Deep Learning — CAE Glossary

It doesn't completely replace them, but its use as a surrogate model (proxy model) is growing. For example, in automotive crash simulation, a single FEM calculation takes hours. But if you pre-train a neural network on hundreds or thousands of simulation results, you can predict stress distributions for new design parameters in milliseconds.

Good observation. Surrogate model accuracy depends heavily on training data quantity and quality. In practice, the mainstream approach is: use surrogate models for design exploration and screening, then validate final candidates with traditional FEM/CFD. For example, screen 10,000 design candidates with a surrogate model, then run full analysis on only the top 100.

PINN stands for Physics-Informed Neural Network. The key feature is incorporating the residual of governing equations into the loss function. Ordinary DL learns only from data, but PINN directly teaches the network that "this solution must satisfy the Navier-Stokes equation" or other physical constraints.

Let's consider steady-state heat conduction. The governing equation is $\nabla^2 T = 0$. In PINN, you compute the second partial derivatives of network output $\hat{T}(x,y)$ using automatic differentiation, then add a penalty term to the loss function to make $\nabla^2 \hat{T}$ approach zero. Physical laws act as constraints even in data-sparse regions, so non-physical solutions are less likely. A common application is inverse problems—for example, inferring material thermal conductivity from sensor data at a few points.

The principle is the same as in image recognition. You reuse the weights from a pre-trained model on a similar physical phenomenon as initial values. For instance, if you have a model trained on 5000 cases of NACA 0012 airfoil aerodynamics, transferring it to NACA 4412 achieves high accuracy with only 200–300 additional data points. From scratch, you'd need about 3000 cases.

Exactly. In the CAE world, generating a single data point takes minutes to hours, so reducing training data generation costs has enormous impact. Transfer learning can reduce that cost by an order of magnitude. However, if the source and target tasks are too dissimilar, you risk negative transfer, so careful assessment of physical similarity is essential.

Great question. There are several approaches, but the most intuitive is SDF (Signed Distance Field) representation. You convert the analysis domain geometry to a grid image where each pixel stores the distance to the object boundary. This becomes an input channel for CNN, and the output channel predicts the stress or temperature contour.

U-Net and Encoder-Decoder architectures are standard. The structure is the same as in image segmentation, but input replaces photos with SDF or boundary condition maps, and output replaces segmentation masks with stress or velocity contours. Skip connections help preserve spatial fine structure. Inference completes in milliseconds, enabling real-time design interfaces.

3D CNN can handle it, but memory grows cubically, making resolutions around $128^3$ the practical limit. Recently, Graph Neural Network (GNN) approaches are emerging, handling unstructured meshes directly as graphs with greater mesh freedom. Point cloud-based techniques using PointNet are also advancing.

The most important is caution regarding extrapolation. Neural networks are accurate within the training data range (interpolation) but predictions break down for unseen conditions—unfamiliar material properties or extreme loads. Physics-based simulators have conservation laws built in and can handle extrapolation reasonably well, but pure data-driven models lack that guarantee. Either incorporate physics like PINN, or explicitly evaluate reliability in extrapolation regions through V&V.

Definitely an ally. Traditional FEM/CFD and deep learning are complementary. FEM/CFD is physically rigorous but slow; deep learning is fast but has limited generalization. The hybrid approachdigital twin and real-time control contexts.