Traditional engineering often hits a wall when dealing with hundreds of design variables, such as the thousands of components involved in vehicle crash safety. Conventional Bayesian optimization, a standard mathematical tool for finding the best configuration in complex systems, struggles because its internal surrogate model must be retrained every time a new design is tested. This process becomes computationally exhausting as the number of variables grows, often making the search for optimal solutions impractical.

To break this bottleneck, MIT researchers introduced a technique called GIT-BO, which leverages a foundation model pre-trained on vast amounts of spreadsheet-style (tabular) data. Much like how a language model understands the relationships between words, this “ChatGPT for spreadsheets” understands the structure of numerical data without needing to be retrained for every specific task. By using this pre-trained intelligence, the system can quickly identify which specific design variables actually impact performance the most.

The results are striking: in realistic engineering benchmarks ranging from power grid optimization to automotive safety, the method outperformed state-of-the-art algorithms by up to 100 times in speed. By focusing the search on high-impact variables rather than exploring every possibility equally, GIT-BO allows engineers to scale their designs to levels previously deemed impossible. This shift suggests that foundation models are evolving from simple text generators into powerful algorithmic engines for scientific discovery and advanced manufacturing.

Traditional engineering often hits a wall when dealing with hundreds of design variables, such as the thousands of components involved in vehicle crash safety. Conventional Bayesian optimization, a standard mathematical tool for finding the best configuration in complex systems, struggles because its internal surrogate model must be retrained every time a new design is tested. This process becomes computationally exhausting as the number of variables grows, often making the search for optimal solutions impractical.

To break this bottleneck, MIT researchers introduced a technique called GIT-BO, which leverages a foundation model pre-trained on vast amounts of spreadsheet-style (tabular) data. Much like how a language model understands the relationships between words, this “ChatGPT for spreadsheets” understands the structure of numerical data without needing to be retrained for every specific task. By using this pre-trained intelligence, the system can quickly identify which specific design variables actually impact performance the most.

The results are striking: in realistic engineering benchmarks ranging from power grid optimization to automotive safety, the method outperformed state-of-the-art algorithms by up to 100 times in speed. By focusing the search on high-impact variables rather than exploring every possibility equally, GIT-BO allows engineers to scale their designs to levels previously deemed impossible. This shift suggests that foundation models are evolving from simple text generators into powerful algorithmic engines for scientific discovery and advanced manufacturing.