Introduction
Finite Element Method (FEM) simulations are essential for automotive safety and crash testing, providing high-fidelity predictions of structural behavior. However, these simulations are computationally intensive, often requiring days to complete for complex vehicle models. To address this, the Finite Element Method Integrated Networks (FEMIN) framework was developed. FEMIN accelerates simulations by replacing non-critical regions of a vehicle's mesh with neural network (NN) surrogates, which predict the forces at the boundary between the NN-replaced region and the remaining FEM mesh.
While effective at reducing computation time, integrating "black-box" neural networks into physics solvers introduces reliability concerns. In a live simulation, there is no ground truth to verify the NN's predictions, making it difficult for engineers to know when the surrogate model is drifting or producing inaccurate results. This paper proposes a probabilistic upgrade to the FEMIN framework by adapting the Deep Variational Bayes Filter (DVBF) to provide real-time confidence metrics alongside force predictions.
Methodology: The Deep Variational Bayes Filter
The authors adapt the DVBF—a state-space model that combines the temporal logic of a Kalman Filter with the expressive power of Deep Learning—to serve as the surrogate model within FEMIN.
Technical Innovations
- Predictive Decoding Step: Standard DVBF architectures are typically generative, where input and output are identical. Since FEMIN requires mapping kinematics (input) to forces (output), the architecture was modified. The model uses an internal transition model to generate a preliminary force estimate, which is then refined by the encoder using physical data.
- Probabilistic Output: Unlike deterministic NNs, this model outputs a distribution characterized by a mean and standard deviation. The mean drives the FEM solver, while the standard deviation serves as a real-time confidence metric (uncertainty estimation).
- Window-based Training: To handle the extensive time-series data of crash events, the model utilizes rolling windows, ensuring scalability for long-duration simulations.
Experimental Setup
The framework was evaluated on two representative automotive load cases generated using the VPS (ESI Group) explicit FEM solver:
- Box-Impact (BI): A sheet metal specimen impacting a rigid cylinder, simulating a standard crash scenario.
- Tension-Compression-Tension (TCT): A cyclic loading test designed to capture complex, history-dependent material behaviors such as plasticity and strain hardening.
Training and validation were conducted using 22 different design variations for each case, involving changes in impact velocities and positions.
Key Results and Insights
The DVBF-integrated FEMIN demonstrated significant improvements in both accuracy and reliability:
- Superior Accuracy: In the Box-Impact case, the DVBF achieved an $R^2$ score of 0.983, outperforming window-based LSTMs (0.969). In the TCT case, it successfully captured complex hardening behaviors that caused standard LSTMs to struggle.
- Reliable Confidence Metrics: The model's predicted standard deviation correlated strongly with actual absolute errors. The Prediction Interval Coverage Probability (PICP) was over 88% for the BI case and 98% for the TCT case, indicating that the true physical values almost always fell within the model's predicted confidence intervals.
- Smoothing and Stability: The DVBF tends to produce smoother force trajectories. This characteristic acts as a form of error correction, leading to more stable and accurate displacement results over the course of the simulation.
- Computational Efficiency: Training the DVBF for the BI case took approximately 38 minutes, a significant reduction compared to the 5+ hours required for a full-sequence LSTM.
Conclusion and Future Work
By transitioning from deterministic to probabilistic surrogate models, the authors have enhanced the trustworthiness of AI-accelerated physics simulations. The ability to provide a "sanity check" via uncertainty estimation is a crucial step toward the industrial adoption of hybrid FEM-AI workflows.
Future research will focus on scaling this approach to full-vehicle crash models and investigating the model's performance when encountering physical phenomena outside its training domain, such as unexpected structural buckling.
Summary updated