MFPC-Net: Multi-fidelity Physics-Constrained Neural Process

Yating Wang, Guang LinarXiv 2020 · arXiv

Abstract

Surrogate modeling in complex engineering domains often faces the challenge of limited high-fidelity data. MFPC-Net (Multi-fidelity Physics-Constrained Neural Process) addresses this by integrating the probabilistic flexibility of Neural Processes with the rigor of physics-informed constraints. By leveraging abundant low-fidelity data alongside sparse high-fidelity observations, MFPC-Net produces surrogate models that are accurate, uncertainty-aware, and physically consistent. Unlike traditional physics-informed neural networks, MFPC-Net generalizes across entire classes of mathematical models without requiring retraining for new scenarios.

Introduction

In fields such as aerospace engineering and climate modeling, high-fidelity simulations are computationally expensive, necessitating the use of fast surrogate models. While low-fidelity data is often cheap and plentiful, effectively combining it with sparse high-fidelity data remains a challenge. Existing approaches like Physics-Informed Neural Networks (PINNs) excel at respecting physical laws but often struggle with uncertainty quantification and generalization to new problem instances.

MFPC-Net bridges this gap by creating a multi-fidelity model that respects underlying physics, provides robust uncertainty estimates, and adapts to new problems within a specific class (e.g., varying source terms in a thermal model) on the fly.

Methods

The MFPC-Net framework integrates three core components: Neural Processes, multi-fidelity data fusion, and physics-based regularization.

The MFPC-Net architecture, illustrating the generation and inference paths alongside the physical constraint and inverse problem modules.

Neural Processes (NP)

At its core, MFPC-Net utilizes the NP architecture to learn a distribution over functions. An encoder transforms a set of context observations into a latent representation $z$. A decoder then uses this representation to make predictions, providing both a mean estimate and a standard deviation to quantify uncertainty.

Multi-fidelity Integration

To enhance predictive power, the network incorporates low-fidelity solutions ($y^L$) as additional inputs alongside physical coordinates ($x$). This allows the model to learn the correlation between fidelity levels, using the low-fidelity data as a global guide for the high-fidelity prediction ($y^H$).

Physics Constraints via Lagrange Multipliers

Physical consistency is enforced by embedding governing equations (e.g., partial differential equations) directly into the loss function. To balance data-driven and physics-based losses, MFPC-Net employs a dynamic optimization strategy using Lagrange multipliers. This ensures the model satisfies physical laws even in regions devoid of high-fidelity data.

Algorithm 1: The training procedure for MFPC-Net, incorporating dynamic Lagrange multipliers for physics constraints.

For inverse problems, such as identifying unknown material properties (e.g., permeability $\kappa$), the architecture includes additional layers to estimate these parameters simultaneously while solving the forward physics.

Results

The performance of MFPC-Net was evaluated on several synthetic engineering benchmarks, demonstrating significant improvements over standard data-driven methods.

Forward Problems

In 2D elliptic PDE scenarios, the inclusion of physics constraints significantly improved accuracy. With only 5 high-fidelity observation points, the relative $L^2$ error was reduced from 13.53% (without physics constraints) to 6.3% (with physics constraints). The model achieved high precision (under 2% error) using only 3% of potential observation points, demonstrating exceptional data efficiency.

Inverse Problems

MFPC-Net successfully identified complex, non-smooth permeability fields, such as "channelized" structures common in geological modeling. These high-contrast features are typically difficult for standard neural networks to resolve, but the physics-constrained NP approach proved robust in identifying these parameters while simultaneously solving the forward problem.

Generalization and Speed

A key advantage of MFPC-Net is its ability to generalize. Once trained on a class of models, it can provide high-fidelity predictions for new scenarios (e.g., a new heat source configuration) near-instantaneously using only a few context points, without the need for retraining. This makes it suitable for real-time optimization and design tasks.

Conclusion

MFPC-Net represents a significant step forward in physics-informed machine learning. By combining multi-fidelity data with probabilistic neural processes and physical constraints, it offers a powerful tool for real-time engineering analysis and uncertainty quantification. Future work will focus on extending the framework to time-dependent systems and improving scalability for high-dimensional 3D simulations.

Summary updated

BibTeX
@misc{wang2020mfpc,
  author = {Yating Wang and Guang Lin},
  title = {{MFPC-Net: Multi-fidelity Physics-Constrained Neural Process}},
  year = {2020},
  eprint = {2010.01378},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2010.01378},
}
Figure