🔢 mathematics

Data-integrated neural networks for solving partial differential equations

This paper proposes DataInNet, a novel architecture that integrates diverse data sources and residual physical learning into a unified neural network framework to efficiently solve partial differential equations with high accuracy while strictly adhering to physical constraints.

Original authors: Jiachun Zheng, Yunqing Huang, Nianyu Yi, Yunlei Yang

Published 2026-02-02

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Jiachun Zheng, Yunqing Huang, Nianyu Yi, Yunlei Yang

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 teach a robot to predict how water flows through a complex pipe system. Traditionally, you would tell the robot the rules of physics (like "water can't disappear") and give it a list of "penalty points" to deduct whenever it breaks a rule. This is how current AI methods, called Physics-Informed Neural Networks (PINNs), usually work. They learn the rules by trial and error, hoping to eventually get the penalty points down to zero.

The paper introduces a new, smarter way to do this called DataInNet. Instead of just giving the robot a list of rules to follow, DataInNet builds the rules directly into the robot's brain.

Here is a simple breakdown of how it works, using everyday analogies:

1. The Problem with the Old Way (PINNs)

Think of the old method like teaching a student to solve a math problem by only showing them the answer key at the very end. You say, "Here is the equation, and here are the boundary rules. If your answer is wrong, you get a bad grade." The student has to guess the whole solution from scratch, checking against the rules constantly. This is slow, and the student often gets stuck on the hard parts (like high-frequency waves or sharp corners).

2. The New Solution: DataInNet

The authors propose splitting the "student" into two specialized helpers working together:

Helper A: The Data Integrator (The "Fact-Checker")
This helper is a special network designed specifically to hold all the known facts: the starting conditions, the boundaries, and the source of the problem.
- The Analogy: Imagine a chef who is given a pre-made sauce (the boundary conditions) and a list of ingredients (the source terms). Instead of the chef trying to figure out what the sauce should taste like, the Data Integrator is the sauce. It doesn't guess; it simply holds the known physical data and makes sure it is always present in the final dish.
- The Innovation: In the old method, the AI had to learn that "at the edge of the pipe, the water stops." In DataInNet, the AI is built so that it physically cannot output a result where water flows out the edge. It structurally excludes impossible answers.
Helper B: The Residual Learner (The "Problem Solver")
This is a standard neural network. Its job is much easier now. It doesn't have to learn the boring, known parts (like the boundaries) because Helper A is already handling those.
- The Analogy: If Helper A has already prepared the crust of a pizza, Helper B only needs to focus on arranging the toppings in the middle. Because it doesn't have to waste energy re-learning the crust, it can focus entirely on the tricky, unknown parts of the solution.

3. The "Global Input" Trick

The paper also highlights a clever way of feeding data to the AI.

The Old Way (Local Input): Imagine telling the AI, "Remember the starting temperature, but only when you are looking at the exact moment time starts (t=0)." This is like only showing a student the first page of a book and expecting them to remember the plot for the rest of the story.
The New Way (Global Input): The new method tells the AI, "Here is the starting temperature for every point in space and time, even if you are looking at the middle of the story." It's like giving the student a map that shows the starting point everywhere they look. This helps the AI understand the "big picture" much faster and more accurately.

4. What Did They Find?

The authors tested this new "two-helper" system on several difficult math problems (like waves vibrating very fast or heat spreading through weirdly shaped rooms).

The Result: The new system was much more accurate and stable than the old methods.
The Numbers: For some very difficult problems involving high-frequency waves (which are usually hard for AI to catch), the new method reduced errors to incredibly tiny levels (about 1 in a million).

Summary

In short, DataInNet changes the game from "AI, please guess the answer and try not to break the rules" to "AI, here is the framework of the rules built into your structure; now just fill in the missing pieces." By hard-wiring the known facts (like boundaries and starting points) directly into the network's architecture, the AI has a much smaller, easier job to do, leading to faster and more accurate solutions.

Technical Summary: Data-Integrated Neural Networks for Solving Partial Differential Equations

Problem Statement
Deep neural networks (DNNs) have become a prominent tool for solving partial differential equations (PDEs), yet they face persistent challenges, particularly "spectral bias" (the tendency to learn low-frequency components before high-frequency ones) and training instability. While Physics-Informed Neural Networks (PINNs) attempt to address these issues by embedding physical laws into the loss function as soft constraints, they often struggle with low numerical accuracy and the difficulty of learning complex physical laws from scratch. Specifically, PINNs treat source terms, initial conditions, and boundary conditions merely as penalty terms in the loss function, requiring the network to learn these known physical components alongside the unknown solution.

