🤖 machine learning

UFO: A Domain-Unification-Free Operator Framework for Generalized Operator Learning

This paper introduces UFO, a novel neural operator framework that achieves discretization decoupling and robust generalization across diverse benchmarks by enabling adaptive, cross-domain interactions among distinct function representations without requiring domain unification.

Original authors: Hanli Qiao, George Em Karniadakis, Muhammad Muniruzzaman

Published 2026-05-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hanli Qiao, George Em Karniadakis, Muhammad Muniruzzaman

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

The Problem: The "One-Size-Fits-All" Translator

Imagine you are trying to teach a robot to predict how water flows through a pipe. In the past, scientists built two main types of "translator" robots:

The Physical Translator (DeepONet): This robot is great at looking at the pipe's physical shape and location. It's like a carpenter who knows exactly where every nail is. However, it struggles to understand the "vibrations" or high-frequency ripples in the water. If the water starts vibrating wildly, this robot gets confused.
The Frequency Translator (FNO): This robot is a musician. It listens to the "notes" (frequencies) of the water flow. It's amazing at understanding complex vibrations. But, it's rigid. It needs the water data to be measured on a perfect, evenly spaced grid (like a piano keyboard). If the measurements are messy or taken from random spots, this robot breaks down.

Both robots are stuck in their own "world." One only speaks the language of place, and the other only speaks the language of sound. They can't easily talk to each other or adapt if the input data changes (like if you measure the water at a different speed or in a different pattern).

The Solution: UFO (The Universal "Bilingual" Bridge)

The authors introduce a new framework called UFO (Domain-Unification-Free Operator). Think of UFO not as a single robot, but as a master translator who speaks two languages fluently at the same time.

Instead of forcing the robot to choose between "Place" or "Sound," UFO builds a bridge between them. It creates a special "handshake" where the physical location and the frequency sound waves talk to each other dynamically.

Here is how UFO works, using a simple analogy:

1. The Two Sides of the Bridge

The Spectral Encoder (The Musician): UFO takes the messy input (the water flow) and turns it into a musical score. It doesn't just listen to the notes; it learns how the notes change based on where they are happening.
The Spatial Basis Network (The Architect): UFO also builds a flexible map of the space where the solution lives. This map isn't stuck to a grid; it can stretch and shrink to fit any shape or size.

2. The Secret Sauce: The "Adaptive Phase Modulation"

This is the most important part. In old models, the Musician and the Architect would just shout their answers at each other and hope they matched.

In UFO, they use a dynamic phase modulator. Imagine the Musician and the Architect are dancing together.

If the input is smooth, they dance a slow, gentle waltz.
If the input is jagged or chaotic (like a sudden shockwave), they instantly switch to a fast, complex tango.
This "dance step" (the phase) is learned on the fly. It allows the two different ways of looking at the problem to blend perfectly, creating a solution that is both physically accurate and spectrally rich.

Why is this a Big Deal? (The "Discretization Decoupling")

The paper claims UFO has a superpower called Discretization Decoupling.

Imagine you are trying to predict the weather.

Old Robots: If you trained the robot on data from 100 weather stations, it could only predict the weather for those exact 100 spots. If you asked for a prediction for a spot in between, or if you only had 10 stations, the robot would fail or give a bad answer. It was "tied" to the grid.
UFO: Because UFO separates the "input view" from the "output view," it doesn't care how you measured the data.
- You can train it with data from 100 stations.
- You can then ask it to predict the weather for a single point, or for 1,000 points, or for points in a weird, irregular pattern.
- It works because it learned the relationship between the physics, not just the specific grid of numbers.

The Proof: Four Tough Tests

The authors tested UFO against the old robots on four very difficult challenges:

The "Step" Test (StepHeat): They used data with sudden, sharp jumps (like a cliff). Old robots either smoothed out the cliff (missing the detail) or got the frequency wrong. UFO kept the sharp edge perfectly.
The "Shift" Test (Delta-Helmholtz): They moved the whole pattern around and changed the spacing of the data. UFO stayed steady and accurate, while the others got distorted or collapsed.
The "Twist" Test (Burgers): They used complex, swirling fluid patterns. UFO preserved the shape of the swirls even when the patterns were bigger or smaller than what it was trained on.
The "Random Noise" Test (GRF-Helmholtz): They used random, chaotic noise. UFO handled the chaos better than the physical-only robot and stayed competitive with the frequency-only robot, proving it can handle both order and chaos.

The Bottom Line

UFO is a new way of teaching computers to solve physics problems. Instead of forcing the computer to look at the world in just one way (either by location or by frequency), UFO lets it look at the world through two lenses simultaneously, blending them together with a flexible, adaptive dance.

This means it can learn from messy, irregular data and still give you a precise, high-resolution answer, even if you ask for the answer in a format it has never seen before. It's a more flexible, robust, and "physically aware" way to teach machines how the world works.

Technical Summary: UFO (Domain-Unification-Free Operator)

1. Problem Statement

Neural operators have emerged as a powerful paradigm for learning mappings between function spaces, particularly for approximating solution operators of parametric partial differential equations (PDEs). However, existing architectures are fundamentally constrained by their realization within a single representational domain.

