Introduction

Surrogate modelling (also known as black-box modelling, meta-modelling or response surface modelling) is a technique used to approximate complex simulations with simpler mathematical models. In the context of 3DEXPERIENCE PLATFORM SIMULIA APPS or SIMULIA ABAQUS and ISIGHT, surrogate modelling is often used to reduce computational cost when running detailed simulations.

Instead of running expensive full simulations for every design variation or parameter set, a surrogate model is trained using a limited number of ABAQUS simulations. This model can then predict outcomes (like stress, deformation, etc.) quickly for new input values.

This post is a simple introduction with an easy-to-understand introduction to surrogate modelling without writing scripts on the 3DEXPERIENCE Platform.

⚙️ How It Works:

1. Design of Experiments (DOE): Choose several combinations of input parameters.
2. Run Simulations: Use ABAQUS to run FEA simulations for each design point.
3. Build Surrogate Model: Fit a simpler model (e.g., polynomial regression, kriging, or neural networks) to the simulation results.
4. Prediction & Optimization: Use the surrogate to perform sensitivity analysis, optimization, or uncertainty quantification efficiently.

This process is often integrated through SIMULIA Process Composer App or SIMULIA Isight, which can work alongside ABAQUS to automate surrogate model creation and design optimization.

📌 Uses of Surrogate Modelling in FEA

• Design Optimization: Quickly explore a large design space and find optimal parameters without needing to run thousands of physics simulations.
• Sensitivity Analysis: Understand how changes in inputs affect outputs.
• Uncertainty Quantification: Model and manage variability in material properties or boundary conditions.
• Real-time Simulations: Use the surrogate model in real-time systems where full simulations would be too slow.
• Multidisciplinary Analysis: Combine outputs from ABAQUS with other simulation tools in a streamlined optimization workflow.

Why integrate FE Simulations with ML? How does that even make sense?

Answer: Time! Time! Time!

Time is money. 

👉If a Formula Student/SAE BAJA team is testing a spaceframe chassis that has multiple load bearing members, and the team is struggling to figure out how to vary each member to reduce over torsional stresses of the chassis.

👉They make a parametric model; varying the length or cartesian position of each member. Every permutation of the configuration of the chassis design leads to a unique design, potentially leading to hundreds and thousands of hours of FE analyses...

(Assuming, they don't have nonlinear materials in the chassis which will be even more computationally expensive)

Surrogate Model

1.Surrogate models are low fidelity empirical models.

2. These are created bottoms up from simulation data

3. Capable of smoothing a noisy response

4. Extremely fast to evaluate!

5. Accurate.

Prerequisites of establishing a surrogate model:

1. Solve an FE Analysis with Geometric Parameters: 

2. In the Optimization Process Composer app, a structured process is constructed to execute a Design of Experiments (DOE), generating a representative set of input-output pairs across the design space. The resulting numerical data feeds directly into a Machine Learning Model training pipeline, where surrogate modeling techniques such as Response Surface Modelling (RSM) or Universal Kriging (UK) are applied to approximate the underlying response behavior.

To enhance predictive accuracy, hyperparameter tuning is integrated into the workflow, employing optimization strategies to minimize the mean approximation error (e.g., Mean Squared Error or Root Mean Squared Error) of the surrogate models.

Once the surrogate model is trained and validated, it is made accessible within the Results Analytics App, enabling advanced post-processing, visualization, and sensitivity analysis using the approximated response surfaces.

3. SIMULIA Results Analytics App
🔍 1. Data Extraction from ABAQUS Simulations

• Physics Results Analytics app allows you to visualize and extract results (e.g., displacements, stresses, temperatures) from simulation outputs.
• It supports efficient post-processing of large datasets—perfect for building surrogate models from a batch of simulation runs.

📊 2. Create and Analyze Datasets

• You can aggregate simulation results from multiple runs (DOE samples).
• Physics Results Analytics app helps in creating custom result metrics that can be used as outputs in the surrogate model.
• Supports filtering, grouping, and comparison across different parameter combinations.

🧠 3. Link to Surrogate Modelling Tools

• Once the simulation data is extracted and prepared, PRA integrates seamlessly with tools like Isight, 3DEXPERIENCE Process Composer, or even external Python/ML tools.
• You can use this clean, organized dataset to train your response surface models, kriging, neural networks, or other surrogate types.
• .

📈 4. Perform Sensitivity & Correlation Analysis

• PRA includes built-in tools to analyze parameter sensitivities and correlations—helpful for feature selection before surrogate model training.
• This guides you in focusing on the most influential inputs, improving model accuracy and reducing complexity.
   

🖥️5. Pre-process training of the model with ease 

🛠️ 6. Integration into Design Exploration

• The analytics and surrogate models can be looped back into the design process for optimization, trade-off studies, or uncertainty quantification.
• PRA provides a user-friendly interface to track design decisions and results over time.

The Approximation model's dashboard can be used this way:
 

Turning the breadth (BR) slider to 38.96

The FEA sensor output values update within 2 seconds with updated values.

Next, I will change the "ext" parameter while resetting the "BR" parameter.
 

The Mass and Displacement values update within 2 seconds.

All in all, we see that without actually going into CAD and FE apps on the platform, we get instant results of the FE analysis, saving time and eliminating repetitive workflows.

✅ Key Benefits

• Reduces time spent on manual data extraction from Simulation output.
• Enhances model-building efficiency and quality.
• Provides a centralized platform for simulation data analysis.
• Supports end-to-end workflows: from simulation to model to optimization.