Machine Learning in Materials Processing & Characterization
Unit 7: Time-series and Process Monitoring
Prof. Dr. Philipp Pelz
FAU Erlangen-Nürnberg
0. The Third Dimension: Time
Beyond Static Snapshots
Today’s Learning Journey:
- Sequences: Why materials processing is dynamic.
- RNNs: Neurons with memory.
- Vanishing Gradients: The limit of basic recurrence.
- LSTMs & GRUs: Gated memory for long-term dependencies.
- Monitoring: Case studies in Additive Manufacturing.
1. Sequences in Materials Science
Process Logs as 1D Data
- Manufacturing: Temperature, pressure, and flow rate time-series.
- In-situ: Melt pool fluctuations in 3D printing.
- Mechanical: Stress-strain evolution during fatigue.
Why standard NNs fail
- MLPs assume inputs are independent.
- CNNs assume spatial, not temporal, locality.
- We need a model that “remembers” the past state to predict the future.
2. Recurrent Neural Networks (RNNs)
The Concept of Recursion
- Output at \(t-1\) becomes input at \(t\).
- Hidden State \(h_t\): The internal memory of the network. McClarren, Ryan G., (2021)
Unrolling the Graph
- Visualizing the RNN as a sequence of identical layers with shared weights.
- \(h_t = \sigma(W_{hh} h_{t-1} + W_{xh} x_t + b_h)\)
3. The Vanishing Gradient Problem
The Limit of Short-Term Memory
- During training (BPTT), we multiply by the same weights many times.
- If weights are small, gradients shrink exponentially.
- Result: The network “forgets” the early part of the sequence. McClarren, Ryan G., (2021)
The Challenge: Linking a process parameter in Step 1 to a defect in Step 1000.
4. LSTMs and GRUs
Gated Memory
- LSTM (Long Short-Term Memory): Uses a “Cell State” conveyor belt.
- Forget Gate: Throws away irrelevant history.
- Input Gate: Stores new sensor features.
- Output Gate: Decides the next prediction.
GRU (Gated Recurrent Unit)
- A simplified, faster version of LSTM with fewer gates.
- Often sufficient for process monitoring tasks.
5. Case Study: Process Monitoring
Additive Manufacturing (AM)
- Monitoring high-speed photodiode signals.
- Goal: Detect “pore initiation” before it happens.
- RNNs capture the oscillatory signatures of unstable melt pools.
Remaining Useful Life (RUL)
- Predicting when a component (e.g., turbine blade) will fail based on historical vibration logs.
6. Summary & Takeaways
Top Takeaways
- Time-series capture the dynamic physics of materials creation.
- RNNs introduce internal state (memory) into machine learning.
- LSTMs/GRUs are required for long sequences to prevent vanishing gradients.
- Autonomous Lab: Sequential models enable real-time feedback and control.
Exercise Handoff
- Load the Fatigue Life dataset.
- Prepare sliding windows of length \(T=50\).
- Build an LSTM to forecast the next stress peak.
References
Machine learning for engineers: Using data to solve problems for physical systems, Ryan G. McClarren
Example Notebook
Week 7: Process Monitoring — TensileTestDataset
© Philipp Pelz - Machine Learning in Materials Processing & Characterization