Neural Network Surrogate Model

A method that utilizes deep neural networks (DNNs) as approximators for the input-output relationships in CAE. It learns nonlinear mappings from large amounts of simulation data, enabling real-time prediction.

Expressing this in a mathematical formula, it looks like this.

Loss function:

Neural network-type surrogates are an important method aiming for the fusion of data-driven approaches and physics-based modeling. While computational cost is a major bottleneck in conventional CAE analysis, introducing neural network-type surrogates can significantly improve the trade-off between computational efficiency and prediction accuracy. The mathematical foundation of this method is based on function approximation theory and statistical learning theory, with theoretical research topics including guarantees of generalization performance and rigorous analysis of convergence. Particularly, dealing with the "curse of dimensionality" when the input dimension is high is a key practical challenge, and approaches like dimensionality reduction and leveraging sparsity are important.

Shows the basic mathematical framework for applying machine learning models to CAE.

In AI×CAE, the loss function is composed as a weighted sum of a data-driven term and a physics constraint term:

Here, $\mathcal{L}_{\text{data}}$ is the squared error with observed data, $\mathcal{L}_{\text{physics}}$ is the residual of the governing equations, and $\mathcal{L}_{\text{reg}}$ is the regularization term. Adjusting the weight parameters $\lambda$ greatly affects learning stability and accuracy.

The biggest challenge for surrogate models is prediction accuracy outside the range of the training data (extrapolation region). Incorporating physical laws can improve extrapolation performance, but complete guarantees are difficult.

When the dimension of the input parameter space is high, the required number of samples increases exponentially. Efficient sample placement through Active Learning or Latin Hypercube Sampling (LHS) is super important.

• The training data sufficiently represents the physics of the analysis target.
• The relationship between input parameters and output is smooth (if there are discontinuities, domain partitioning is necessary).
• Reducing computational cost is the main purpose; conventional solvers should be used in conjunction for final verification requiring high accuracy.
• If the quality of the training data (mesh-converged, V&V completed) is insufficient, the model's reliability decreases.

Understanding the dimensionless parameters governing the physical phenomenon being analyzed forms the basis for appropriate model selection and parameter setting.

• Péclet Number Pe: Relative importance of convection vs. diffusion. Pe >> 1 indicates convection-dominated (stabilization methods required).
• Reynolds Number Re: Ratio of inertial forces to viscous forces. A fundamental parameter for fluid problems.
• Biot Number Bi: Ratio of internal conduction to surface convection. For Bi < 0.1, the lumped capacitance method is applicable.
• Courant Number CFL: Indicator of numerical stability. For explicit methods, CFL ≤ 1 is required.

For order-of-magnitude estimation of analysis results, dimensional analysis based on Buckingham's Π theorem is effective. Using characteristic length $L$, characteristic velocity $U$, and characteristic time $T = L/U$, the order of each physical quantity is estimated beforehand to confirm the validity of the analysis results.

Choosing appropriate boundary conditions is directly linked to solution uniqueness and physical validity. Insufficient boundary conditions lead to an ill-posed problem, while excessive ones create contradictions.

Yeah, you're doing great! Actually getting your hands dirty is the best way to learn. If you don't understand something, feel free to ask anytime.

Explains numerical methods and algorithms for implementing neural network-type surrogates.

As data preprocessing, normalization/standardization of input features is crucial. Since CAE data have vastly different scales for each physical quantity, it's necessary to appropriately choose methods like Min-Max normalization or Z-score normalization. In selecting the learning algorithm, choose an appropriate method according to data volume, dimensionality, and degree of nonlinearity.

Implementation using the Python ecosystem (scikit-learn, PyTorch, TensorFlow) is common. Key implementation aspects include learning acceleration via GPU parallelization, automatic hyperparameter tuning, and preventing overfitting through cross-validation. Utilizing the HDF5 format is recommended for efficient I/O processing of large-scale CAE data.

It's important to use k-fold cross-validation, Leave-One-Out method, and holdout method appropriately according to the purpose, and to evaluate prediction performance comprehensively using determination coefficient R², RMSE, MAE, and maximum error.

Ensure code quality and experiment reproducibility by introducing version control (Git), automated testing (pytest), and CI/CD pipelines. Strictly enforce dependency library version pinning (requirements.txt) to make rebuilding the computational environment easy. Ensuring result reproducibility by fixing random seeds is also an important implementation practice.