AI in simulation: the adoption paradox

85% of the organizations surveyed have begun testing or deploying AI in their simulation workflows. Of those, only 26% say they are satisfied operationally, a sign that scaling from pilot to production is still hard.

Almost half of respondents name the lack of usable data and the difficulty of reaching production as their main barriers to deployment.

Traditional architectures (GNN, PINN, ROM) struggle to scale to industrial meshes and start over on every project, with no way to reuse the data built up across earlier simulation campaigns.

Architectures built on pre-trained transformers cut the amount of data needed by relying on pre-training tailored to a specific area of physics. Even so, they are still very rarely used in industrial simulation workflows today.

5% of the engineers surveyed have experimented with transformer architectures in their simulation workflows.

Manufacturers are aiming beyond faster design tools. They want to move toward horizontal applications that can pool data from many different sources. The use case they rate highest is hybrid modeling, which combines simulation data with field measurements and feeds straight into production.

These applications are only feasible with AI-based simulators that can correlate heterogeneous data streams (3D displacement fields, IoT sensor time series, manufacturing metadata), which traditional tools cannot handle.