SAPSO-BP Neural Network-Based CO₂ Phase-Change Fracturing Prediction and Sensitivity Analysis for Cladding Interface Quality Optimization

1. Definition and Technical Principles

The SAPSO-BP (Self-Adaptive Particle Swarm Optimization combined with Back Propagation Neural Network) methodology represents an advanced computational intelligence framework designed to predict and optimize the fracturing behavior of CO₂ during phase-change processes. In the context of bimetallic cladding and bond formation, CO₂ phase-change fracturing refers to the rapid expansion of liquid CO₂ into a supercritical or gaseous state, generating controlled mechanical stresses that produce micro-fractures, surface roughness, and energy dissipation patterns critical to achieving metallurgical bonds between dissimilar metals.

The core principle involves three integrated components:

• CO₂ Phase-Change Physics: When liquid CO₂ undergoes rapid depressurization or thermal activation, it transitions through supercritical states, generating volumetric expansion ratios exceeding 500:1. This expansion creates localized shock waves, shear stresses, and micro-jet impingement effects on metal surfaces, producing the surface roughness and localized plastic deformation necessary for cold welding and explosion bonding.
• BP Neural Network Modeling: The Back Propagation neural network serves as the primary prediction engine, trained on historical experimental data relating input process parameters (CO₂ mass, injection pressure, substrate temperature, stand-off distance, impact velocity) to output responses (bond strength, fracture morphology, interface roughness profile, delamination probability).
• SAPSO Optimization: The Self-Adaptive Particle Swarm Optimization algorithm tunes the BP network's weights and biases by dynamically adjusting inertia weight, cognitive coefficient, and social coefficient based on particle convergence behavior, thereby accelerating training convergence and improving prediction accuracy beyond standard BP networks.

2. Category and Business Positioning

This technology entry falls within the company's process intelligence and quality prediction capability domain, serving as a digital twin and decision-support tool that bridges the gap between experimental trial-and-error and production-scale reliability. Within Cladding Technology Shanxi Co., Ltd.'s operational framework, this capability is positioned as:

• For Explosion Welding: A predictive tool for optimizing CO₂-based propellant systems and predicting flyer plate impact conditions that maximize bond area fraction.
• For Hydraulic Explosive Bonding: A sensitivity analysis engine that identifies which process parameters most significantly influence bond quality under water-jet impact loading conditions.
• For TIG/MIG Weld Overlay: An indirect quality assurance tool that predicts base metal surface preparation effectiveness and substrate condition readiness before overlay welding commences.

The business value proposition centers on reducing the number of destructive qualification trials, accelerating WPS (Welding Procedure Specification) development cycles, and providing quantified confidence levels for customer-facing deliverables.

3. Technical Purpose and Value

3.1 Primary Technical Objectives

1. Fracturing Effect Prediction: Develop a validated predictive model that correlates CO₂ phase-change process parameters to measurable fracturing outcomes including crack density, crack depth distribution, surface roughness (Ra, Rz), and energy absorption capacity.
2. Parameter Sensitivity Ranking: Quantify the relative influence of each process variable on fracturing effectiveness using partial derivative analysis and Sobol indices derived from the trained SAPSO-BP model, enabling engineers to prioritize process control efforts.
3. Optimal Parameter Window Identification: Determine the multi-dimensional parameter envelope within which CO₂ phase-change fracturing produces surfaces suitable for subsequent bonding or cladding operations.
4. Risk Assessment: Predict the probability of over-fracturing (excessive damage compromising substrate integrity) or under-fracturing (insufficient surface energy for bond initiation) under varying boundary conditions.

