Synthetic Engineering Data: Can AI Learn From Designs That Never Existed?

Artificial intelligence has a data problem.

‍

The performance of many AI systems depends heavily on the quantity, quality and diversity of the data available to them. In engineering and manufacturing, however, good data can be difficult and expensive to obtain.

‍

A company may have decades of CAD models, technical specifications, requirements, test reports and product configurations, but this information is often distributed across different tools, programs and formats. And when engineers are developing a completely new product, an even more obvious problem appears:

How do you train or inform AI with data about something that does not exist yet?

‍

This is where synthetic engineering data becomes particularly interesting.

‍

Instead of relying exclusively on historical projects or physical observations, engineering organizations can increasingly generate useful data computationally — from CAD models, parametric definitions, engineering rules, design constraints and generated configurations.

‍

The idea opens a different perspective on Engineering AI:

What if AI could learn not only from what engineers have already designed, but also from technically meaningful designs that have never existed before?

‍

What is synthetic engineering data?

‍

Synthetic engineering data is artificially generated information that represents plausible engineering objects, configurations, conditions or scenarios without being collected directly from an equivalent physical product or historical project.

‍

Depending on the application, synthetic engineering data can include:

• Images rendered from CAD models;
• Geometric variations of components;
• Alternative system architectures;
• Different component arrangements;
• Routing configurations;
• Virtual manufacturing conditions;
• Artificially generated sensor signals;
• Rare or abnormal operating conditions;
• Automatically labelled datasets;
• Generated design alternatives.

‍

Synthetic data is already becoming an important research topic in manufacturing.

‍

A 2026 review published in Procedia CIRP examined 18 recent studies on synthetic data generation for manufacturing AI. The applications covered defect detection, predictive maintenance, process modelling, material design and human–robot collaboration, using methods ranging from simulation to GANs, VAEs and diffusion models. ( Synthetic Data Generation in Manufacturing: A Review of Methods, Domains, and Emerging Trends — Procedia CIRP, 2026)

‍

But engineering offers an opportunity that goes beyond generating additional images or sensor data.

‍

Engineering already contains structured representations of how products can exist.

‍

CAD geometry, system architectures, parameters, constraints and engineering rules can therefore become foundations for generating new data.

‍

From synthetic images to synthetic designs

‍

One of the clearest industrial applications of synthetic data today is computer vision.

‍

Imagine a company developing an AI system that must recognize hundreds of different industrial components. The conventional approach requires collecting and manually labelling large numbers of photographs of those components.

‍

Synthetic data provides another possibility.

‍

If the component already exists digitally in CAD, images can be generated from the model before large real-world image datasets have been collected.

‍

A 2026 study published in Discover Applied Sciences demonstrated exactly this approach. Researchers automatically generated images of industrial components directly inside a CAD environment, viewing the objects from multiple perspectives.

They then trained a convolutional neural network exclusively with the synthetic images and tested it using photographs of the real components.

The model reached 79.67% accuracy on the real-world test set. The researchers also observed improved internal accuracy as the number of synthetic training images increased, while highlighting that the synthetic-to-real gap remained an important limitation.(Classification of Industrial Components by Artificial Intelligence Using Synthetic Data from a Computer-Aided Design Software — Discover Applied Sciences, 2026 )

‍

This example shows something important. A CAD model is no longer necessarily only the digital definition of an object. It can also become a source from which additional AI training data is produced.

‍

But what if we take this principle one step further?

‍

Instead of generating multiple images of one existing design, what if engineering systems generate multiple designs themselves?

‍

The design space itself can become a dataset

‍

Consider an engineer designing a battery architecture. A traditional project might eventually produce only a small number of mature configurations.

‍

Yet the engineering problem itself may contain thousands of possible combinations: different numbers of modules, different positions, different dimensions, different interfaces, different packaging strategies and different architectural choices.

‍

Most of these combinations will never become physical products.Historically, many of them would never even be explicitly created. Computational design exploration changes this.

‍

When design engineers define a design problem through parameters, constraints and rules, algorithms can generate and evaluate large populations of potential solutions. Each generated configuration contains information.

‍

It may have:

• A specific architecture;
• A set of parameter values;
• Geometric characteristics;
• Interfaces and dependencies;
• Constraint results;
• Performance indicators;
• Reasons for being accepted or rejected.

‍

Viewed through the lens of synthetic data, generated engineering alternatives can therefore become a new source of structured design data.

‍

This interpretation is particularly interesting as research into AI-based CAD generation grows.

