🤖 machine learning

Learning PDE Solvers with Physics and Data: A Unifying View of Physics-Informed Neural Networks and Neural Operators

This paper proposes a unifying framework that categorizes Physics-Informed Neural Networks (PINNs) and Neural Operators (NOs) along three fundamental dimensions—what is learned, how physics is integrated, and how computational load is amortized—to clarify their relationships, limitations, and roles in developing reliable learning-based PDE solvers.

Original authors: Yilong Dai, Shengyu Chen, Ziyi Wang, Xiaowei Jia, Yiqun Xie, Vipin Kumar, Runlong Yu

Published 2026-02-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yilong Dai, Shengyu Chen, Ziyi Wang, Xiaowei Jia, Yiqun Xie, Vipin Kumar, Runlong Yu

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 water flows through a river, how heat spreads through a metal plate, or how a virus spreads through a city. Scientists use complex mathematical recipes called Partial Differential Equations (PDEs) to describe these things.

For decades, solving these recipes has been like trying to solve a massive jigsaw puzzle by hand, piece by piece, every single time the weather changes or the river bends. It's accurate, but incredibly slow and expensive.

Recently, scientists started using Artificial Intelligence (AI) to speed this up. But there are two main schools of thought on how to teach the AI, and they often argue about which is better. This paper is like a wise mediator stepping in to say: "Stop fighting. You are actually two different tools in the same toolbox, designed for different jobs."

Here is the breakdown of the paper using simple analogies:

1. The Two Main Characters: The "Tutor" vs. The "Translator"

The paper focuses on two popular AI methods: PINNs and Neural Operators (NOs).

PINNs (Physics-Informed Neural Networks): The "Strict Tutor"

How it works: Imagine you have a student (the AI) taking a test for one specific river flow. The teacher (the physics laws) stands over their shoulder, checking every single answer. If the student's answer violates the laws of physics (e.g., water flowing uphill), the teacher gives them a "red mark" (a penalty). The student keeps studying until they get the perfect answer for that one specific river.
Best for: When you have very little data, or when you need to figure out hidden secrets (like finding a leak in a pipe based on a few noisy sensors).
The Downside: It's slow. You have to make the student study from scratch for every single new river. If you want to simulate 1,000 different rivers, you have to teach the student 1,000 times.

Neural Operators (NOs): The "Super Translator"

How it works: Imagine a translator who has read thousands of books about rivers. They don't solve a specific river; they learn the pattern of how rivers behave. Once trained, if you show them a new river map they've never seen before, they can instantly "translate" it into a flow prediction without studying it first.
Best for: When you have lots of data and need to run thousands of simulations quickly (like designing a new airplane wing or predicting weather for the next year).
The Downside: It needs a huge library of examples to learn from. If the new river looks totally different from anything in its library (a "contract shift"), it might guess wrong without realizing it.

2. The "Unified View": It's Not a Competition, It's a Continuum

The authors argue that we shouldn't treat these as rivals. Instead, think of them as points on a sliding scale:

The "Data-Sparse" End: You have almost no data, but you know the physics perfectly. You need the Tutor (PINN) to guide the AI step-by-step.
The "Data-Rich" End: You have massive amounts of data and need speed. You need the Translator (NO) to memorize the patterns and work instantly.
The "Hybrid" Middle: Sometimes you need a translator who also has a tutor standing by to double-check the work. This is the future: an AI that guesses quickly (NO) but checks its own work against the laws of physics (PINN) to ensure it's safe.

3. The Three Big Hurdles (Why is this hard?)

The paper identifies three "monsters" that both methods have to fight, regardless of which one they are:

The "Zoom" Problem (Multiscale): Real-world physics has huge waves and tiny ripples. AI often learns the big waves easily but misses the tiny ripples, or gets confused by the noise. It's like trying to hear a whisper in a hurricane.
The "Shape-Shifter" Problem (Geometry): If you train an AI on a square room and then ask it to predict airflow in a round cave, it might get confused. The AI needs to understand that the shape of the world matters, not just the numbers.
The "Trust" Problem (Contract Shift): This is the most important one. If you train an AI on summer weather, will it work in winter? If it guesses wrong, will it tell you? The paper says we need "trust signals"—like a dashboard warning light—that tell us when the AI is guessing outside its comfort zone.

