MIT researchers have developed a new artificial intelligence approach that could help engineers solve extremely complex design problems much faster, from optimizing power grids to improving vehicle safety. The method has been described as similar to a “ChatGPT for spreadsheets” because it uses a foundation model trained on tabular data, the kind of structured information commonly used in engineering and scientific design.
Many engineering problems involve hundreds of variables, and each test can be expensive or time-consuming. In vehicle safety design, for example, engineers may need to evaluate thousands of parts and design choices to understand how a car would perform in a crash. Traditional optimization tools can struggle when the number of possible combinations becomes too large, especially when each evaluation requires costly simulations or physical testing.
How the New AI Method Works
The MIT team built its approach around Bayesian optimization, a proven method often used when engineers need to find the best design but cannot afford to test every possible option. Bayesian optimization usually works by building a simpler prediction model that estimates which design choice should be tested next, while also considering uncertainty. The problem is that this model often needs to be retrained after each step, which becomes slow and inefficient when the design space has hundreds of variables.
To solve that issue, the researchers used a tabular foundation model as the prediction engine inside the optimization process. Unlike many traditional models, this foundation model has already been trained on large amounts of tabular data and can be used directly without being retrained each time. This makes the optimization process faster and more reusable across different engineering problems.
Why It Is Being Compared to ChatGPT
The comparison to ChatGPT comes from the way the system handles structured data. Large language models are trained on massive amounts of text and can adapt to many language tasks. In a similar way, a tabular foundation model is trained on large amounts of spreadsheet-like data and can adapt to many prediction and optimization tasks involving rows, columns and variables.
For engineers, this is important because much of their work is already stored in tables, simulations, measurements and design records. Instead of creating a new model from scratch for every project, the new approach can use its general understanding of tabular patterns to help identify which variables matter most in a specific design problem.
Finding the Most Important Variables
One of the key strengths of the method is its ability to focus on the variables that have the greatest impact on the final result. In a car design problem, for example, there may be hundreds of design criteria, but not all of them strongly affect crash safety. The AI system can estimate which features or combinations of features are most likely to improve performance, then focus the search on those areas instead of wasting time exploring everything equally.
This ability matters because complex engineering problems often become difficult not only because they have many possible answers, but because most variables are less important than a few critical ones. By identifying the most influential parts of the design space, the system can move toward better solutions much more efficiently.
Faster Results for Harder Problems
In tests on 60 benchmark problems, including realistic engineering-style challenges such as power grid design and car crash testing, the new method found top solutions between 10 and 100 times faster than several widely used optimization algorithms. The researchers also found that the system became especially useful as problems grew more complicated and included more dimensions.
That makes the method promising for demanding fields such as materials development, drug discovery, naval ship design, energy systems and advanced manufacturing. These are areas where engineers and scientists often face enormous design spaces, expensive simulations and limited opportunities to test each possible option.
Limits and Future Potential
The method did not outperform other approaches in every scenario. For example, the researchers found that it was less successful on some problems such as robotic path planning, possibly because that type of task was not well represented in the foundation model’s training data. This suggests that the strength of the method still depends partly on how well the model understands the kind of data it is being asked to analyze.
In the future, the researchers want to improve the performance of tabular foundation models and apply the technique to even larger problems with thousands or millions of variables. If successful, the approach could make advanced design optimization more accessible to engineers who need faster tools for real-world challenges.
Final Outlook
The new MIT method points to a broader shift in how artificial intelligence may be used in engineering. Instead of only generating text or images, foundation models could become powerful engines inside scientific and design tools, helping classical optimization methods work at scales that were previously difficult or impractical.
For engineers, the value is clear. A reusable AI model that can read spreadsheet-like data, identify the most important variables and guide the search for better designs could reduce wasted simulations, speed up development and make complex systems easier to improve.