‍

A 2026 survey in Advanced Engineering Informatics reviewed 78 publications on learning-based 3D CAD generation. The authors identified three major application areas — CAD generation, CAD reconstruction and CAD reverse engineering — and highlighted design-space exploration and data enhancement among the potential applications of these technologies in mechanical engineering.( Learning-Based 3D CAD Model Generation in Mechanical Engineering: A Survey — Advanced Engineering Informatics, 2026)

‍

The same survey also identifies the availability of large-scale, high-quality CAD datasets as an important challenge for further development of learning-based CAD generation.

‍

This creates an interesting loop:

Engineering data helps AI generate designs.Generated designs can, in turn, create new engineering data.

‍

A design does not have to be manufactured to contain useful information

‍

Industrial engineering data is naturally biased toward the past. Companies accumulate information about products that were designed, selected, manufactured and tested.

‍

That data is extremely valuable. But it represents only part of the possible engineering landscape.

‍

A final product architecture tells us what engineers eventually chose. It does not necessarily tell us about the hundreds of architectures they could have considered.

‍

A validated harness route tells us how one routing problem was solved. It does not describe every geometrically feasible alternative. And a successful product contains relatively little information about the combinations that were rejected before the final architecture emerged.

‍

Synthetic engineering data could expand this view. Consider a routing application.

‍

The historical dataset may contain the approved routes from previous programs.

‍

But if design  engineers can formalize:

• Connection points;
• Exclusion zones;
• Minimum bend radii;
• Separation constraints;
• Component geometry;
• Routing rules;

‍

Then a computational system can potentially explore many additional possibilities.

‍

Those configurations do not have to become manufactured harnesses to contain useful information about the design space.

‍

The same reasoning can apply to:

• Electrical harness routing;
• Pipes and ducts;
• Battery pack architecture;
• Component placement;
• Fuel tank design;
• Propulsion architectures;
• Mechanical integration;
• System architecture;
• Industrial layout.

‍

The goal is not to create thousands of arbitrary designs.

‍

It is to generate meaningful engineering alternatives inside a controlled design space.

‍

Synthetic does not mean fictional

‍

This distinction is essential. The word synthetic can easily create the impression that the objective is simply to produce more data.

‍

But in engineering, more data is not automatically better data.

‍

A generated mechanical configuration that intersects another component may be useless.

‍

A harness route that ignores minimum bend radius is not valuable just because it increases the size of a dataset.

‍

A component layout that violates accessibility requirements might look geometrically plausible while remaining industrially unacceptable.

‍

Synthetic engineering data therefore requires more than variation.

‍

It requires engineering context. That context can include several layers.

‍

• Geometry

Generated configurations need to exist inside the actual geometric environment.

• Engineering rules

Clearance, accessibility, compatibility, separation, manufacturing constraints and domain-specific rules determine whether configurations are meaningful.

• System relationships

Engineering components are interconnected. Interfaces, dependencies and functional relationships matter.

• Parameters and requirements

Synthetic configurations become more valuable when the conditions under which they were generated remain attached to them.

• Evaluation criteria

A generated alternative becomes far more informative when engineers can understand why it succeeds, fails or performs differently from another configuration.

‍

This leads to an important principle: Useful synthetic engineering data is not random data. It is data generated within engineering boundaries.

‍

Why synthetic engineering data matters for AI

‍

The most obvious benefit of synthetic data is scale. Engineers can generate far more examples of specific configurations, components or operating conditions than historical projects can provide.

‍

But its value goes beyond volume.

‍

1. It exposes AI to rare cases

‍

Industrial datasets are often unbalanced. Normal operating conditions are common, while failures and unusual situations are much rarer.

‍

A 2026 systematic review of 86 peer-reviewed studies on synthetic data for predictive maintenance identified data scarcity, class imbalance and the cost of collecting real failure examples as major drivers for synthetic data generation. It also highlighted the relevance of hybrid and physics-informed approaches in safety-critical applications. ( Synthetic Data for Predictive Maintenance — Journal of Intelligent Manufacturing, 2026

‍

The same principle applies more broadly: synthetic generation can deliberately expose AI to cases that historical data does not adequately represent.

‍

2. Data can exist before the product does

‍

New engineering programs have little historical data by definition.

‍

Yet if requirements, CAD environments, system relationships, parameters and engineering rules already exist, synthetic engineering data can be generated before the final product is physically built.

‍

This is particularly powerful in engineering because much of the generation environment already exists inside the product’s digital definition.

