Physically consistent model learning for reaction-diffusion systems

Erion Morina, Martin HollerarXiv 2025 · DOI · arXiv

Introduction

Reaction-Diffusion (RD) systems serve as the mathematical foundation for modeling complex phenomena across chemistry, biology, and physics. These systems describe how the concentrations of various species evolve over time through the dual processes of local chemical reactions and spatial diffusion. While the advent of Scientific Machine Learning (SciML) has provided powerful tools for learning these dynamics from data, standard "black-box" models, such as vanilla neural networks, often struggle to respect the underlying physics.

Specifically, unconstrained models frequently violate fundamental physical principles, such as the non-negativity of chemical concentrations or the conservation of mass. Furthermore, they may produce mathematically "ill-posed" equations that lack stable or unique solutions, rendering them unreliable for long-term forecasting. This paper introduces a rigorous mathematical framework to address these issues by embedding physical consistency directly into the model architecture.

The Framework: Hard-Coding Physical Laws

A common approach in SciML is to enforce physical constraints via "soft" penalties in the loss function. However, this paper argues for a more robust "hard constraint" approach. By systematically modifying the structure of the neural network $f_\theta$ into a physically consistent class $\bar{f}_\theta$, the authors ensure that the learned model cannot violate physical laws, regardless of the optimization process.

Key Architectural Modifications

  1. The Transition Function ($\chi$): To prevent the prediction of negative concentrations—a property known as quasipositivity—the authors introduce a smooth "switch" function. This function modifies the reaction term near the boundaries of the positive orthant. If a concentration approaches zero, the transition function ensures the reaction can only produce that species, never consume it.
  2. Mass Control and Growth Conditions: To prevent the total mass of the system from growing to infinity, the framework applies mathematical constraints on the growth rates of the reaction terms. This ensures that the system remains dissipative or mass-conserving, aligning with the second law of thermodynamics and mass balance.
  3. Preserving Approximation Capacity: A critical concern when constraining a model is the potential loss of expressivity. The authors provide approximation theorems proving that these modified structures retain the "universal approximation" property. They can still represent any valid physical reaction given sufficient data.

Mathematical Well-Posedness and Convergence

Beyond physical realism, the paper emphasizes the necessity of mathematical well-posedness. A learned model is only scientifically useful if the resulting Partial Differential Equation (PDE) is guaranteed to have a unique, stable solution.

The authors treat the learning process as a regularized inverse problem. By utilizing an "all-at-once" optimization framework, they provide the first rigorous proofs for this class of constrained models:

  • Theorem 15 (Well-Posedness): This theorem establishes that the proposed modifications lead to RD systems that possess unique, global classical (or weak) solutions. This ensures that the model remains stable over arbitrary time horizons.
  • Theorem 20 (Convergence): This represents the paper's primary result, proving that as the volume of clean data increases, the learned parameters of the constrained model converge to the true underlying physical system. This provides a theoretical guarantee that the learning process is consistent.

Discussion and Future Work

This work is primarily foundational and theoretical, providing the necessary proofs to ensure that future empirical applications of SciML to RD systems are grounded in physical reality. By shifting the burden of physical consistency from the loss function to the model architecture, the authors provide a template for more reliable and interpretable scientific AI.

However, several avenues for future research remain. As a theoretical contribution, the paper does not provide an optimized numerical solver or a specific software implementation. Future work will need to address the computational efficiency of these modified terms in large-scale simulations. Additionally, while the framework currently focuses on mass conservation and positivity in RD systems, extending these hard constraints to other physical symmetries—such as rotational invariance or momentum conservation in fluid dynamics—remains a significant challenge for the field.

Summary updated

Abstract

This paper addresses the problem of learning reaction-diffusion (RD) systems from data while ensuring physical consistency and well-posedness of the learned models. Building on a regularization-based framework for structured model learning, we focus on learning parameterized reaction terms and investigate how to incorporate key physical properties, such as mass conservation and quasipositivity, directly into the learning process. Our main contributions are twofold: First, we propose techniques to systematically modify a given class of parameterized reaction terms such that the resulting terms inherently satisfy mass conservation and quasipositivity, ensuring that the learned RD systems preserve non-negativity and adhere to physical principles. These modifications also guarantee well-posedness of the resulting PDEs under additional regularity and growth conditions. Second, we extend existing theoretical results on regularization-based model learning to RD systems using these physically consistent reaction terms. Specifically, we prove that solutions to the learning problem converge to a unique, regularization-minimizing solution of a limit system even when conservation laws and quasipositivity are enforced. In addition, we provide approximation results for quasipositive functions, essential for constructing physically consistent parameterizations. These results advance the development of interpretable and reliable data-driven models for RD systems that align with fundamental physical laws.
BibTeX
@misc{morina2025physically,
  author = {Erion Morina and Martin Holler},
  title = {{Physically consistent model learning for reaction-diffusion systems}},
  year = {2025},
  eprint = {2512.14240},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.14240},
}
Figure