New ML framework combines physics with AI for better models
Researchers propose 'Orthogonal Discrepancy Kernels' to merge physics equations with machine learning for incomplete systems
Researchers Swapnil Manna, Timothy J. Rogers, and Lawrence Bull have proposed a novel semi-parametric framework called Orthogonal Discrepancy Kernels, designed to enhance system identification by merging physics-based components with machine learning. This approach decouples discrepancy functions (black-box learning) from physics-based rules (white-box modeling), allowing for interpretable models even with incomplete physics data.
The framework leverages orthogonal Gaussian process regression to balance sparse parameter selection with discrepancy learning, producing more accurate and interpretable models. This method is particularly useful in scenarios where physical laws are partially known or noisy, such as in robotics or signal processing. The paper, published on arXiv (arXiv:2606.21199), highlights its potential to improve predictive modeling in fields where traditional physics-based models fall short.
- Orthogonal Discrepancy Kernels combine physics-based models with machine learning for interpretable predictions
- Uses orthogonal Gaussian process regression to balance white-box and black-box components
- Published on arXiv (arXiv:2606.21199) with potential applications in robotics and signal processing
Why It Matters
Bridges the gap between physics-based and AI-driven modeling, enabling more accurate predictions in data-scarce or noisy environments.