Skip to content
Neural Concept Connect 2026 · Bay Area ✓ · Next: Munich (Nov 24) · Tokyo (Feb 3, 2027) Apply for your spot
Neural Concept
Get started

How Many Simulations Do You Need To Train Your Predictive Model?

Anthony Massobrio

CFD Expert & AI for CAE Contributor

2 min read Published

In short

The first question engineering teams ask about machine learning is how many simulations are needed to train the model. There is no single answer: the number varies greatly between use cases, and the dominant factor is the size of the design space the predictive model must cover.

Key takeaways

  • Dataset size is driven by the design space to be explored, more than by any other single factor.
  • There is no universal number, so any answer stated without the use case should be treated cautiously.
  • A graph drawn from project experience gives teams a starting range.
  • The question matters because it determines whether a machine learning approach is affordable at all.

Editor’s note: this article was first published in 2021. Since then, Neural Concept has evolved beyond physics prediction into a platform that brings together the tools, infrastructure and skills AI agents need to understand and act on engineering processes. Engineers can then design better-performing physical products, faster, through AI-native engineering.

When investigating Machine Learning approaches for engineering design cycles, the first question of the engineering team, most of the time, is regarding the number of simulations needed to train the model. There is not a single answer, as this number greatly varies from one use-case to another. It is influenced by numerous aspects however, the main factor is the design space to be explored by the predictive model.

From project experience, the following graph gives an order of magnitude of the dataset size needed to start using machine learning as a predictive model in the next design cycles.

Chart of CAE cases needed to train AI vs number of design parameters, geometric deep learning compared with traditional machine learning
Number of cases needed for AI vs. Number of Design Parameters

Would you like to discuss a particular application? Feel free to get in touch with us: contact@neuralconcept.com

Conversational AI trained on text returns an unverified value, while physics AI trained on a simulated pressure field returns a value with a measured errorTech Reads 27 min

Why human-like AI fails without quantitative grounding

Generative AI vs predictive AI: one input branching into many candidate designs vs many observations converging to one scoreTech Reads 16 min

Generative AI vs Predictive AI in Engineering Contexts

Comparison of System 1 and System 2 thinkingTech Reads 17 min

The future of engineering decision making: trends, tools, careers