🔢 mathematics

PI-DOSnet: A Physics-Informed Deep Operator-Splitting Network for Evolution Partial Differential Equations

This paper introduces PI-DOSnet, a physics-informed deep operator-splitting network that enables data-free training and stable long-time inference for evolution partial differential equations by integrating physical constraints and operator splitting techniques.

Original authors: Jizu Huang, Yue Qian, Tao Zhou

Published 2026-06-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jizu Huang, Yue Qian, Tao Zhou

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict how a drop of ink spreads in a glass of water, or how a fire spreads through a forest. These are problems described by Evolution Partial Differential Equations (PDEs). For decades, scientists have used traditional math tools to solve these, but they can be slow, especially for complex, high-dimensional problems.

In recent years, AI (specifically deep learning) has offered a new way to solve these. However, most AI methods have a big catch: they need massive amounts of data (like thousands of videos of ink spreading) to learn. If you don't have that data, or if the physics are too complex to simulate enough times to get the data, these AI models fail.

This paper introduces a new AI model called PI-DOSnet. Think of it as a "smart physics student" that doesn't need a library of past examples to learn; instead, it learns directly from the laws of physics themselves.

Here is how it works, broken down into simple concepts:

1. The "Splitting" Strategy (The Chef's Knife)

Imagine you are a chef trying to cook a complex dish that requires chopping vegetables, boiling water, and frying meat. Doing all three at once is messy and hard.

Traditional AI tries to learn the whole dish in one giant, confusing step.
PI-DOSnet uses a technique called Operator Splitting. It acts like a chef who separates the tasks: "First, I'll just chop the veggies. Then, I'll just boil the water."
In the math world, the model splits the complex equation into a linear part (easy, predictable stuff like diffusion) and a non-linear part (tricky stuff like reactions). It handles them separately and then combines the results. This makes the problem much easier for the AI to digest.

2. The "Physics-Informed" Brain (No Textbooks Needed)

Most AI models are like students who only learn by memorizing answer keys (input-output data pairs). If the teacher asks a question not in the book, the student fails.

PI-DOSnet is different. It is Physics-Informed. Instead of memorizing answers, it is given the "textbook" (the actual mathematical equation governing the system) and told, "You must obey these rules."
It checks its own work against the laws of physics. If its prediction breaks a physical law (like conservation of energy), it knows it's wrong and fixes itself. This means it can learn without needing any pre-recorded data. It just needs the equation and the starting conditions.

3. The "Time Machine" Feature (Seeing the Future)

Traditional AI models for these problems usually learn to jump from "Time 0" to "Time 10" in one giant leap. If you ask them what happens at "Time 5," they often get confused or need to be retrained.

PI-DOSnet treats Time as a specific ingredient it can taste. You can tell it, "Show me the state at Time 5," or "Show me Time 100," and it will calculate it directly.
To handle very long times (like predicting the weather for a whole year), it uses an iterative strategy. It predicts the next hour, uses that result as the starting point for the next hour, and repeats. It's like walking a long path by taking one step, then using that new position to take the next step, rather than trying to jump the whole distance at once.

4. Why It's Better Than the Old Ways

The paper claims PI-DOSnet solves three major headaches:

No Data Needed: It doesn't need expensive, pre-calculated datasets. It learns from the math itself.
Stability: When predicting far into the future, many AI models start to "drift" and become nonsense. PI-DOSnet is designed to stay stable, even when taking large "steps" in time.
Efficiency: It is faster and uses fewer computer resources than traditional methods, especially for high-resolution simulations.

Real-World Tests in the Paper

The authors tested this "smart student" on several classic physics problems:

Convection (Moving Fluid): Like wind blowing smoke. The model predicted the smoke's path perfectly for a long time without needing to be retrained.
Diffusion-Reaction: Like a chemical spreading and reacting. It handled this accurately even with large time steps.
Allen-Cahn Equation: This models how materials separate (like oil and water). The model was able to predict the energy of the system correctly, showing it respects physical laws.
Gross-Pitaevskii Equation: This describes quantum particles (Bose-Einstein condensates). The model successfully predicted complex, wavy patterns in these quantum systems.

The Bottom Line

PI-DOSnet is a new type of AI that solves complex, time-changing physical problems by:

Breaking the problem into smaller, manageable pieces (Splitting).
Learning directly from the laws of physics rather than memorizing data (Physics-Informed).
Stepping through time carefully to ensure long-term accuracy.

It offers a way to simulate complex physical systems quickly and accurately, even when we don't have a massive database of past examples to learn from.

