Creating a new product often begins with a simple two-dimensional drawing before engineers turn it into a detailed three-dimensional model using computer-aided design (CAD) software.
These 3D models are essential because they allow engineers to test how products such as airplanes, cars and household appliances will perform before they are ever built.
Now, researchers from the Massachusetts Institute of Technology and their collaborators have developed an artificial intelligence (AI) system that can convert 2D designs into accurate 3D CAD models more efficiently by learning from its own mistakes.
The research was presented at the International Conference on Machine Learning (ICML).
Modern AI systems known as vision-language models can already understand images and text together.
Researchers have been adapting these models to generate CAD programs that create 3D objects from simple sketches or images. However, today’s systems often produce models that are too simple or contain errors, making them unsuitable for real engineering work.
The MIT team found that one of the biggest reasons for these limitations is a shortage of high-quality training data. AI systems need many examples to learn how to create reliable CAD models, but collecting and preparing these examples is expensive and time-consuming.
To solve this problem, the researchers developed a new training method called GIFT, short for Geometric Inference Feedback Tuning. Instead of depending on people to create more training examples, GIFT allows the AI to generate new learning material on its own.
The system works by asking the AI to solve the same CAD design problem several times. Some of the answers are completely correct, while others are close but contain small mistakes that prevent the design from working properly. Rather than throwing away these imperfect attempts, GIFT carefully studies them, fixes the errors and adds both the corrected designs and the successful ones to a new training dataset.
By learning from these “almost right” answers, the AI gradually becomes better at solving the types of problems that previously caused it to fail. This allows the system to improve without requiring engineers to manually correct thousands of designs.
The researchers found that these near-perfect examples are especially valuable because they reveal exactly where the AI struggles. If a design task is always solved correctly, there is little new to learn. But when the AI succeeds only some of the time, those mistakes become useful lessons that help strengthen its future performance.
Another advantage of GIFT is that it improves existing AI models without requiring them to be retrained from scratch. Instead, it uses a technique called inference-time scaling, which lets users decide how much computing power they want to spend improving the results. This makes the system flexible enough to fit different budgets and time requirements.
When tested against several competing methods, GIFT produced more accurate CAD programs while using only about 20 percent of the computing resources required by other approaches. The resulting 3D models also matched the intended shapes much more closely.
The researchers say geometry was their first priority because if the shape of a product is wrong, nothing else about the design will work correctly. In the future, they hope to expand GIFT so it can also help improve how easily products can be manufactured and how well they perform in real-world conditions.
As AI continues to become a more important tool in engineering, systems like GIFT could make product design faster, cheaper and more reliable. By allowing AI to improve itself through its own experience, the technology could help engineers develop better products while reducing the time and cost involved in creating new designs.