4. How to Choose Your Tool?

The paper gives a simple guide on which tool to pick based on your "Job Description":

Pick the Tutor (PINN) if: You are a detective solving a mystery with very few clues, or you need to find hidden parameters (like "where is the crack in the dam?").
Pick the Translator (NO) if: You are an engineer designing 10,000 different car parts and need to know which one is fastest right now.
Pick the Hybrid if: You are running a nuclear power plant or a weather forecast. You need the speed of the Translator, but you need the Safety Check of the Tutor to ensure you don't make a catastrophic mistake.

The Big Takeaway

The paper concludes that the future of scientific computing isn't about picking a winner between these two methods. It's about building a unified system where:

We clearly define what the AI is allowed to do (the "contract").
We mix the speed of data-driven learning with the safety of physics laws.
We build "warning systems" so we know when the AI is about to fail.

Think of it like building a self-driving car. You don't just want a car that learns from driving data (NO); you also need the hard-coded rules of traffic laws (PINN) to ensure it doesn't drive off a cliff. This paper provides the blueprint for building that perfect, safe, and fast hybrid system.

1. Problem Statement

Partial Differential Equations (PDEs) are fundamental to scientific modeling in fields like fluid dynamics, climate science, and materials science. While classical numerical solvers are effective for single, high-fidelity simulations, modern scientific workflows increasingly require multi-query scenarios (e.g., design optimization, uncertainty quantification, inverse problems, and data assimilation).

In these scenarios, repeatedly solving PDEs with varying parameters, boundary conditions, or geometries using traditional discretization methods becomes a computational bottleneck. This has spurred the development of learning-based PDE solvers. However, the field is currently fragmented into two dominant paradigms:

Physics-Informed Neural Networks (PINNs): Learn instance-specific solutions by minimizing PDE residuals.
Neural Operators (NOs): Learn mappings from input specifications (e.g., coefficients, boundary conditions) to solution fields, enabling amortized inference.

The paper argues that the community often views these as competing methods, lacking a unified perspective to understand their relationships, structural limitations, and appropriate roles within scientific workflows.

2. Methodology: A Unifying Design Space

The authors propose a unifying taxonomy that places PINNs and NOs within a shared design space defined by three fundamental structural choices:

T1: Object of Learning
- PINNs: Learn an instance-specific solution $u_\phi(x, t)$ for a single set of parameters.
- NOs: Learn a family-level operator $G(a)$ that maps instance specifications $a$ to solution fields $u$ , covering a family of PDEs.
T2: Supervision Interface
- PINNs: Treat physics constraints (governing equations) as the primary supervision signal, often augmented by sparse data.
- NOs: Treat paired data (input-output pairs) as primary supervision, with physics injected via architecture, regularization, or refinement steps.
T3: Amortization Locus
- PINNs: Amortize supervision (learning the physics) but require per-instance training (optimization for every new query).
- NOs: Require offline training and data generation but provide amortized inference (fast evaluation for new instances without retraining).

The paper further defines Workflow Contracts, which specify the perturbations allowed at deployment (e.g., geometry changes, resolution shifts, parameter regimes) and the invariants that must be preserved. The choice between PINNs and NOs depends on which contract the solver must satisfy.

