AI‑Driven Simulation: Expert FAQ on Optimization and Data Fusion

How Can Data From Measurements and Simulations Be Combined?

At CADFEM, we use methods that link different data sources. This so-called multi-fidelity modeling makes it possible to use measurements from experiments and results from simulations together. In experiments, the same quantities are usually measured as those calculated in simulations. However, it is also possible to include different quantities. By combining both data sources, a powerful model is created that learns internally from the data and can predict both simulation results and experimental measurements.

The big advantage: Just a few experiments are enough to calibrate a solid model. Additionally, experiments are used to examine correlations between simulation and reality. Simulation data, on the other hand, can be generated quickly and cost-effectively in large quantities – resulting in a model that combines efficiency and accuracy.

How Do You Handle Uncertainties in Measurement and Simulation Accuracy in Your Workflow?

Anyone working with experimental data usually has to deal with uncertainties. By using probabilistic models, these can also be managed in simulation results with different model qualities reflected in the uncertainties. While this is generally less critical in simulations, singularities or other phenomena can still occur. Therefore, a probabilistic approach is helpful because a key part of training probabilistic models is distinguishing actual correlations in the data from statistical variations (noise). 

This is a major advantage compared to neural networks, which cannot do this. A probabilistic model recognizes that variance in the data comes from noise and tries to uncover the true physics behind it. 

What Benefits Does a Generative AI Model Offer When Solving an Optimization Problem With Multiple Objectives and Selecting an Optimal Design?

The big advantage of generative AI is that it can be combined with predictive AI – a model that predicts results for a given geometry without requiring parameterization. It’s important to note that parameterization always imposes restrictions on geometry. Once you parameterize a geometry, you can only obtain geometries that fit that parameterization. 

By incorporating generative AI into geometry optimization, geometries can be combined that could never be combined based on parameterization because no corresponding parameterization is possible. Therefore, the non-parametric approach offers much greater flexibility and more diverse generative design options for desired geometries. 

However, you do need training geometries to train the model, which is not required for parametric optimizations. There, you can start immediately after parameterizing the geometry. 

When Should Bayesian Optimization Use the True Function (e.g., FEM/CFD Simulation) and When the Surrogate Model? If We Rely on the Surrogate Model, How High Should the Prediction Accuracy (e.g., R²) Be to Work With This Approach?

Bayesian optimization combines both: surrogate model and true function. It’s not an either/or but an iterative process: 

1. Start: Initially, only a few data points are available, e.g., from five initial FEM/CFD simulations.
2. Surrogate model: Based on this data, a probabilistic model (e.g., Gaussian process) is trained to approximate the true function and quantify uncertainties.
3. Candidate selection: The surrogate model suggests the next promising design point via an acquisition function (e.g., Expected Improvement).
4. Validation: This point is evaluated with the true function (simulation). The result flows back into the model – and the cycle starts again. 

The strategy is to use the surrogate model to select new simulations and thus steer the experimental design. However, every suggestion is always validated by the true function. It is also possible to optimize only with the surrogate. 

How Accurate Does the Surrogate Model Need To Be?

A high R² value (e.g., > 0.9) is only required if you want to rely exclusively on the surrogate model – e.g., for purely virtual optimizations without simulations. In typical Bayesian optimization, however, a “good enough” model that recognizes the main relationships and accounts for uncertainties is sufficient. Even with an R² of 0.7–0.85, the model can efficiently lead to the optimum because it improves with each new simulation.

How Are Changes to Models, Datasets, and Configurations Tracked?

Tracking changes to data, models, and configurations is always project-specific and individual. Depending on customer requirements, various approaches are useful: from simple versioning strategies to advanced machine-learning operations workflows. If tools or strategies are already in use at the customer, our models can be flexibly adapted to them. 

This can also be implemented with existing Ansys software, such as the data management system Ansys Minerva, to create a tailored solution. Formal tracking is particularly useful when many changes are made and an overview of model and data versions needs to be maintained.