Methodology: DataInNet Architecture
To address these limitations, the authors propose Data-Integrated Neural Networks (DataInNet). Unlike traditional PINNs, DataInNet reformulates the problem by directly integrating known physical data (source terms, initial conditions, and boundary conditions) into the network architecture itself, rather than relying solely on loss function penalties.

The architecture consists of two distinct subnetworks:

Data Integration Neural Network ( $u^d_{nn}$ ): This network is specifically designed to accept and fuse various types of physical data. It employs a multi-branch parallel structure where specific branches ( $Q_{n,i}$ ) incorporate specific data types (source term $f$ , initial condition $I$ , or boundary condition $B$ ) via learnable weights and gating mechanisms. These branches interact with an intrinsic feature learning branch ( $G_n$ ) through element-wise multiplication. The network inherently excludes function classes that violate known physical constraints, effectively narrowing the solution search space. To handle large magnitude data (e.g., high-frequency source terms), a normalization strategy is applied.
Auxiliary Neural Network ( $u_{nn}$ ): A standard fully connected network tasked with learning the residual physical information not captured by the data integration network.

The final approximation is the sum of these two components: $u^*_{nn} = u_{nn} + u^d_{nn}$ . The model is trained by minimizing a loss function that measures the residual of the PDE and the boundary/initial conditions against this combined output.

A key methodological innovation highlighted in the numerical experiments is the Global-Input Strategy. Instead of feeding initial condition data only at $t=0$ (Local-Input), the DataInNet treats initial and boundary conditions as fields defined over the entire spatiotemporal domain. This allows the model to perceive global physical constraints at every interior point during training, significantly improving convergence and accuracy.

Key Contributions
The paper outlines three primary contributions:

Data Integration Architecture: A novel paradigm that moves beyond soft constraints in the loss function by embedding physical priors (source terms, initial/boundary conditions) directly into the network structure.
Reduction of Solution Space: By structurally embedding known constraints, the method inherently excludes solutions that violate physical laws. This focuses the optimization process on feasible function classes, enhancing accuracy, computational efficiency, and generalization.
Dual-Branch Structure: The explicit separation of known information (handled by the data integration network) and unknown residual information (handled by the auxiliary network) facilitates collaborative learning between known physics and the solution to be found.

Numerical Results
The authors evaluated DataInNet on several challenging PDEs, comparing performance against standard PINNs and varying network configurations (number of modules, training points, and activation functions):

Heat Equation: Demonstrated that the Global-Input strategy yields significantly lower errors (relative $L_2$ error $\approx 2.98 \times 10^{-4}$ ) compared to the Local-Input strategy, which failed to propagate constraints into the domain interior.
Poisson Equation (2D & 3D): Successfully solved problems with sharp peaks and high-frequency components. On a 2D Poisson equation with a sharp peak, DataInNet achieved a maximum absolute error of $1.77 \times 10^{-3}$ , whereas PINNs failed to capture the variation. In 3D high-frequency cases, relative $L_2$ errors reached $1.93 \times 10^{-3}$ .
Helmholtz Equation: Achieved a relative $L_2$ error of $3.23 \times 10^{-6}$ , demonstrating superior performance in high-frequency regimes compared to PINNs.
Wave Equation: Successfully approximated solutions with high-frequency components ( $O(10^{-5})$ relative $L_2$ error), with the sigmoid activation function proving optimal for this specific case.

Across all experiments, increasing the number of modules in the data integration network and the number of training points consistently improved accuracy. The choice of activation function (e.g., $\sin(x)$ , $\tanh(x)$ , $\sigma(x)$ ) was found to be problem-dependent, with no single function universally optimal.

Significance and Claims
The paper claims that DataInNet offers a "novel approach" to leveraging physical data by shifting from a penalty-based paradigm to an architectural integration paradigm. The authors assert that this method substantially narrows the solution search space, thereby mitigating optimization challenges common in traditional PINNs. The results suggest that DataInNet delivers superior performance on difficult problems, including those with high-frequency solutions and complex geometries (e.g., L-shaped domains). The authors conclude that while the method is effective, future work will focus on devising more efficient data integration methods and assessing applicability in even more challenging scenarios. They modestly note that when all physical data (source, initial, boundary) are constant, DataInNet offers no essential distinction from fully connected neural networks.