MIT Develops AI Model to Enhance Engineering Optimization

Researchers at the Massachusetts Institute of Technology (MIT) have developed a new artificial intelligence approach that accelerates and improves the process of solving complex engineering challenges. Their technique leverages a foundation model—a large, pre-trained AI model—designed to analyze tabular data, making it especially effective for applications in fields such as power grid optimization, materials development, and automotive safety.

Traditional engineering tasks, such as designing safer vehicles or optimizing electrical grids, often require exploring vast combinations of variables. Each trial can be costly or time-consuming, making comprehensive testing impractical. While classic optimization tools like Bayesian optimization are widely used, they struggle as the number of variables grows, since models typically need to be rebuilt or retrained for each new scenario or with every iteration.

The MIT team, led by graduate student Rosen Yu and joined by Cyril Picard and professor Faez Ahmed, has re-envisioned Bayesian optimization by embedding a generative AI system called a tabular foundation model as its core surrogate model. This approach enables the algorithm to automatically prioritize the most influential variables, honing in on optimal solutions more efficiently and without repeated retraining.

A tabular foundation model—described by the team as akin to "ChatGPT for spreadsheets"—is trained on vast datasets composed of rows and columns, rather than language. This is particularly useful in engineering, where much of the data is structured in tables rather than text. Once pre-trained, the model is ready to be deployed for a variety of optimization and prediction tasks, without the need for scenario-specific retraining.

To improve accuracy and efficiency, the researchers introduced a method allowing the model to identify which variables, or combinations of variables, have the greatest impact on the result. For example, in automotive design, only a handful of criteria among hundreds may significantly influence passenger safety. Their algorithm focuses optimization on these critical features, saving time and computational resources.

Testing the approach against five state-of-the-art optimization algorithms over 60 benchmark problems—including power grid design and car crash simulations—the MIT method found optimal solutions between 10 and 100 times faster than its competitors. The results were most pronounced as problem complexity increased, although the authors note that model performance may be limited in domains where training data is sparse, such as certain robotics scenarios.

The research is scheduled for presentation at the International Conference on Learning Representations, reflecting the academic interest in scaling classical scientific methods through AI.

Wei Chen, chair of the Department of Mechanical Engineering at Northwestern University and not involved in the research, described the approach as a promising way to make advanced optimization more accessible in real-world contexts. Looking ahead, the MIT team aims to extend their method to even higher-dimensional problems, such as naval ship design, and further enhance the capabilities of tabular foundation models themselves.

As AI continues to evolve from language and perception tasks to acting as a core component of engineering and scientific workflows, advancements like this could make sophisticated, data-driven design more efficient and widespread.

Source: news.mit.edu