Machine Learning for Characterization and Processing
Unit 14: Integration, limits, and reflection

AI 4 Materials / KI-Materialtechnologie

Prof. Dr. Philipp Pelz

FAU Erlangen-Nürnberg

01. The AI-Driven Materials Scientist

• Final Recap: From atoms to bits, and back to materials discovery.
• We’ve covered:
  • Modalities & Physics (Week 2).
  • Pipeline Integrity (Week 3).
  • Deep Learning (Weeks 5-7).
  • Physics-Informed Models (Week 13).
• Goal: Reflecting on the “Why” and “What’s Next”.

02. Opening the Black Box: Explainability

• In engineering, knowing “what” will happen is not enough.
• Trust: An operator won’t stop a machine because a model said “99% error” without a reason.
• Explainability: Providing human-interpretable evidence for a model’s decision.

03. Sensitivity Analysis (Neuer Ch 7.2)

• Local Perturbation:
  • Input: \(x_i\).
  • Disturb: \(x_i + \epsilon\).
  • Measure: \(\Delta y\).
• If a \(1^\circ\)C change in temperature leads to a 50% change in prediction, the model is either very insightful or very brittle.

04. Global Explanations: SHAP and LIME

• SHAP (Shapley Additive Explanations): Based on game theory.
• Quantifies the “fair share” of each feature toward the final prediction.
• Materials Use: Identifying which chemical element is the primary driver of corrosion resistance in a high-entropy alloy.

05. Causality in the Process Chain

• Correlation != Causality (Recap from Week 1).
• Causal Process Chain:
  • Anomaly detected at Step 10.
  • Cause originated at Step 2.
• Prediction vs. Detection: AI must move earlier in the chain to allow for intervention (Neuer Ch 7.3.3).

06. Materials Ontologies: Digitizing Meaning

• Does the computer know what “Quenching” means?
• Ontology: A semantic map of materials concepts.
• By connecting “Cooling Rate” to “Dislocation Density”, we help the algorithm “reason” like a metallurgist.

07. The Limits of AI in Materials

• Data Bias: Models are only as good as the history they’ve seen.
• Success Bias: Negative experimental results are rarely published, so AI is blind to failure modes.
• Physical Hallucinations: Large models can produce patterns that look plausible but violate thermodynamics.

08. The Ethical Cost of AI

• Training massive models consumes energy.
• In scientific ML, Efficiency is an ethical requirement.
• PINNs (Week 13) are more data-efficient and environmentally “greener” than brute-force NNs.

09. The Role of the Expert in 2030

• AI handles the Tedious: Peak picking, segmentation, data cleaning.
• AI explores the Vast: 10-dimensional process maps.
• The Human handles the Question: What material do we need to solve the climate crisis?
• The Human handles the Interpretation: Does this discovery make physical sense?

10. Conclusion: AI 4 Materials

• AI is not a replacement for domain knowledge; it is an amplifier.
• Your greatest asset is your ability to bridge the gap between physics and the algorithm.
• Final Thought: The best models are those that work in the lab, not just on a benchmark.

11. Recap: Unit 14

• Explainability builds engineering trust.
• Causality is the engine of discovery.
• Be aware of the limits: data bias and hallucinations.
• Next: Your mini-projects and the final exam!

12. References & Further Reading

• Neuer (2024): Ch. 7 (Explainability and Semantics)
• Sandfeld (2024): Ch. 1 (Materials Data Science)
• McClarren (2021): Ch. 10 (Future Directions)