Technical Summary: PI-DOSnet: A Physics-Informed Deep Operator-Splitting Network for Evolution Partial Differential Equations

Problem Statement
Evolution partial differential equations (PDEs) describe time-dependent physical systems but are computationally expensive to solve using traditional numerical methods (e.g., FEM, FVM) and existing deep learning approaches. While data-driven operator learning methods (such as DeepONet and FNO) offer efficient inference by learning mappings between function spaces, they typically require large datasets of paired input-output observations, which are costly to generate for time-dependent problems. Furthermore, purely data-driven models often lack explicit enforcement of physical laws, leading to error accumulation during long-time integration and poor generalization to arbitrary time steps. Conversely, Physics-Informed Neural Networks (PINNs) eliminate the need for paired data but often suffer from slow convergence and instability in long-time simulations. There remains a need for a framework that combines the efficiency of operator splitting, the data-free nature of physics-informed learning, and the ability to perform stable long-time inference without retraining on massive datasets.

Methodology: PI-DOSnet
The authors propose PI-DOSnet (Physics-Informed Deep Operator-Splitting Network), a framework that integrates the Deep Operator-Splitting Network (DOSnet) architecture with physics-informed constraints.

Architecture and Operator Splitting: PI-DOSnet adopts the "Autoﬂow" structure of DOSnet, decomposing the PDE evolution into a sequence of Operator Splitting Blocks (OSBs). Each block alternates between linear and nonlinear layers, mimicking the Lie-Trotter or Strang splitting methods.
Physics-Informed Linear Approximation: Unlike DOSnet, which relies on data to learn the linear operator, PI-DOSnet approximates the linear evolution operator $e^{\tau L}$ using a second-order Taylor expansion: $I + \tau L_\theta + \frac{\tau^2}{2}L_\theta^2$ . Here, $L_\theta$ is a learnable convolutional approximation of the linear operator $L$ . This explicit formulation allows the network to compute time derivatives via automatic differentiation.
Input Design: PI-DOSnet explicitly incorporates the time variable $t$ as a network input, alongside the initial condition $u_0(x)$ . This contrasts with standard DOSnet, which only outputs solutions at a fixed terminal time. This design enables the prediction of solutions at arbitrary times $t \in [0, T]$ without increasing the network depth.
Training Strategy: The model is trained using a physics-informed loss function comprising the PDE residual and boundary conditions, eliminating the need for paired input-output data. Spatial derivatives are approximated using finite difference schemes, while time derivatives are computed via automatic differentiation.
Long-Time Inference and Retraining: To predict solutions beyond the training interval $[0, T]$ , PI-DOSnet employs an iterative strategy where the output at time $T$ serves as the input for the next interval. To mitigate error accumulation and oscillations during long-time inference, the authors introduce a retraining mechanism: if the mean square error (MSE) of the residual exceeds a tolerance threshold, the inferred states are added to the training set, and the network is retrained.

Key Contributions

Data-Free Operator Learning: PI-DOSnet is a physics-informed operator learning framework that does not require paired input-output training data, relying solely on the governing PDE and initial/boundary conditions.
Arbitrary Time Inference: By treating time as an explicit input and utilizing Taylor expansions, the network can predict solutions at any time step, overcoming the fixed-time limitation of standard operator splitting networks.
Stability and Efficiency: The paper provides a linear stability analysis showing that the learned linear operator is less sensitive to spatial mesh refinement than traditional discretizations. This allows PI-DOSnet to use larger time steps and fewer network blocks while maintaining stability, reducing both training and inference costs.
Iterative Retraining Strategy: A robust inference protocol is established that dynamically augments the training set with inferred states to maintain accuracy over extended time horizons.

Numerical Results
The authors validate PI-DOSnet on four benchmark problems:

Convection Equation: PI-DOSnet achieved high accuracy (relative $L_2$ error $\approx 3.15 \times 10^{-3}$ ) over long time intervals ( $t=10T$ ). The second-order Taylor expansion was shown to be superior to first-order approximations for long-time inference.
Diffusion-Reaction Equation: The method demonstrated that increasing the time step size per block did not significantly degrade accuracy, contrasting with traditional numerical methods. The optimal balance between accuracy and efficiency was found with 2 blocks.
Allen-Cahn Equation: PI-DOSnet successfully solved both 1D and 2D cases up to $t=10T$ . The retraining strategy effectively reduced inference errors. Notably, the network preserved the energy dissipation law of the system, with predicted energy closely matching reference solutions. Stability analysis revealed that the learned operator's spectral radius grows slowly with mesh resolution, unlike traditional schemes which require drastically smaller time steps (CFL condition) as resolution increases.
Gross-Pitaevskii Equation: The framework handled complex-valued solutions for this nonlinear Schrödinger-type equation, achieving accurate predictions for solitary-wave identification tasks.

Significance and Claims
The paper claims that PI-DOSnet offers a robust and efficient framework for solving evolution PDEs by bridging the gap between data-driven operator learning and physics-informed methods. Its primary significance lies in its ability to perform long-time inference without explicit time-step integration and without requiring large paired datasets. The authors emphasize that the method achieves energy-stable solutions even with large time steps and maintains accuracy across varying spatial resolutions where traditional methods would fail due to stability constraints.

The authors modestly acknowledge remaining challenges, noting that the current framework faces difficulties with PDEs on intricate geometric domains, high-dimensional problems, and those exhibiting low-regularity solutions, which are identified as directions for future work.