It is now the era of Digital Transformation! In today’s rapidly evolving engineering landscape, organizations are rethinking how products are designed, simulated, and validated by integrating cutting-edge technologies that reduce cycle time, enhance accuracy, and accelerate innovation.

With the adoption of Xipa’s AI/ML technology, engineering teams can achieve 50–80% time savings across key CAE Model Build processes, enabling faster decision-making and improved operational efficiency. Digital transformation is no longer optional; it has become a strategic imperative.

Drastically cutting down Model-Build among the many stages in the product development process is crucial. The current Model-Build phase remains a critical bottleneck, and accelerating this phase is vital enabler of success, demanding efficiency and scalability.

Even though Xitadel has demonstrated the successful application of AI/ML technology to drastically cut down the time to reduce mesh and connections generation, there are inherent challenges associated with the application of AI/ML.

The Shift Towards Intelligent CAE Model-Build

Traditionally, CAE Model-Build phase of the simulation process has been one of the most time-consuming and error prone steps in Virtual Product development. Engineers spend significant effort preparing geometry, generating analysis-specific and solver specific meshes, applying loads and boundary conditions, and validating simulation models.

However, AI/ML technology with intelligent feature recognition, including Connections, has the transformational potential to impact this current paradigm. By leveraging intelligent automation, organizations can drastically reduce manual intervention, standardize processes, and improve model quality. This empowers engineers to focus on innovation rather than model iteration.

Machine Learning (ML) models can now be custom trained to:

1. Identify geometric features
2. Identify and fix modelling errors
3. Generate Meshes and Connections that
   1. Comply with stipulated meshing specifications
   2. Are solver specific (eg. Crash, Durability, etc., NASTRAN, LS-DYNA, etc.)
   3. Capture organization’s Best Practices for meshing

Current CAE Model-Build Approach has limitations

1. Excessive Manual Effort: Geometry cleanup, defeaturing, and meshing require significant human intervention, increasing turnaround time.
2. Inconsistent Quality: Variation in engineers’ modelling practices leads to inconsistent model quality and results.
3. Limited Automation: Current scripts or templates handle only repetitive tasks; complex geometry still needs manual inputs.

4. Skill Dependency: Heavy reliance on experienced analysts for geometry interpretation and mesh preparation, limiting scalability.

 ML Driven CAE Model Build Enhancement

Challenges to Implementing Recognition and Classification of CAD Data with Machine Learning

1. Data Availability & Labelling: Lack of sufficiently large and labelled datasets of CAD models for ML training.
2. Geometry Complexity: High geometric diversity, continuously evolving features.
3. Feature Standardization: Features across different feature families may appear geometrically similar and can only be differentiated through certain non-visible attributes. The absence or inconsistency of these attributes often leads to misclassification during feature recognition and categorization.
4. Feature Specific Mesh: Developing and refining meshing algorithms that generate attribute- and feature-specific meshes for each feature family, while continuously updating these algorithms to handle varying feature representations and newly encountered variations without impacting previously validated scenarios.
5. Model Generalization: ML models trained on one domain (e.g., Crash) may not perform well on another (e.g., Durability) without retraining.
6. Data Imbalance: The uneven distribution of feature types where some features occur in large numbers while others are smaller in numbers creates data imbalance, which can lead to poor recognition accuracy for underrepresented features.

Approaches to Overcoming the Challenges

1. Data Collection and Augmentation:
   Build a comprehensive dataset by collecting CAD models across multiple domains (Crash, NVH, Durability). Apply data augmentation techniques such as feature transformation, mirroring, and synthetic data generation to improve ML training diversity.
2. Automated Labelling and Annotation Tools:
   Develop semi-automated or rule-assisted labelling workflows to accelerate dataset creation, ensuring consistent and accurate feature tagging for ML training.
3. Feature Attribute Mapping:
   Standardize key geometric and non-geometric attributes (e.g., thickness, material, attachment type) across feature families to help the ML system differentiate between visually similar features.
4. Adaptive Meshing Algorithms:
   Implement learning-based mesh generation rules that automatically adjust parameters based on feature geometry, part type (plastic or sheet metal), and previously validated meshing strategies.
5. Domain-Specific Model Training:
   Train and fine-tune separate ML models for different application domains (Crash, Durability, NVH) while maintaining a shared base model for cross-domain learning.
6. Balancing Feature Representation:
   Use techniques such as data resampling, synthetic feature creation, or class-weight adjustments to reduce the impact of feature imbalance and improve recognition accuracy.
7. Continuous Learning Framework:
   Establish a feedback mechanism where misclassified or unrecognized features are periodically reviewed and added to the training dataset, enabling continuous improvement of recognition accuracy.

Expected Performance Improvements

1. Reduced Model-Build Time:
   With ML training and customization, overall CAE model build time required for meshing setup and rule-based preprocessing can be reduced by 50–80%, enabling faster CAE turnaround.
2. Overcoming Automation Limitations:
   Machine learning helps overcome the constraints of traditional rule-based automation by learning complex feature patterns, enabling “visual” recognition of diverse and unseen geometries with minimal manual intervention.
3. Ease of Operation:
   Offers a clean, minimal GUI with one-time setup and fully automated batch execution, eliminating the need for user intervention.
4. Continuous Improvement:
   ML-driven systems improve over time by learning from past projects, gradually increasing automation accuracy and reducing human intervention.

Connect with us to know more!