.
• Cross-Axis Modal Separation: Maximizing the frequency gap between the primary mode and transverse/torsional modes to avoid signal interference.

Input and Output Parameters

Input Geometric Parameters

• Beam Height: Vertical dimension of the flexure.
• Beam Length: Longitudinal dimension of the flexure.
• Fillet Size: Radius of the fillets to reduce stress concentrations.
• Beam Width: Lateral dimension of the flexure.

Output Physical Responses

• Modal Frequencies (1-4): The first four natural frequencies of the structure (Hz).
• Maximum Deformation: The peak displacement under design load (mm).
• Maximum Stress: The peak von Mises stress to ensure the material remains within the elastic limit (MPa).

Repository Structure

Core Scripts

• model_build.py: Full pipeline for data loading, scaling, ensemble training, and performance metrics.
• predict.py: Contains the SurrogatePredictor class for making fast inferences with uncertainty.
• dataset_gen.py: Generates the input parameter grid based on design bounds.
• sweep.py: Automates the execution of Ansys Workbench to populate the dataset.
• ansys_runner.py: Handles the communication and parameter updates within Ansys Workbench.
• optimization.py: Performs design optimization using the surrogate model.
• clean_logs.py: Maintenance script to clear temporary log files.
• process_runner.py: Utility to execute scripts and save their terminal output to markdown files.
• main.py: Main entry point for the framework.

Analysis and Visualization

• manifold-extraction.py: Uses UMAP/PCA to visualize the output manifold and weight space.
• model-analysis.py: Conducts Jacobian sensitivity analysis to see how inputs affect outputs.

Folders

• models/: Stores the trained ensemble (ensemble_20_models.pkl) and scaling parameters.
• dataset/: Contains the dataset.csv used for training and validation.
• ansys_files/: Project files and parameter definitions for Ansys Workbench.
• plots/: Stores all generated figures and analysis results.

Installation and Setup

Prerequisites

• Python 3.x
• Required Libraries: numpy, pandas, scikit-learn, matplotlib, joblib
• Optional: umap-learn (for manifold extraction), termcolor (for formatted logs).

Running the Pipeline

1. Data Generation: Run dataset_gen.py to create the input grid in dataset/dataset.csv.

   python dataset_gen.py

2. FEA Data Collection: Ensure Ansys Workbench is open and the server port is known. Run sweep.py and enter the port number when prompted.

   python sweep.py

3. Model Training: Train the ensemble of 20 MLP regressors.

   python model_build.py

4. Inference: Use the SurrogatePredictor in predict.py to get results for new geometries.

Performance Metrics

The model is evaluated using $R^2$ and Mean Absolute Percentage Error (MAPE).

Target Parameter	R2 Score	MAPE (%)
Modal Frequency 1	0.9961	1.60%
Modal Frequency 2	0.9995	0.47%
Modal Frequency 3	0.9987	0.61%
Modal Frequency 4	0.9948	1.29%
Max Deformation	0.9849	3.56%
Max Stress	0.9686	4.00%