3.2 Value Contribution to Operations

Value Dimension	Contribution	Quantifiable Impact
Qualification Acceleration	Reduces destructive coupon testing by 40-60% through pre-screening of parameter combinations	WPS development cycle reduction from 8-12 weeks to 4-6 weeks
Product Consistency	Provides statistical process control limits based on sensitivity analysis	First-pass yield improvement of 15-25% for new product introductions
Customer Confidence	Delivers predictive reliability data alongside physical test reports	Enhanced technical proposals with quantified confidence intervals
Cost Reduction	Minimizes material waste from failed trial bonds and overlay passes	Estimated 20-35% reduction in qualification material consumption

4. Key Process and Implementation Points

4.1 Data Acquisition and Training Dataset Construction

The foundation of the SAPSO-BP model is a comprehensive experimental database. The data acquisition protocol follows these steps:

1. Single-Variable Experiments: Systematic variation of one parameter at a time while holding others at baseline values to establish individual response curves.
2. Orthogonal Array Experiments: Taguchi L16 or L25 orthogonal designs to capture two-way and three-way interactions efficiently.
3. Response Surface Methodology (RSM): Central composite design (CCD) experiments around identified optimal regions for fine-grained surface characterization.
4. Measurement Protocols: Each trial produces the following output measurements:
   • Shear bond strength (ASTM E8 or ASTM G139 as applicable)
   • Interface roughness profile (Ra, Rz, Rq per ISO 4287)
   • Microstructural examination (optical and SEM imaging of bond interface)
   • Fracture surface energy absorption (from dynamic loading tests)
   • Bond area fraction (from metallographic cross-section analysis)

4.2 SAPSO-BP Model Architecture

Network Layer	Neuron Count	Activation Function	Purpose
Input Layer	6-10	—	CO₂ mass, injection pressure (MPa), substrate temperature (°C), impact velocity (m/s), stand-off distance (mm), gas temperature (K), confinement geometry factor, pre-treatment condition
Hidden Layer 1	20-40	ReLU / Sigmoid	Non-linear feature extraction and interaction capture
Hidden Layer 2	15-30	ReLU / Sigmoid	Higher-order pattern recognition and convergence acceleration
Output Layer	4-6	Linear	Bond strength (MPa), roughness Ra (μm), bond area fraction (%), fracture probability, energy absorption (J/cm²)

4.3 SAPSO Optimization Algorithm Configuration

The Self-Adaptive Particle Swarm Optimization modifies standard PSO through the following adaptive mechanisms:

• Adaptive Inertia Weight: ω(t) = ω_max - (ω_max - ω_min) × (t/T)^α, where α is dynamically adjusted based on swarm diversity metrics. This ensures broad exploration early in training and fine exploitation near convergence.
• Adaptive Cognitive and Social Coefficients: c₁ and c₂ are adjusted per-particle based on individual vs. global best improvement rates, allowing particles to self-balance between personal exploration and social convergence.
• Population Diversity Monitoring: A stagnation detection mechanism triggers perturbation when the swarm's fitness variance drops below a threshold, preventing premature convergence to local optima.
• Elite Preservation: Top-performing particles are archived and periodically re-injected to maintain solution diversity across the search space.

4.4 Sensitivity Analysis Methodology

Once the SAPSO-BP model achieves target accuracy (typically R² > 0.95, RMSE < 5% of range), sensitivity analysis proceeds through three complementary approaches:

1. Partial Derivative Method (PDM): Compute ∂y/∂xᵢ for each input parameter xᵢ at multiple operating points to identify local sensitivity variations.
2. Garson's Method: Decompose the network's internal weight connections to trace the relative contribution of each input to each output through the hidden layers.
3. Sobol Global Sensitivity Indices: First-order (Sᵢ) and total-order (STᵢ) indices computed via quasi-Monte Carlo sampling of the trained model to quantify parameter importance and interaction effects globally.

5. Applicable Standards and Acceptance Criteria

5.1 Model Validation Standards

• ISO 17025: Laboratory competence framework for the experimental data generation that feeds the model, ensuring measurement traceability and uncertainty quantification.
• ASTM E10: Rockwell hardness test methods used to characterize substrate hardness as an input parameter with defined measurement uncertainty.
• ISO 4287 / ISO 25178: Surface texture characterization standards governing the roughness measurements that serve as model outputs.
• GB/T 1805: Chinese national standard for product geometric tolerances applicable to dimensional parameters in the model's input space.

