Proprietary dual-model system combining improved SRSS physics with custom gradient boosting regression. Trained on 560 Salome/Code_Aster FEM simulations with bolt-type-specific calibrations.
Open-source CAD and mesh generation. Procedural geometry construction via Python scripting. Parametric bolt patterns, plate geometry, and weld definitions.
Advanced nonlinear FEM solver. Material plasticity, contact mechanics, bolt pretension modeling. EN 1993-1-8 material models.
Proprietary in-house development. Dual gradient boosting regressors with physics-informed feature engineering. 13-algorithm competition per bolt type.
Enhanced Square Root Sum of Squares (SRSS) with 7 fitted coupling coefficients models multiaxial load interactions. Each bolt configuration (Pretension μ=0.5, Snug-Tight μ=0.0) employs calibrated coefficients optimized via least-squares regression on FEM ground truth.
Proprietary gradient boosting regressor developed in-house learns systematic deviations between physics baseline and FEM truth. 23 engineered features (14 load ratios + 8 physics-derived + 1 SRSS prediction) input to ensemble model. 13 competing algorithms (Ridge, ElasticNet, SVR, GBM, RF, KNN, GPR, etc.) evaluated via 5-fold cross-validation per bolt type.
Final resistance combines calibrated physics baseline with learned correction term. Bolt-type-specific routing ensures Pretension and Snug-Tight connections receive appropriate coupling coefficients and ML corrections. Classification layer (Random Forest, 200 trees) maps continuous resistance to PASS/FAIL categories with 96.4% accuracy.
HEA120-HEA120 beam-column connection with 6×M20-8.8 bolts, 15mm S355 endplates. Geometry procedurally generated via Salome Python API for Code_Aster meshing.
Parametric Python script generates HEA120-HEA120 connection. Bolt holes, endplate thickness, weld geometry defined procedurally.
Salome automated meshing: 8-node hexahedral elements for plates/beams, refined mesh near bolt holes and welds.
Nonlinear FEM analysis: material plasticity (ε,pl ≤ 5%), contact mechanics, bolt pretension (Pretension: 155kN, Snug-Tight: 0kN).
Extract resistance %, bolt forces, plate von Mises stresses, plastic strains from .rmed output files.
Random Forest classifier (200 trees, max_depth=15) achieves 96.4% accuracy across 560 FEM cases. Only 6 borderline misclassifications (100-105% zone).
Hybrid model reduces MAE from 14.7% (physics alone) to 6.95% (physics+ML). 53% error reduction via custom gradient boosting residual correction.
Borderline design zone (85-115% resistance) shows 8.1% average error. FEM verification mandatory for final approval in this range.
| Resistance Range | Avg Error | Max Error | Recommendation |
|---|---|---|---|
| Fail (<85%) | 4.2% | 12.8% | High confidence - no FEM needed |
| Critical (85-115%) | 8.1% | 15.2% | FEM verification required |
| Pass (115-200%) | 3.9% | 9.4% | High confidence - safe for preliminary design |
| Deep Pass (>200%) | 11.6% | 25.3% | Use classification only (PASS reliable, R% less accurate) |
ConnectAI is a preliminary design screening tool, not a replacement for code-compliant FEM analysis. Follow this protocol for responsible engineering practice:
Use AI predictions to rapidly evaluate 50-100 design alternatives. Eliminate clear failures (R < 85%) and identify promising candidates (R > 115%). No FEM needed at this stage.
Iterate quickly with AI (10-20 variants/day), then run Salome/Code_Aster FEM on top 2-3 finalists. Verify borderline cases (95-105%) with full nonlinear analysis.
Always verify critical connections with full FEM before approval. High-consequence structures, borderline resistances, or loads near capacity limits require Code_Aster validation and PE sign-off.
Test real HEA120 connection with Salome-generated geometry and custom ML predictions