This technology offers a novel machine learning approach for predicting variability in situations where the same input conditions yield varying outputs. It addresses common challenges in fields like manufacturing and materials science, where understanding and managing variability is critical. By predicting both average behavior and the range of possible outcomes, the method provides deeper insight into the associated variability and supports better decision-making in design, optimization, and control of processes and experiments.
Description
Researchers at Oak Ridge National Laboratory and General Motors have developed a data-driven method for predicting variability in systems where repeated experiments or operations result in a range of outcomes under identical inputs—a “one-to-many” scenario. Traditional models typically focus on average outcomes and overlook inherent variability. This invention introduces a new machine learning framework that accounts for variations across replicates and learns how such variability evolves with different input conditions. The approach uses a tailored neural network architecture designed to learn from complex, high-dimensional datasets and deliver not just a point prediction, but insights into the range of possible results. This capability allows engineers and researchers to better understand, predict, and control variability in industrial and experimental settings.
Benefits
- Captures complex one-to-many relationships often missed by conventional models
- Enables variability-informed decision-making in design and process control
- Scalable to large, high-dimensional industrial datasets
Applications and Industries
- Process optimization and quality assurance in manufacturing, including materials joining and additive manufacturing in automotive and aerospace industries.
- Experiment design and uncertainty quantification in materials science, chemistry, and other research fields. Product development and testing in energy storage, power generation systems and beyond—where performance consistency and reliability are critical.
Contact
To learn more about this technology, email [email protected] or call 865-574-1051.
Researchers
Contact
To learn more about this technology, email [email protected] or call 865-574-1051.