Researchers at the Massachusetts Institute of Technology (MIT) and partner institutions have developed a new artificial intelligence framework that improves how machines convert two-dimensional(2D) engineering designs into functional three-dimensional(3D) computer-aided design(CAD) models.
The system helps AI models generate more accurate CAD models while using far less computing power than existing methods. The advance aims to simplify rapid prototyping, reduce development costs, and support engineers working on complex product designs.
Engineers rely on CAD software to create digital 3D models before manufacturing physical products. These digital models allow designers to test strength, durability, and performance through simulations before building real prototypes. Better CAD generation also shortens development time and reduces expensive design revisions.
Smarter Training Method
The research team introduced a framework called Geometric Inference Feedback Tuning(GIFT). Instead of relying primarily on large collections of manually created training data, GIFT enables an AI model to improve by learning from its own mistakes. The researchers recently presented the work at the International Conference on Machine Learning.
The project was led by Giorgio Giannone, a research affiliate in MIT’s Design Computation and Digital Engineering Lab and principal research scientist at Red Hat. The research team also included Anna Claire Doris, Amin Heyrani Nobari, Kai Xu, Akash Srivastava, and Faez Ahmed. The work received support from the MIT-IBM Computing Research Lab.
READ ALSO: Green Rail Revolution: India’s First Indigenous Hydrogen Train Hits the Tracks
Vision-language models combine image recognition with language understanding. In this case, the AI receives a 2D engineering drawing along with text instructions and generates Python code that runs inside CAD software. That code creates a complete 3D model that engineers can inspect and test.
The researchers found that existing AI systems often struggle due to a lack of high-quality CAD training data. Many available datasets lack sufficient variety to teach models to solve difficult engineering design problems. This limits the quality of the generated CAD programs.
Traditional data augmentation methods create additional training examples by making small changes to existing images. Developers may adjust colors, shapes, or object sizes to produce new samples. While effective for many vision tasks, that approach offers limited benefits for CAD code generation.
GIFT employs a different strategy, creating training data based on the AI model’s own performance. The system first evaluates where the model succeeds and where it fails. It then builds new examples that specifically address those weaknesses.
Learning From Errors
The framework asks the AI model to solve the same CAD generation task multiple times simultaneously. Each generated answer is checked to determine whether the CAD code works correctly. This process helps identify patterns in successful and unsuccessful attempts.
Some outputs come very close to producing valid CAD programs but still contain small errors. GIFT automatically fixes these near-complete solutions and saves them alongside fully correct versions. These corrected examples become valuable training material for future learning.
READ ALSO: Skyroot’s Vikram-1 Becomes India’s First-Ever Private Orbital Rocket, Joining US and China
According to lead author Giorgio Giannone, the most useful learning opportunities occur when the model solves a problem only part of the time. If every attempt succeeds, there is little new information to learn. If every attempt fails, the system has little guidance on how to improve.
By focusing on these middle cases, GIFT teaches the model how to correct specific mistakes. The framework also exposes the AI to multiple correct solutions for the same engineering task. This broader experience improves the model’s understanding of CAD programming.
An important feature of the framework is that it works without human correction. Engineers do not need to manually review or repair the generated code before the system learns from it. This automation reduces time and effort during model improvement.
MIT Team Tackles Efficiency
GIFT improves AI performance through a method known as inference-time scaling. Rather than retraining an entire AI model, the framework improves results while the existing model continues generating answers. This approach avoids the high computing costs normally associated with retraining large AI systems.
Users can also choose how much computing power they want to dedicate to the process. This flexibility allows organizations to balance speed, cost, and accuracy according to their available resources. Smaller engineering teams may benefit from lower computing requirements.
In testing, the framework outperformed several competing approaches. AI models using GIFT produced more accurate CAD programs while requiring only about 20 percent of the computing resources used by other methods. The resulting 3D models also matched reference designs more closely.
The improved accuracy matters because even small geometric errors can affect engineering analysis. A design with incorrect dimensions may fail simulation tests or produce unreliable manufacturing results. More accurate CAD models therefore improve confidence throughout the product development process.
The researchers believe this technology can benefit industries that depend heavily on engineering design. Aerospace companies, automotive manufacturers, consumer electronics producers, industrial equipment makers, and robotics developers all rely on accurate CAD models during product development. Faster design generation may help these industries accelerate innovation while controlling engineering costs.
READ ALSO: RTX Brings AI to Pratt & Whitney Engine Inspections for Faster and Smarter Aircraft Maintenance
Future Engineering Applications
The team plans to expand GIFT beyond geometric accuracy. Future versions may teach AI systems to generate CAD programs that also improve manufacturability and product performance. This means AI could eventually suggest designs that are easier to build and perform better under real operating conditions.
The researchers also want to apply the framework to larger AI models and more diverse CAD generation tasks. As these systems continue to improve, they may support increasingly complex engineering projects involving multiple components and advanced manufacturing techniques. This broader capability would increase the usefulness of AI-assisted design across industries.
Associate Professor Faez Ahmed said nearly every physical product begins as a CAD model. He noted that today’s AI tools often generate only simple shapes that are not suitable for real engineering work. He said enabling models to learn from their own mistakes moves AI-assisted design closer to everyday industrial use.
The growing use of artificial intelligence in engineering reflects a wider shift toward digital product development. Companies increasingly rely on virtual testing before building physical prototypes because it saves both time and resources. Better AI-generated CAD models strengthen that digital workflow and help engineers move from concept to production more efficiently.
As manufacturers seek faster development cycles and more efficient design processes, tools that improve CAD generation are becoming increasingly important. Systems like GIFT show how AI can strengthen existing engineering software instead of replacing it. Continued research in this area may make rapid prototyping more accurate, affordable, and accessible for engineering teams around the world.
