Synthetic data is artificially generated data that mimics some statistical properties of real data. In industry it is used to train and test AI models when real examples, such as rare defects or failures, are scarce or sensitive.
NIST defines synthetic data generation as a process that uses seed data to create artificial data with some of the statistical characteristics of the seed data. In manufacturing, the main methods are rendering images from 3D models in simulation, augmenting real images (rotation, lighting, noise), physics-based simulation of sensor signals, and generative models.
The appeal is coverage and labeling. Rare defects and failures can be generated in large numbers, and labels come free because the generator knows what it created. Robotics and vision teams use simulation to train perception and grasping before touching real hardware.
The main limit is the domain gap: models trained only on synthetic data often perform worse on real data. Surface textures, lighting, sensor noise, and real defect shapes are hard to reproduce. Most practical pipelines mix synthetic and real data and validate on real data only. Synthetic data can also encode the generator's assumptions, hiding the very variations that cause problems on the line.
Questions about Synthetic data
Can you train a model on synthetic data only?
You can, but models trained only on synthetic data often perform worse on real data. This is called the domain gap. Most practical pipelines mix synthetic and real data and validate on real data only.
How is synthetic data generated in manufacturing?
The main methods are rendering images from 3D models in simulation, augmenting real images with rotation, lighting or noise, physics-based simulation of sensor signals, and generative models.
When is synthetic data worth using?
When real examples are scarce or sensitive, such as rare defects or failures. These can be generated in large numbers, and the labels come free because the generator knows what it created.
Synthetic data in the digest
Related terms
Computer vision inspection
Computer vision inspection is the automated checking of parts or products for defects using cameras and image analysis, increasingly with deep learning models.
Digital twin
A digital twin is a virtual model of a physical asset, process, or system that is kept in sync with its real counterpart through live data.
Foundation models for robotics
Foundation models for robotics are large models pretrained on broad data, such as images, text, and many robots' demonstrations, that can be adapted to control different robots across many tasks, instead of training one narrow model per task.
Physical AI
Physical AI is AI that perceives, reasons about, and acts in the physical world, typically through robots, vehicles, or machines.
Machine vision
Machine vision is the use of cameras, lighting, and image-processing software to inspect, measure, identify, or guide automatically in industrial settings.