NP-ODE: Neural Process Aided Ordinary Differential Equations for Uncertainty Quantification of Finite Element Analysis

Yinan Wang, Kaiwen Wang, Wenjun Cai, Xiaowei YuearXiv 2020 · arXiv

Introduction

Finite Element Analysis (FEA) is the industry standard for simulating complex physical systems, ranging from heat transfer in engines to the structural integrity of aircraft. Despite its ubiquity, FEA faces two significant hurdles: high computational cost and inherent determinism. High-fidelity simulations often require hours or days for a single run, and they typically fail to account for the uncertainties present in real-world inputs, such as material properties or environmental conditions.

While data-driven surrogate models like Gaussian Processes (GPs) or standard Neural Networks (NNs) can accelerate simulations, they often lack physical insight. These models do not inherently "know" that the underlying system is governed by differential equations, which can lead to poor generalization when training data is scarce. To bridge this gap, researchers have developed NP-ODE, a novel surrogate model that combines the uncertainty-capturing power of Neural Processes with the physics-aligned structure of Neural Ordinary Differential Equations (Neural-ODEs).

Methodology: The NP-ODE Framework

The NP-ODE architecture utilizes an encoder-decoder framework designed to learn the underlying dynamics of a system while quantifying uncertainty.

  • The Encoder: This component takes observed data points and projects them into a latent feature space. It employs a dual-path approach: a deterministic path with an attention mechanism to identify relevant context, and a stochastic path to capture system-wide uncertainties.
  • The Decoder (Neural-ODE): Unlike traditional decoders that use Multi-Layer Perceptrons (MLPs), NP-ODE uses a Neural-ODE.
Comparison between standard discrete Neural Networks and continuous Neural-ODEs.

As illustrated in Figure 2, while a standard NN treats layers as discrete transformation steps, a Neural-ODE treats the transformation as a continuous process. By parameterizing the derivative of the hidden state, the model effectively learns the differential equation describing the system. This makes the model more parameter-efficient and physically consistent with the FEA it aims to emulate.

Experimental Validation and Pipeline

The model's performance was evaluated through a structured pipeline that begins with FEA data generation and validation against experimental results.

Overview of the NP-ODE pipeline from FEA data generation to prediction and uncertainty quantification (UQ).

The NP-ODE was validated on two distinct datasets:

  1. Synthetic 2D Spirals: Data generated from a known linear ODE, tested with varying levels of Gaussian noise ((\sigma = 0.01, 0.02, 0.1)) to assess robustness.
  2. Tribocorrosion FEA: A real-world dataset comprising 106 simulations of material degradation, involving six input variables (mechanical and electrochemical properties) and one output (corrosion rate).

Key Results

The NP-ODE was compared against standard Neural Processes (NPs) and Gaussian Processes (GPs), yielding several key insights:

  • High Accuracy: In the tribocorrosion case study, NP-ODE achieved a Mean Absolute Percentage Error (MAPE) of 2.12% with only 80 training samples, consistently outperforming NPs (~3.2%) and GPs (2.4% to 4.4%).
  • Robust Uncertainty Quantification (UQ): Unlike deterministic Neural-ODEs, NP-ODE generates a probability distribution. This allows engineers to visualize 95% confidence intervals for every prediction. While GPs often under- or over-estimated uncertainty depending on the kernel, NP-ODE provided reliable intervals that accurately covered ground-truth data.
  • Computational Efficiency: Once trained, the NP-ODE reduced the time required for new predictions from hours (in FEA) to just 0.2 seconds, enabling near real-time analysis.

Conclusion and Future Work

NP-ODE represents a significant step toward physics-informed surrogate modeling. By adopting an ODE-based architecture, the model naturally aligns with the mathematical language of the physical world without requiring hard-coded physical laws.

Future work will address current limitations, such as the computational overhead of the ODE solver during training. Potential improvements include implementing adaptive step-size solvers and extending the framework to handle Partial Differential Equations (PDEs), which would allow the model to scale to even more complex FEA problems.

Summary updated

BibTeX
@misc{wang2020np,
  author = {Yinan Wang and Kaiwen Wang and Wenjun Cai and Xiaowei Yue},
  title = {{NP-ODE: Neural Process Aided Ordinary Differential Equations for Uncertainty Quantification of Finite Element Analysis}},
  year = {2020},
  eprint = {2012.06914},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2012.06914},
}
Figure