3. Key Contributions

The paper makes three primary contributions:

A Unifying Taxonomy: It organizes the field based on structural design choices (what is learned, how physics is integrated, and where costs are amortized) rather than just algorithmic variations.
Cross-Paradigm Analysis of Structural Properties: It identifies three universal challenges ("pressure points") that affect all learning-based PDE solvers, regardless of the paradigm:
- C1: Multiscale Representation: Solutions often contain sharp boundary layers and broadband spectra. Learning signals (residuals or paired data) often allocate error unevenly across frequencies, leading to aliasing or smearing of high-frequency content.
- C2: Geometry and Boundary Feasibility: Solvers must handle variations in domain shapes and boundary conditions. Mismatches here often lead to "silent" failures where average error is low, but physical feasibility (e.g., boundary conditions) is violated under deployment perturbations.
- C3: Trustworthiness under Contract Shift: Reliability requires more than average accuracy; it demands uncertainty quantification, out-of-distribution detection, and mechanisms to trigger fallback or refinement when the workflow contract is violated.
Application-Oriented Method Choice: The paper provides a decision framework for selecting PINNs, NOs, or Hybrids based on specific workflow constraints (e.g., data availability, query frequency, and geometry complexity).

4. Detailed Review of Paradigms

Physics-Informed Neural Networks (PINNs)

Mechanism: Minimizes a composite loss function comprising PDE residuals, boundary/initial condition violations, and data mismatch.
Key Challenges: Poor conditioning for stiff PDEs, gradient conflicts between loss terms, and the high cost of automatic differentiation for high-order derivatives.
Advancements:
- Residuals: Moving from strong-form to weak/variational forms for better conditioning.
- Constraints: Using hard/semi-hard constraints (e.g., parameterizations that satisfy boundary conditions by construction) rather than soft penalties.
- Sampling: Adaptive collocation strategies to focus on regions of high complexity.
- Geometry: Graph-based PINNs and manifold-aware coordinate systems to handle complex domains.

Neural Operators (NOs)

Mechanism: Learns a map $u = G(a)$ using paired data (e.g., DeepONet, Fourier Neural Operators).
Key Challenges: Distribution shift in input parameters, discretization mismatch (training on one grid, testing on another), and long-horizon error accumulation.
Advancements:
- Geometry/Discretization: Mesh/graph-based operators for irregular domains; coordinate-free kernels for domain variation.
- Multiscale: Spectral truncation variants and wavelet-based operators to capture fine-scale features.
- Physics Injection: "Hybrid" approaches where the operator provides a fast proposal, and physics constraints (residuals) are used for correction, refinement, or validation (e.g., PINO).
- Foundation Models: Pre-training operators across multiple physics domains to enable transfer learning.

5. Results and Synthesis

The paper does not present new experimental benchmarks but synthesizes existing literature (2017–2026) to draw conclusions:

Hybridization is the Future: Pure PINNs are data-efficient but slow for multi-query tasks. Pure NOs are fast but require massive data and struggle with strict physical constraints. Hybrid approaches (e.g., Operator proposes + Physics refines) offer the best balance, combining amortized speed with physical validity.
Workflow Dictates Method:
- Use NOs when paired simulation data is available and the goal is rapid inference for many queries (e.g., design loops, digital twins).
- Use PINNs when data is sparse/noisy, parameters are unknown (inverse problems), or strict feasibility under constraints is paramount.
- Use Hybrids for long-horizon rollouts or when operating near the boundaries of the training distribution.

6. Significance and Future Directions

This survey is significant because it moves the field beyond "which method is better" to "which method fits the workflow contract." It emphasizes that reliability in scientific computing requires:

Controlled Extrapolation: Explicitly defining generalization contracts and using diagnostics to detect when a solver is operating out-of-contract.
Physics-First Design: Treating physical laws (conservation, symmetry) as explicit interfaces enforced by architecture, not just soft penalties.
Cost Transparency: Developing end-to-end cost models that account for data generation, training, and adaptation, rather than just inference latency.

The authors conclude that the path forward lies in shared abstractions and evaluation signals that principledly connect physics and data, enabling the development of reliable, reusable, and decision-driven PDE solvers for scientific discovery.