A new study published on arXiv investigates the performance of neural network surrogate models in capturing the full distribution of solutions for stochastic problems, particularly focusing on the tails of the distribution. Researchers found that prediction errors at the extreme ends can be an order of magnitude larger than mean field errors, often due to the networks extrapolating beyond training data. The study compared feed-forward networks and Deep Operator Networks, suggesting that a fully connected neural network trained with a weak form residual loss performed best in handling these extrapolated inputs. AI
IMPACT Highlights limitations of current neural networks in predicting extreme outcomes, crucial for risk assessment in physical modeling.
RANK_REASON The cluster contains an academic paper detailing a numerical study on neural network performance.
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