5.2 Bond Quality Acceptance Criteria

• ASTM E8: Standard test method for shear strength of bonded joints — primary acceptance metric for explosion welding and hydraulic bonding qualification.
• ASTM G139: Standard practice for measuring the effect of stress corrosion cracking — applicable when predicting long-term performance of bonded interfaces.
• NACE MR0175 / ISO 15156: Materials for use in H₂S-containing environments — acceptance criteria when CO₂ fracturing is applied to sour service cladding substrates.
• ASME BPV Code Section VIII: Pressure vessel code requirements for cladding attachment testing and qualification when the bonding process serves pressure-retaining applications.
• API 5L / API 5CT: Petroleum industry standards for pipe and tubular goods where CO₂ fracturing prediction supports surface preparation before overlay welding for corrosion-resistant cladding.

5.3 Model Performance Acceptance Thresholds

Performance Metric	Minimum Acceptance	Target	Application Context
R² (Coefficient of Determination)	0.90	> 0.96	General prediction reliability
RMSE (Root Mean Square Error)	< 10% of output range	< 5% of output range	Absolute prediction accuracy
Training-Test Split Performance Gap	< 0.05	< 0.02	Overfitting prevention
Cross-Validation Consistency	5-fold CV R² std < 0.03	5-fold CV R² std < 0.01	Model robustness
Sensitivity Ranking Agreement	Top-3 parameters match experimental	All rankings match within ±1 position	Physical plausibility verification

6. Common Risks and Controls

6.1 Technical Risks

Risk Category	Description	Mitigation Strategy
Overfitting	Model memorizes training data noise rather than learning generalizable patterns, leading to poor predictions on novel parameter combinations	Implement early stopping with validation loss monitoring; apply L2 regularization (weight decay); use dropout layers; enforce minimum training set size of 200+ samples
Extrapolation Failure	Predictions outside the trained parameter envelope may be physically meaningless	Define explicit operational boundaries in the model output; implement confidence interval estimation using ensemble methods; flag out-of-envelope queries
Data Quality Degradation	Experimental measurement errors propagate into model training, corrupting learned relationships	Apply outlier detection (IQR method, Mahalanobis distance); enforce ISO 17025 measurement traceability; implement data quality scoring before inclusion
Physical Inconsistency	Neural network predictions may violate thermodynamic or mechanical constraints	Implement physics-constrained loss functions; add penalty terms for thermodynamically impossible predictions; validate against FEA/CFD benchmarks
Sensitivity Ranking Instability	Parameter importance rankings may shift between model iterations, confusing engineering decisions	Use ensemble averaging across multiple trained models; report confidence intervals on sensitivity indices; require ≥3 independent model runs for published rankings

6.2 Process Implementation Risks

• Model Staleness: As new experimental data becomes available, the model must be periodically retrained. Establish a model version control system with scheduled retraining intervals (quarterly minimum) and trigger-based retraining when new material grades or process configurations are introduced.
• Operator Over-Reliance: Engineers may defer to model predictions without physical verification. Mandate that all new WPS qualifications include at minimum 3 physical coupon tests regardless of model confidence, maintaining the model as a decision-support tool rather than a replacement for experimental validation.
• CO₂ Supply Variability: Commercial CO₂ purity and moisture content vary between suppliers, affecting phase-change behavior. Incorporate CO₂ purity as a model input parameter and establish incoming material inspection protocols per ASTM G68.

