NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·3h ago
AI Research
arXiv CS.AI·May 21·5 min read
Prediction Is Not Physics: Learning and Evaluating Conserved Quantities in Neural Simulators
TL;DR
Neural networks often struggle to preserve conservation laws in physics simulations, as evidenced by a diffusion model trained on Hamiltonian trajectories showing significant discrepancies in energy conservation. This study explores whether these networks can learn or identify globally conserved quantities from physical data.
✦ Why It Matters
Engineers can improve neural network designs to better respect conservation laws in physical simulations.
Key Takeaways
- Explore neural network architectures that prioritize conservation laws.
- Investigate methods for selecting conserved quantities in training.
- Evaluate model performance using physical metrics like energy standard deviation.
- Consider implications of conservation failures in simulation applications.
Full Summary
In physics simulations, preserving conservation laws, such as energy conservation, is crucial for accurate modeling. A diffusion model was trained on Hamiltonian trajectories, which describe the evolution of a physical system over time.
Despite achieving a mean squared error (MSE) close to $10^{-3}$, the model's energy standard deviation was found to be between 7500 and 36000 times larger than the true values. This significant deviation raises questions about the ability of neural networks to learn conserved quantities from physical trajectories.
The research investigates methods for selecting or learning these quantities, aiming to improve the fidelity of neural simulators. Findings suggest that while neural networks can approximate certain dynamics, they may not inherently respect fundamental physical laws.
This has implications for engineers and researchers working on simulations that require adherence to conservation principles.