Bond Graphs for multi-physics informed Neural Networks for multi-variate time series

Alexis-Raja Brachet, Pierre-Yves Richard, Céline HudelotarXiv 2024 · DOI · arXiv

Abstract

Real-world engineering systems are inherently multi-physical, yet most Physics-Informed Machine Learning (PIML) methods are tailored for single-domain problems. This paper introduces the Neural Bond graph Encoder (NBgE), a novel architecture that bridges classical multi-physics modeling with deep learning. By translating Bond Graphs—a domain-agnostic representation of energy flow—into a graph neural network framework, NBgE creates "physics-aware" data representations. Evaluated on both simulated and real-world datasets, NBgE consistently improves forecasting accuracy and data efficiency across various standard neural architectures.

Introduction

Physical-Informed Machine Learning (PIML) has gained significant traction, but existing approaches like Physics-Informed Neural Networks (PINNs) often struggle with systems that span multiple physical domains (e.g., electrical, mechanical, and fluid). These systems require a unified language to describe energy conservation and causality.

The NBgE architecture integrates Bond Graph knowledge with raw data to produce physics-informed representations for downstream prediction tasks.

The proposed Neural Bond graph Encoder (NBgE) addresses this gap by acting as a task-agnostic encoder. It extracts features from raw multi-physics data that respect underlying physical laws, providing a robust inductive bias for subsequent time-series forecasting tasks.

Methodology

The NBgE workflow consists of three primary stages designed to transform physical knowledge into a learnable graph structure:

  1. Bond Matrix Construction: The system's Bond Graph—a diagram representing power exchange between components—is first encoded into a structured matrix. This captures the fundamental "skeleton" of the physical system.
  2. Dual Graph Generation: Standard Graph Neural Networks (GNNs) typically operate on nodes, whereas Bond Graphs store critical information on edges (bonds). To resolve this, the authors transform the Bond Graph into a "Dual Graph." In this representation, physical variables (e.g., voltage, force) are mapped to nodes, while physical relations (e.g., resistance, inertia) are mapped to edges.
  3. Bond Graph Convolution (BGC): The core engine of NBgE performs message passing in the frequency domain using Discrete Fourier Transforms (DFT). This approach simplifies complex temporal operations like integration and derivation into simple multiplications, allowing the network to learn physical constraints more effectively.
The translation process from a physical Bond Graph to a Bond Matrix and finally to a Dual Graph suitable for GNN processing.

Experimental Evaluation

The researchers evaluated NBgE on two distinct multi-domain systems:

  • Direct Current (DC) Motor: A simulated environment involving electrical and mechanical domains where the physics is fully known.
  • Human Respiratory System (RS): A real-world public dataset involving fluid mechanics (airflow) and mechanical deformation (chest/abdomen movement). This represents a "partially-known" system with complex, non-linear dynamics.

NBgE was tested as a "booster" for four common baseline models: Linear, MLP, GraphSAGE, and Transformers.

Key Results and Discussion

Experimental results demonstrate that NBgE provides a consistent performance uplift across all tested models and datasets. For the DC Motor, the Mean Absolute Error (MAE) of a standard Linear model dropped from 3.91 to 1.64 when informed by the NBgE encoder.

Key insights include:

  • Data Efficiency: NBgE was particularly effective in scenarios with limited historical data, helping large models like Transformers generalize better by providing a physical inductive bias that prevents overfitting.
  • Shape Preservation: Using Soft Dynamic Time Warping (SDTW) as a metric, the authors found that NBgE-informed models produced forecasts that more closely followed the physical "shape" and rhythm of real signals compared to their non-informed counterparts.
  • Domain Agnosticism: By leveraging the universal language of Bond Graphs (treating various physical quantities as "effort" and "flow"), the architecture remains robust across diverse physical domains without requiring manual redesign.

Conclusion

NBgE represents a significant step toward unifying classical engineering principles with modern deep learning. While currently focused on 1D bond graphs and linear physical relations, the framework's modularity allows for future expansion into non-linear dynamics and hybrid architectures. By relaxing rigid symbolic constraints, NBgE allows neural networks to learn unmodeled effects while remaining grounded in the laws of physics.

Summary updated

BibTeX
@misc{brachet2024bond,
  author = {Alexis-Raja Brachet and Pierre-Yves Richard and Céline Hudelot},
  title = {{Bond Graphs for multi-physics informed Neural Networks for multi-variate time series}},
  year = {2024},
  eprint = {2405.13586},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2405.13586},
}
Figure