‍

3. It reduces labelling effort

‍

Machine-learning datasets often require costly manual annotation.

‍

With synthetic data, labels can be generated automatically because the system already knows what was created, where it is located and which parameters define it.

‍

This makes CAD-generated synthetic datasets particularly attractive for industrial computer vision. ( Discover Applied Sciences — CAD-Based Synthetic Data Study

‍

4. It reveals the unexplored design space

‍

This may be the most important opportunity.

Instead of asking:

- What designs have we built before?

- Design engineers can ask:

- What designs are possible under these requirements and constraints?

‍

Large populations of generated alternatives can reveal which parameters matter most, which architectures repeatedly satisfy requirements, where constraints eliminate solution families and which unusual configurations remain feasible.

Synthetic engineering data therefore becomes more than additional training data. It becomes a way to learn about the engineering problem itself.

‍

Engineering knowledge becomes the generator

‍

This is where synthetic engineering data becomes particularly relevant from Dessia's perspective.

‍

At Dessia, we do not believe that the future of Engineering AI depends simply on applying increasingly large AI models to increasingly large historical datasets.

‍

Engineering has another asset: knowledge.

‍

Engineering rules, parameters, system relationships, product requirements and design logic can be transformed into computational representations.

‍

Once engineering knowledge becomes computable, it can be used not only to analyze existing designs but also to generate and evaluate new possibilities.

‍

Dessia's approach to generative engineering is built around this principle.

‍

Instead of manually creating and testing individual configurations one after another, design engineers can describe a design space using elements such as:

• System architectures;
• Component parameters;
• Engineering constraints;
• Performance targets;
• Manufacturing rules.

‍

Algorithms can then explore this design space and generate multiple feasible configurations. Viewed from the perspective of synthetic engineering data, something interesting happens.

‍

Every generated alternative is no longer only an output.

‍

It can also become information about the design space :
‍

• A generated battery architecture provides information about one combination of parameters and constraints.
• A generated routing solution provides information about another possible path through a geometric environment.
• A generated system architecture represents another technically evaluated combination.

‍

Not every generated configuration needs to become the final product to teach engineers something about the problem.

‍

Could synthetic data become part of the engineering digital thread?

‍

Engineering organizations usually preserve the final product definition, but much of the knowledge created during design exploration is lost: alternatives considered, architectures rejected, trade-offs made and constraints that shaped the final solution.

‍

Computational design exploration could preserve more of that reasoning. Alongside the final design, organizations could retain structured information about explored configurations, evaluated parameters, applied rules, rejected alternatives, results and the links between requirements, constraints and architecture.

‍

Future programs could then inherit not only the geometry of previous products, but also knowledge about the design space around them. Design exploration itself becomes reusable engineering knowledge.

‍

Synthetic engineering data is not about replacing reality

‍

Synthetic data does not replace physical testing, validated CAD models, real products or engineering expertise. Its value lies in creating useful information before every physical example exists.

‍

Research already shows synthetic data being applied to manufacturing computer vision and predictive maintenance, while generative CAD is expanding into design-space exploration and data generation. Together, these developments suggest that future Engineering AI could learn from three complementary sources: what design engineers have built, what design engineers know and what engineering computation can generate.

‍

For industrial organizations, competitive advantage may therefore depend less on simply owning the largest historical dataset and more on the ability to transform engineering knowledge into new, structured and meaningful data.

‍

The most valuable design in a future dataset may be one that was never physically built.

‍

Frequently Asked Questions
‍

What is synthetic engineering data?

Synthetic engineering data is artificially generated information representing engineering objects, geometries, configurations, scenarios or operating conditions. It can be generated from CAD models, virtual environments, parametric models, engineering rules or computational design systems rather than collected exclusively from physical products.

‍

How is synthetic data used in engineering?

Synthetic data can support AI model training, computer vision, predictive maintenance, rare-case generation, data augmentation, design exploration and engineering analysis. Research in manufacturing increasingly investigates its use where real datasets are scarce, expensive or highly imbalanced.

‍

Can synthetic data replace real engineering data?

No. Synthetic data should complement real engineering information. Its usefulness depends on the quality of the models, rules and assumptions used to generate it.

‍

What is the difference between synthetic data and generative engineering?

Synthetic data refers to artificially generated information. Generative engineering refers to computational methods that create engineering design alternatives according to defined parameters, rules and constraints.

Generated engineering alternatives can potentially become a source of structured synthetic design data.

‍