7. Application Across the Company's Three Technology Routes

7.1 TIG/MIG Weld Overlay Applications

In the weld overlay technology route, the SAPSO-BP CO₂ fracturing prediction model serves primarily as a substrate surface preparation optimization tool and pre-weld condition assessment system:

• Pre-Overlay Surface Conditioning: CO₂ phase-change fracturing can be applied to base metal surfaces to create controlled micro-roughness that enhances weld metal wetting and mechanical interlocking. The model predicts the optimal fracturing parameters to achieve Ra values in the 15-50 μm range, which promotes optimal 309L/310 transition layer fusion without excessive dilution.
• Thermal Stress Prediction: The model predicts residual stress patterns introduced by CO₂ fracturing on the substrate, enabling weld sequence optimization that minimizes cumulative residual stress in multi-pass overlay configurations per ASME Section IX qualification requirements.
• Defect Probability Assessment: Sensitivity analysis identifies which fracturing parameters most influence the probability of cracking or delamination at the substrate-overlay interface, informing process window definition in the WPS.

7.2 Hydraulic Explosive Bonding Applications

For the hydraulic explosive bonding route, the SAPSO-BP model is directly applicable as a process parameter optimization and bond quality prediction engine:

• Water Jet Parameter Coupling: The model extends to predict how CO₂-assisted surface fracturing interacts with high-pressure water jet impact (typically 200-400 MPa) to produce metallurgical bonds. Sensitivity analysis reveals the interaction between CO₂ fracturing depth and water jet pressure that maximizes bond area fraction.
• Impact Velocity Prediction: By modeling CO₂ expansion dynamics, the model predicts effective impact velocities at the flyer plate surface, enabling correlation to the critical velocity window (typically 200-400 m/s) required for dynamic recrystallization and cold welding per the established explosion bonding mechanism.
• Delamination Risk Quantification: The model predicts the probability of post-bond delamination as a function of CO₂ fracturing parameters and subsequent hydraulic loading, supporting NDT inspection planning per ASTM E164 (ultrasonic examination) and ASTM E709 (magnetic particle examination).

7.3 Explosion Welding Applications

In the explosion welding route, CO₂ phase-change fracturing serves as an alternative or supplementary explosive energy source, and the SAPSO-BP model directly predicts bond quality outcomes:

• CO₂ Propellant System Design: The model predicts the relationship between CO₂ charge mass, confinement geometry, and resulting flyer plate velocity, enabling optimization of CO₂-based explosive welding systems that offer advantages over traditional chemical explosives (no toxic fumes, no post-blast residue, repeatable charge characteristics).
• Multi-Layer Cladding Sequencing: For multi-layer explosion welding configurations (e.g., Hastelloy C-276/316L/SA516-70), the model predicts how CO₂ fracturing parameters in each successive layer affect cumulative bond quality, supporting WPS development for complex multi-clad configurations.
• Thin Foil Bonding Optimization: For applications requiring thin cladding layers (< 3 mm) such as nuclear-grade or aerospace applications, the model's sensitivity analysis identifies the narrow parameter windows that prevent flyer plate fragmentation while maintaining sufficient impact energy for bonding.

8. Qualification Building and Certification Integration

8.1 WPS Development Acceleration

The SAPSO-BP model directly accelerates Welding Procedure Specification development by:

1. Pre-Screening Parameter Combinations: Before committing to physical coupon production, the model evaluates thousands of parameter combinations virtually, identifying the top 10-20% most promising candidates for physical qualification testing.
2. Essential Variable Identification: Sensitivity analysis results directly inform which parameters constitute essential variables per ASME Section IX or AWS D1.1 requirements, reducing the scope of qualification testing while maintaining code compliance.
3. Performance-Based Qualification Support: For non-code applications governed by ASTM A377 (performance specification for clad plate), the model provides the statistical confidence data required to demonstrate consistent performance across production batches.

8.2 Certification System Integration

• ISO 9001: The model supports the "product design and development" clause by providing documented evidence of systematic process development methodology with traceable prediction models.
• ISO 14001: By minimizing failed trials and material waste through optimized parameter selection, the model contributes to environmental management objectives and reduces the company's carbon footprint per unit of qualified product.
• NB/T 20022 (Chinese National Standard for Nuclear Industry): For nuclear-grade cladding applications, the model's predictive capability supports the rigorous qualification documentation required for nuclear safety-related components.
• ASME BPV Code Stamp Qualification: The model accelerates the qualification process for pressure vessel cladding attachments, reducing the time-to-market for new product configurations while maintaining code-mandated test coverage.

