09/2026 – now
Independent project, RAG system
Geoberlin · Berlin
- Designing and building a retrieval-augmented generation system for a surveying engineering firm. Ongoing.
- Type
- Master's thesis
- Partner
- Siemens Energy
- Period
- Feb – Aug 2026
- Team
- Individual
Gas turbines · fracture mechanics · hybrid ML
Physics-informed ML for short-crack growth
A neural network that predicts crack-growth curves for gas-turbine components with manufacturing defects, trained on experiments and held to the shape of an established fracture-mechanics model (IBESS) through its loss.
- Problem
- The physical short-crack model is interpretable but systematically underestimated fatigue life on the experiments; with 56 specimens a free black-box model would overfit.
- My part
- Individual thesis in the Predictive Analytics team at Siemens Energy: model design, physics loss, nested cross-validation and all implementation.
- Result
- −20.1% mean absolute log error vs. the physics baseline (0.965 → 0.771), each of the 56 specimens predicted blind.
- Limit
- 29 of 56 predictions now overestimate life (baseline: 8), the unsafe direction. A feasibility study, not a certified lifetime assessment.
Full case study: model design, validation, limits
Starting point
The physical short-crack model (IBESS) is interpretable but, on the experimental data, it systematically underestimated fatigue life. With only 56 specimens from one material at one nominal stress ratio (R = 0.1), a free black-box model would overfit and lose physical meaning.
What the network predicts
Instead of free values at every point, the network outputs 16 decoder parameters per specimen: 7 density knots each for the depth and surface crack fronts, one terminal surface-to-depth ratio, and one cycle "warp" multiplier. Integrating and normalising the knots guarantees monotonically growing crack curves by construction.
Inputs
Initial defect size (a₀, c₀), max/min stress, specimen width and thickness, defect type (surface or embedded, one-hot), plus derived aspect ratio, stress range, stress ratio (nominally 0.1 throughout, so it carries little information) and log-transformed size and load terms. All standardised. The physics curves are not network inputs: they supervise the curve shape through the loss, and the physics model's threshold cycle sets the time scale that the warp multiplier corrects.
Loss function
Total loss = experimental error (MSE of log cycles) + four physics penalties: curve shape vs. IBESS, terminal crack-front ratio, concavity of the threshold curve (checked at 17 points), and a cap on the warp correction so timing cannot absorb all the error.
Validation
- Nested cross-validation with 5 outer folds, so each of the 56 specimens gets exactly one blind prediction.
- Loss weights tuned by repeated inner validation inside each outer training fold.
- Preprocessing, feature construction and model selection all fitted inside training folds only, to rule out leakage.
- After prediction, physical quantities (geometry, crack closure, J-integral, ΔK, da/dN) are recomputed as a consistency check.
Results
| Metric | Baseline | Thesis | Change |
|---|---|---|---|
| MAE (log) | 0.965 | 0.771 | −20.1% |
| RMSE (log) | 1.171 | 1.062 | −9.3% |
| Late predictions | 8 / 56 | 29 / 56 | +21 |
Errors are measured in natural-log cycle space, so the typical miss shrank from a factor of about 2.6 to about 2.2 in cycles. The error distribution moved from a consistent underestimate to centred on zero, and the median prediction ratio moved close to 1. The smaller RMSE gain shows that typical predictions improved more than the worst ones.
Limits I documented. A "late" prediction overestimates life, which is the unsafe direction, so the model is less conservative than the baseline. One specimen was predicted at more than three times its measured life, showing instability with so little data. A deployable version would need a lower bound, a one-sided loss or a safety factor, and data across more materials and stress ratios. The thesis is a feasibility study, not a certified lifetime assessment.
Public code (2026). The repository re-implements the method on synthetic specimens (the experiments and the partner's physics code are private, and nothing in it reproduces the numbers above). It benchmarks the decoder against simpler corrections of the same physics model on identical cross-validation folds. On that synthetic task a 15-parameter ridge regression on the log lifetime correction beats the 1,280-parameter decoder (MAE in log cycles 0.265 vs. 0.314 over three fold seeds), while the decoder leaves fewer late predictions (26 vs. 31 of 56). The physics penalties change the curve shapes but barely move the lifetime error, because the lifetime is one scalar warp. I kept this negative result; the repository is a method demonstration, not evidence that the architecture is the best lifetime predictor. It also fixes a NaN gradient in the interpolation, refuses to score runs with invalid predictions, and ships a pinned environment, run manifests and CI.