Physics-Guided ML for Fabric GSM
Predicting the weight of knitted fabric from machine settings — accurately, and in a way textile engineers can check.
- Role
- Co-author (2nd of 6), presenting author
- Team
- 6 authors — Fab Lab IUB
- Timeline
- Feb 2026 – Sep 2026
- Organisation
- Independent University, Bangladesh
- Code
- Private repository
- Reading time
- 3 min
Research paper accepted at IEEE ICACT 2026. On 5,648 production samples, an Extra Trees model reached a test RMSE of 4.10 g/m² (R² 0.952, MAPE under 0.5%) after adding a physics-derived tightness factor; SHAP and LIME confirmed the model learned textile physics.
Overview
In knitwear factories, GSM — grams per square metre — drives quality control, costing and production planning. It is usually measured after knitting, when fixing a wrong setting means wasted fabric. We asked whether it can be predicted reliably from the machine and yarn settings before production — and whether the predictions make physical sense.
The problem
- Individual features correlate only weakly with GSM, so naive models struggle.
- Black-box accuracy is not enough: engineers need to see why a prediction was made before trusting it on the factory floor.
Context & constraints
We used the Single Jersey subset (5,648 samples, GSM roughly 115–260 g/m²) of a 12,569-sample structured knitting dataset collected from a Bangladeshi textile mill. The work began as a group project in the Artificial Intelligence course (CSE425) and was preceded by a 50-page systematic literature review.
Approach
- Physics-guided features — alongside yarn count (tex), stitch length, machine diameter, gauge and feeders, we added the tightness factor K = √tex / stitch length from Pierce's fabric geometry.
- Broad benchmark — 22 regression models, from interpretable (OLS, ridge, lasso, polynomial ridge, GAM, decision trees) to ensembles (random forest, Extra Trees, gradient boosting, XGBoost, LightGBM, CatBoost) and stacking/voting.
- Honest evaluation — a 70/15/15 split with a fixed seed, scaling only for non-tree models, randomised search with cross-validation, and stability checks across seeds.
- Explainability — SHAP (global and local) and LIME (local), plus partial-dependence checks against textile theory.
Key capabilities
- Tightness-factor feature from Pierce’s geometry
- 22-model benchmark with tuned and untuned variants
- Interpretable baseline (degree-3 polynomial ridge) alongside the best black box
- SHAP and LIME explanations validated against physics
Architecture
Data
Modelling
Explainability
A reproducible modelling pipeline from factory data to physically checked explanations. Select a component to trace its connections.
Key decisions
ADR-01Engineer physics into features, not the loss
- Context
- Physics-informed losses and PINNs are powerful but heavy for tabular factory data.
- Decision
- Encode domain knowledge as a derived feature — the tightness factor — and let standard models use it.
- Trade-offs
- Simpler and portable; removing the feature raises RMSE from about 4.1 to 4.4 g/m².
ADR-02Report an interpretable model next to the best one
- Context
- Factories may prefer a formula they can audit over a forest of trees.
- Decision
- Publish the degree-3 polynomial ridge (RMSE 7.35, R² 0.845) alongside Extra Trees.
- Trade-offs
- Readers can choose between transparency and accuracy with numbers for both.
Challenges
Weak individual correlations
Problem. No single setting predicts GSM well on its own.
Solution. Combining physics-derived interactions with non-linear ensembles recovered strong accuracy.
Quality & delivery
Testing
Five-fold cross-validation put Extra Trees at 2.81 ± 0.47 g/m² RMSE, confirming robustness beyond a single split. Tuning did not beat the default Extra Trees (4.33 vs 4.10), which we report rather than hide.
Outcome
- Test RMSE (Extra Trees)
- 4.10 g/m²
- R² 0.952, MAPE 0.467%
- Cross-validated RMSE
- 2.81 ± 0.47
- g/m²
- Models benchmarked
- 22
- Interpretable to ensembles
- Importance in 3 features
- >86%
- Yarn count, tightness, stitch length
Accepted at the 2026 IEEE 3rd International Conference on Advanced Computing Technologies (ICACT 2026) and presented in the session on Explainable & Trustworthy AI Across Domains. Limitation: Single Jersey fabric from one manufacturer — generalising across fabric types is the next step.
Lessons
- A single well-chosen physical feature can beat a lot of hyper-parameter tuning.
- Explanations are a validation tool: when SHAP agrees with textile theory, engineers start to trust the model.
Building something similar?
Tell me what you’re working on — I reply to every serious message.
Related work
Smart Aquaponics with a Digital Twin
A course research paper (3rd of 4 authors) designing ESP32 sensor nodes for water quality, a Firebase telemetry backbone, a React monitoring and control dashboard, a digital twin for scenario testing and LightGBM-based anomaly prediction.
- Python
- C / C++
- React
- Firebase & Firestore
- ESP32