Physics-Informed Spatiotemporal Deep Learning for Emulating Coupled Dynamical Systems
summary
A data-driven emulator uses a spatiotemporal LSTM with dual memories, gradient‑highway units and physics-informed loss terms (first and second spatial/temporal derivatives) to predict coupled stress and binary damage fields from only three initial frames. Trained on 41 high‑fidelity simulations (≈1,600 CPU‑hours each) with 10 validation and 10 test runs, the pipeline applies a coarse-grained damage filter, outperforms simpler CNN/LSTM baselines at frame prediction and crack statistics, but systematically underpredicts peak stress magnitudes and slightly overpredicts early damage.
AI-written summary · about
chapters (20)
- 0:00Intro
- 1:09Overview: Emulation of Coupled Dynamical Systems
- 2:41Motivation: Brittle Material Fracture Modeling
- 3:14Microscale High-Fidelity Modeling
- 4:47Bridging Scales through Fast Emulation
- 5:53Data from Microscale Simulations
- 7:33Problem Statement and Quantities of Interest
- 9:59Challenges
- 15:28Approach: Overview
- 17:50Approach: Coarsening Damage Field
- 19:52Model: Modified ST-LSTM
- 21:45Physics-Informed Input and Loss Functions
- 23:46Results: Entire Field Predictions
- 24:56Comparison of Predictive Accuracy
- 26:44Qualitative Comparison on One Frame
- 29:57Discussion: Damage Prediction
- 31:23Discussion: Crack Lengths
- 31:55Discussion: Maximum Stress
- 33:16Future Steps
- 34:13Conclusions
related talks
01h 23mReal-Time Structural Health Monitoring Using Physics-Driven Digital Twins
34mData-Driven Solutions in Engineering
06mExtraction and Interpretation of Deep Autoencoder-based Temporal Features from Wearables for ...
51mThe Brain as an Efficient, Robust Adaptive Learner
52mPredictive Coding Models of Perception
25mLooking beyond LSTMs: Alternatives to Time Series Modelling using Neural Nets
07mFine-Grained Air Pollution Inference with Mobile Sensing Systems: A Weather-Related Deep ...
31mThe Advantages of Deep Learning with Recurrent Neural Networks