Physics Informed Constrained Learning of Dynamics from Static Data

Pengtao Dang, Tingbo Guo, Melissa Fishel, Guang Lin, Wenzhuo Wu, Sha Cao, Chi ZhangarXiv 2025 · DOI · arXiv

TL;DR

Standard Physics-Informed Neural Networks (PINNs) require continuous time-series data to model system dynamics, which is often impossible to collect in real-world scenarios like biology or traffic monitoring. This paper introduces "Constrained Learning" and the MPOCtrL framework, enabling PINNs to accurately estimate dynamic properties (like metabolic flux) using only static, partially observed snapshot data. By combining neural networks with a novel message-passing optimization algorithm, the method enforces physical laws globally and dramatically outperforms existing state-of-the-art flux estimators.

What Problem Does This Paper Address?

Physics-Informed Neural Networks (PINNs) have revolutionized how we model complex systems by baking physical laws (like fluid dynamics or mass conservation) directly into the neural network's loss function. This reduces the need for massive datasets. However, PINNs have a fundamental limitation: they rely on calculating time derivatives, meaning they require continuous time-course data.

In many fields, getting this data is prohibitively expensive or physically impossible. For example, in systems biology, you cannot continuously monitor the metabolic flux inside a living human cell over time; you can only take static, destructive snapshots (like single-cell RNA sequencing). Similarly, in traffic monitoring, you might only have periodic photos or indirect noise measurements rather than continuous high-speed camera feeds. The authors tackle the challenge of approximating these dynamic, mass-carrying fluxes over a network when the available data is entirely static and only partially observed.

Methods

The authors propose a new learning paradigm called Constrained Learning, which shifts the focus away from time derivatives and instead relies heavily on two principles: coherency (strict adherence to physical laws, like steady-state mass conservation) and parsimony (model simplicity to prevent overfitting).

To implement this, they introduce MPOCtrL (Message Passing Optimization-based Constrained Learning):

  • Directed Factor Graphs (DFG): The dynamic system is modeled as a DFG, where nodes represent variables (e.g., metabolites) and edges represent the flux/flow between them.
  • Self-Supervised PINNs: The system uses neural networks to generate initial flux estimates based on the static input data.
  • Message Passing Optimizer (MPO): This is the core algorithmic breakthrough. Standard gradient descent struggles to co-optimize neural network weights and strict physical constraints (like ensuring influx perfectly equals outflux across a massive network). Inspired by Belief Propagation, MPO passes "messages" back and forth between nodes and factors. It iteratively adjusts the network's weights to globally balance the graph, ensuring physical laws are met with a highly efficient $O(MN)$ time complexity.

Datasets

  • Synthetic Data: The authors simulated static observation data using two non-linear formulas (NLF1, NLF2). They applied these to four real-world biological reaction networks (APRN, GSLRN, CMMRN, GGSLRN) and one highly complex synthetic network (SDFG) featuring 204 factors and 18 cycles.
  • Real-World Data: Single-cell RNA sequencing (scRNA-seq) data from the NCBI GEO database (GSE173433). This dataset profiles a pancreatic cancer cell line (Pa03c) under control and drug-inhibition conditions. Crucially, it includes matched mitochondrial assay data, providing a rare experimental "ground truth" for the reaction flux of nine metabolites.

Key Insights and Ideas

  • Static Data Can Reveal Dynamics: You don't strictly need time derivatives to infer system dynamics. If you can map the system to a Directed Factor Graph and strictly enforce global physical constraints (like steady-state flux balance), static snapshots contain enough latent information to reverse-engineer the flow.
  • Message Passing Beats Gradient Descent for Constraints: When a loss function contains complex, interconnected physical constraints, standard gradient descent is unstable and slow. Borrowing Belief Propagation from probabilistic graphical models allows the network to reach a physically valid state much faster.
  • Global vs. Local Balancing: Existing graph-balancing methods only look at immediate parent/child nodes. MPO passes messages globally. When it updates a node, it considers the imbalance of the entire network, preventing the model from getting stuck in physically impossible local minima.

Key Results

  • Massive Accuracy Leaps on Synthetic Data: On complex networks with multiple cycles (like CMMRN), MPOCtrL achieved a cosine similarity of >0.83 between its predictions and the ground truth. Competing state-of-the-art methods (scFEA and Compass) failed to handle the complexity, scoring around 0.25.
  • Real-World Validation: When predicting how metabolic flux changes in cancer cells after a drug is applied, MPOCtrL's predictions strongly correlated with actual lab experiments, achieving a Pearson Correlation Coefficient of 0.75 ($p=0.019$).
  • Baseline Failures: On the same real-world task, the baseline scFEA showed a negative correlation with reality, and Compass predicted zero changes across the board.
  • Efficiency: Thanks to message caching, the MPO algorithm scales linearly ($O(MN)$), making it highly practical for large-scale networks.

Limitations and Future Work

  • Trivial Solutions: The authors note that the model requires careful handling, such as early stopping strategies, to prevent the network from collapsing into "trivial solutions" (e.g., predicting zero flux everywhere, which technically perfectly satisfies the mathematical balance constraints but is physically meaningless).
  • Future Work: The primary experiments focus on Equilibrium Steady State (ESS) systems (where variables are unchanged under a steady state). However, the authors outline theoretical mathematical formulations in the appendix to extend this framework to Non-Equilibrium Steady States (NESS) and Dynamic States (DS), paving the way for modeling systems that are actively fluctuating without time-series data.

Summary updated

Abstract

A physics-informed neural network (PINN) models the dynamics of a system by integrating the governing physical laws into the architecture of a neural network. By enforcing physical laws as constraints, PINN overcomes challenges with data scarsity and potentially high dimensionality. Existing PINN frameworks rely on fully observed time-course data, the acquisition of which could be prohibitive for many systems. In this study, we developed a new PINN learning paradigm, namely Constrained Learning, that enables the approximation of first-order derivatives or motions using non-time course or partially observed data. Computational principles and a general mathematical formulation of Constrained Learning were developed. We further introduced MPOCtrL (Message Passing Optimization-based Constrained Learning) an optimization approach tailored for the Constrained Learning framework that strives to balance the fitting of physical models and observed data. Its code is available at github link: https://github.com/ptdang1001/MPOCtrL Experiments on synthetic and real-world data demonstrated that MPOCtrL can effectively detect the nonlinear dependency between observed data and the underlying physical properties of the system. In particular, on the task of metabolic flux analysis, MPOCtrL outperforms all existing data-driven flux estimators.
BibTeX
@misc{dang2025physics,
  author = {Pengtao Dang and Tingbo Guo and Melissa Fishel and Guang Lin and Wenzhuo Wu and Sha Cao and Chi Zhang},
  title = {{Physics Informed Constrained Learning of Dynamics from Static Data}},
  year = {2025},
  eprint = {2504.12675},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2504.12675},
  howpublished = {arXiv},
}
Figure