The number of design variables alone does not determine the amount of data needed to train a machine-learning model that predicts computer-aided engineering (CAE) simulation results. According to an article published by Semiconductor Engineering on September 3, 2026, the more important factor is the nature of the response the model is trying to learn: continuous, smooth responses are easier to predict than results that change when a threshold is crossed or that contain rapidly oscillating signals.
The article is based on four vehicle-crash studies. In this type of workflow, each design experiment represents a complete explicit crash simulation, making construction of the training set the most costly part before reaching a model capable of evaluating new designs in seconds instead of hours.
The Number of Variables Is Not a Sufficient Measure
A comparison between two studies shows that dataset size does not directly scale with the number of design variables. A study of the side-threshold structure of an electric vehicle required four times as many experiments as the 25-experiment occupant-safety study, even though the first study used a simpler representation involving the thicknesses of two panels and the position of one of them, while the occupant study varied several variables in the restraint system, such as dummy-to-seatbelt friction, seatbelt-sensor timing, and slip-ring position.
The article also found no direct relationship between model size or loading-case severity and the number of experiments required. The decisive factor emerged when predictors were trained from the same set of 100 experiments using the same three design variables. Prediction of the threshold-structure mass achieved a predictive-capability score of 0.98, whereas the score was 0.64 for predicting the number of damaged battery cells.
The Nature of the Response Determines Learning Difficulty
The explanation for this difference is related to physics. Mass is a smooth, nearly monotonic response affected by panel thickness, whereas the number of damaged cells is an integer value resulting from exceeding a specified stress threshold that causes an internal short circuit. A small geometric change may move a cell from an undamaged state to a damaged state, or may not, forcing the model to learn many separate boundary locations.
The pattern was repeated in a frontal-crash study consisting of 60 experiments. Prediction errors for intrusion locations, compared with finite-element (FE) analysis, ranged from 0.9% to 12.6%, while errors in predicting accelerations at seat-mounting points reached 16.8%. The article relates this to the fact that intrusion is a smoother displacement, whereas acceleration is the second derivative of displacement and contains high-frequency components.
Rich Outputs Do Not Always Mean More Data
The frontal study presents another result that may seem counterintuitive. Predictors for scalar values required all 60 experiments, but predictors for two-dimensional curves and three-dimensional field results were trained on only 20 experiments. These fields include displacement and plastic strain at three integration points through the thickness and at every time step.
The possible explanation mentioned in the article is that a single experiment provides a field predictor with far more training data than a predictor handling one scalar value. An experiment provides a single number for maximum intrusion at the footwell location, but at the same time provides a spatially correlated field across the model and at all time instants.
What Changes in Experiment Planning in Practice?
The proposed practical rule is to design the simulation campaign according to the least-smooth response, not according to the most complex form of the outputs. Before committing the full solution budget, the article recommends beginning with a pilot phase and analyzing the initial data.
- Examine the correlation matrix and scatter plots to determine which variables drive each response.
- Use the predictive-capability score, which captures linear and nonlinear relationships on a scale from 0 to 1, to estimate response difficulty before training a full predictor.
- Read the learning curve linking data size to accuracy to determine whether adding another 20 experiments will help or whether performance has reached a plateau.
- Associate every prediction with its error bounds; if a prediction falls outside the reported bounds or its confidence range, it should be treated as a signal to run a new solution rather than as a reliable answer.
- Use design-variable-importance maps to reduce spending on variables with limited influence and redirect experiments toward influential variables.
The article explains that the difference between a 25-experiment campaign and a 100-experiment campaign may mean four times the explicit-solution time and number of calendar days. Conversely, reducing the data too aggressively leads to validation failure and loss of confidence in the method, while inflating the dataset incurs the cost of experiments that the learning curve would have shown to be unnecessary.
In an applied example, the predictors were used as response-surface models to drive an optimization with a simulated-annealing algorithm over 500 iterations in a few minutes, because predicting injury criteria took seconds instead of running a full simulation for every iteration.
certi.news Reading
The core value here is not a promise that machine-learning models will eliminate simulation, but that they provide a better criterion for allocating the experiment budget: the difficulty of the physical target matters more than the number of inputs or the richness of the output. However, the reported results come from a sponsored blog article and present specific studies without the full methodological details in the available text; therefore, the rules for selecting data size should be regarded as a starting point for experimentation and validation, not as a substitute for independent engineering validation or for solving cases that fall outside the confidence bounds.