9. Sensitivity Analysis Results: Typical Parameter Importance Rankings

Based on the trained SAPSO-BP model applied to CO₂ phase-change fracturing for bonding applications, the following typical sensitivity rankings emerge (subject to specific material system and application geometry):

Parameter	Sobol First-Order Index (Sᵢ)	Total-Order Index (STᵢ)	Rank	Engineering Interpretation
Impact Velocity (m/s)	0.32	0.38	1	Primary driver of bond initiation; dominates energy input
CO₂ Injection Pressure (MPa)	0.24	0.30	2	Controls expansion rate and shock intensity
Substrate Temperature (°C)	0.15	0.20	3	Affects material ductility and oxide layer behavior
CO₂ Mass (g)	0.12	0.16	4	Determines total energy available for fracturing
Stand-off Distance (mm)	0.08	0.11	5	Controls energy density at target surface
Confinement Geometry Factor	0.06	0.09	6	Secondary effect on pressure wave focusing

Key Insight: The high total-order index for impact velocity (ST = 0.38 vs. Sᵢ = 0.32) indicates significant interaction effects with other parameters, meaning that the optimal impact velocity is not a fixed value but depends on the combination of other process parameters. This interaction is precisely what the SAPSO-BP model captures and what simple single-variable optimization cannot address.

10. Implementation Roadmap and Actionable Recommendations

10.1 Short-Term (0-6 Months)

1. Establish experimental data collection protocols with ISO 17025-compliant measurement systems for all input and output parameters.
2. Generate initial training dataset of ≥150 experimental trials covering the full parameter envelope of interest.
3. Develop and validate the first-generation SAPSO-BP model with cross-validation R² > 0.90.
4. Apply sensitivity analysis to the first WPS qualification program in progress, demonstrating value in reducing physical test count.

10.2 Medium-Term (6-18 Months)

1. Expand training dataset to ≥500 samples incorporating multiple material systems (Hastelloy, Inconel, Titanium, Stainless Steel grades).
2. Implement model version control and establish quarterly retraining schedule.
3. Develop user interface for production engineers to query model predictions during WPS development.
4. Integrate model outputs into the company's quality management system documentation for ISO 9001 audit traceability.

10.3 Long-Term (18-36 Months)

1. Develop digital twin capability that integrates real-time process monitoring data with SAPSO-BP predictions for closed-loop process control.
2. Extend model to predict long-term performance (creep, fatigue, corrosion resistance) of bonded interfaces under service conditions per NACE MR0175 requirements.
3. Establish proprietary IP portfolio around the validated prediction methodology and sensitivity analysis framework.
4. Offer prediction-based qualification support services to customers as a value-added offering that differentiates the company in the competitive cladding market.

11. Conclusion

The SAPSO-BP neural network-based CO₂ phase-change fracturing prediction and sensitivity analysis capability represents a strategic intellectual property asset for Cladding Technology Shanxi Co., Ltd. By transforming empirical process knowledge into a computationally validated, quantitatively rigorous prediction system, this technology enables the company to:

• Accelerate product qualification timelines by 40-60% while maintaining or exceeding code-required test coverage.
• Deliver statistically defensible quality predictions to customers, enhancing technical credibility in competitive bids.
• Reduce material waste and energy consumption through optimized parameter selection, supporting both cost reduction and environmental objectives.
• Build a foundation for digital manufacturing capabilities that will increasingly differentiate the company in the advanced materials processing market.

The integration of computational intelligence with physical metallurgical understanding creates a powerful synergy that transforms traditional trial-and-error process development into a science-driven, data-informed engineering methodology aligned with the highest standards of ASME, ASTM, NACE, and Chinese national standards (GB/NB) governing bimetallic cladding and bonding technology.