Uncertainty-Guided Data Augmentation for Engineers | Deep Dive - Yongmin Kwon
This session details a data-efficient method for training engineering surrogate models by using uncertainty quantification (UQ) to guide geometric data augmentation. Instead of random deformations, the approach lets the deep ensemble model identify its own knowledge gaps (epistemic uncertainty), then uses Free-Form Deformation (FFD) to generate new shapes specifically in those uncertain regions. This ensures every expensive simulation run yields maximally informative data, significantly improving model accuracy for a fixed computational budget across domains like structural mechanics and aerodynamics.
The Challenge of Data Scarcity in Engineering Simulation
High-fidelity engineering simulations, while essential for performance verification, are computationally expensive and time-consuming. This cost creates a significant bottleneck, resulting in data scarcity. Deep learning-based surrogate models (metamodels) offer a way to accelerate this process, but their accuracy is contingent on having a large and diverse training dataset—a luxury rarely available in real-world industrial projects.
The conventional response to data scarcity is data augmentation: deforming existing shapes to create new synthetic data. However, this approach often operates on a flawed assumption.
Design Space vs. Performance Space
A key insight presented is that geometric diversity in the design space does not translate to informational diversity in the performance space. The speaker illustrates this with examples where shapes sampled for maximum geometric variety all map to a small, clustered region of performance metrics.
“Designing diversity in the design space simply doesn’t translate to diversity in performance… We need a method that targets performance diversity, not just geometric diversity.”
This disconnect means that expensive FEM or CFD simulations are often wasted on generating data for regions the model already understands well, leading to redundancy and inefficiency.
A New Paradigm: Uncertainty-Guided Augmentation
The core proposal is to flip the traditional approach. Instead of randomly generating shapes and then checking their performance, the model’s own knowledge gaps are used to guide the generation of new data. The central principle is simple yet powerful: “Augment where the model is uncertain, not everywhere.”
This method shifts the focus from form leading function to function leading form, where performance gaps (epistemic uncertainty) directly drive the creation of new geometries. This ensures that every new simulation run provides maximally informative data to the model.
The UQ-Guided Augmentation Pipeline
The framework operates as a continuous, three-step cycle:
- Train and Quantify Uncertainty: A deep ensemble model is trained on a small initial dataset. This model acts as a surrogate and, crucially, provides a spatial map of its own epistemic uncertainty—the gaps in its knowledge.
- Targeted Deformation: The uncertainty map is used to guide geometric augmentation via Free-Form Deformation (FFD). FFD control points in regions of high uncertainty are moved more aggressively, while points in confident regions are moved less. This creates new shapes that are purposefully designed to challenge the model.
- Simulate and Retrain: The newly generated shapes are labeled using high-fidelity simulations. This new, highly informative data is added to the training set, and the surrogate model is retrained. This iterative cycle progressively improves the model’s accuracy and coverage with minimal data.
Key Technical Contributions
The research highlights three core contributions:
- Shared Coordinate System: The uncertainty map and the FFD control points are aligned in the same lattice space, creating a direct, shared coordinate system between uncertainty and deformation.
- Uncertainty-Driven Deformation: The magnitude of deformation is not random or uniform; it is directly proportional to the model’s epistemic uncertainty at each control point.
- Cross-Domain Generality: The framework was successfully applied to both a structural mechanics problem (the DeepJet bracket dataset) and an aerodynamics problem (the DriveAirNet car dataset) without any domain-specific tuning.
Results and Practical Impact
The UQ-guided method demonstrated significant improvements over conventional random augmentation.
- For the DeepJet structural bracket case, the R² value of the surrogate model improved by 14%.
- For the DriveAirNet aerodynamics case, the method was even more effective, outperforming even scenarios where additional ground-truth data was used. The advantage was more pronounced in aerodynamics, likely because performance is highly sensitive to local surface features, which the method naturally targets.
The practical value is clear: engineers can achieve a more accurate surrogate model with the same simulation budget. No additional FEM or CFD runs are required; the existing budget is simply allocated to more informative samples.
Future Vision: Performance-Aware Generative Design
The speaker concludes by looking toward the future, suggesting a powerful synergy between this local refinement technique and the global exploration capabilities of deep generative models.
“Deep generative models are great at proposing globally novel shapes, large creative leaps in geometry. Our framework is good at the opposite, locally refining where the model is uncertain.”
Combining these two approaches—global exploration from generative models and UQ-guided local refinement—presents a promising path toward a truly “performance-aware generative design” system.