Deep Operator Networks (DeepONets) operate primarily in the physical domain, utilizing branch-trunk networks. While flexible for irregular geometries, they often suffer from spectral bias, struggling to capture high-frequency modes associated with highly oscillatory or non-smooth solutions.
Fourier Neural Operators (FNOs) and related spectral methods operate in the spectral domain. While efficient for global interactions and high-frequency modes, they typically rely on regular grids and fixed spectral representations, limiting their applicability to heterogeneous domains and irregular sampling.

Furthermore, most existing approaches exhibit discretization dependence. DeepONet models often rely on fixed input sensor sets, while FNOs require compatible input-output grid representations. This restricts their utility in scenarios where dense measurements are unavailable or expensive, yet high-resolution predictions are required. Current attempts to address these issues often involve architectural augmentations (e.g., latent spaces, learned deformations) but still rely on fixed fusion rules (concatenation, attention) within a predefined realization mechanism, rather than fundamentally rethinking how operators are realized across domains.

2. Methodology: The UFO Framework

The authors introduce UFO (Domain-Unification-Free Operator), a cross-domain neural operator framework. Unlike previous methods, UFO does not unify representations into a single domain. Instead, it realizes operators through adaptive, jointly conditioned interactions among representations defined on distinct domains.

2.1 Theoretical Foundation

UFO is grounded in the concept of Cross-Domain Operator Learning (CDOL). Formally, a neural operator $G_\alpha$ employs CDOL if it is realized through a non-separable, jointly-conditioned interaction:
$G_\alpha(f)(x) = C_\alpha(\Phi_S(x), \Psi_H(f))$
where:

$\Phi_S(x) \in S$ is a spatial representation of the query coordinate $x$ (Solution Domain).
$\Psi_H(f) \in H$ is a global representation of the input function $f$ (Input Domain), where $S \neq H$ .
$C_\alpha$ is a learnable coupling operator whose realization depends jointly on both representations and cannot be reduced to independent mappings.

2.2 Architecture

The UFO architecture consists of three tightly coupled modules:

Spectral Encoder (SE):
- Constructs a global representation of the input function in a domain distinct from physical space.
- It lifts input observations $\{x'_i\}$ into a latent space, applies a spectral transformation (e.g., FFT-based), and modulates the coefficients using a coordinate-conditioned weighting function $\omega_\theta(x'_i)$ .
- The global representation $\Psi_H(f)$ is obtained via mean aggregation and nonlinear mapping of real and imaginary components. This design allows for discretization-agnostic and resolution-invariant input representations.
Spatial Basis Network (SBN):
- Constructs a continuous representation of the solution domain $\Phi_S(x)$ for arbitrary query coordinates $x$ .
- Implemented as a multilayer perceptron (MLP) that maps coordinates directly to a high-dimensional feature space.
- Unlike grid-based methods, the SBN does not rely on convolution or interpolation, enabling arbitrary query resolutions and irregular locations.
Adaptive Phase-Modulated Coupling Operator:
- This is the core mechanism for cross-domain realization. It models structured interactions between the spectral input representation $\Psi_H(f) = a + ib$ and the spatial basis $\Phi_S(x)$ .
- A learnable mapping $\gamma_\theta$ generates a coupling phase $\alpha$ based on the joint feature $\eta(\Phi_S(x), \Psi_H(f))$ .
- The final output is computed via a phase-modulated interaction:
  $G_\alpha(f)(x) = \langle \Phi_S(x), \cos \alpha \odot a + \sin \alpha \odot b \rangle$
- This mechanism ensures the operator realization is non-separable and adapts dynamically to the interaction between input and solution spaces.

3. Key Contributions

Novel Realization Mechanism: The introduction of UFO, which shifts operator learning from single-domain approximation to cross-domain realization via adaptive, phase-modulated coupling.
Discretization Decoupling: The architecture enables input functions to be observed at resolutions or locations different from those used during training, while solutions can be queried at arbitrary output resolutions.
Comprehensive Benchmarking: The design of four complementary benchmarks to evaluate not just pointwise accuracy, but also spectral consistency, structural coherence, and generalization under distribution shifts:
- StepHeat: Discontinuous inputs to test spectral bias.
- $\delta$ -Helmholtz: Irregular sampling and spectrum mismatch to test translation consistency.
- 2D Burgers: Nonlinear dynamics to test structure preservation under bidirectional extrapolation.
- GRF-Helmholtz: Stochastic high-frequency random fields to test frequency-dominated operators.

4. Experimental Results

The paper evaluates UFO against DeepONet and FNO across the four benchmarks:

StepHeat (Spectral Bias): UFO achieves the lowest relative $L_2$ error in 4 out of 6 discontinuity cases and the best Barron norm error (spectral consistency) in 3 cases. It successfully captures and propagates high-frequency modes induced by discontinuities, outperforming DeepONet (which suffers spectral bias) and FNO (which shows instability with discontinuous inputs).
$\delta$ -Helmholtz (Translation Consistency): Under strong global shift extrapolation ( $\delta = \pm 30.8$ ) and irregular sampling, UFO maintains translation consistency and stable generalization. DeepONet exhibits severe structural distortion, and FNO collapses under large shifts.
2D Burgers (Nonlinear Structure): UFO preserves nonlinear level-set structures and contour geometry better than baselines during bidirectional out-of-distribution (OOD) extrapolation. While FNO performs well in pointwise accuracy within the training distribution, it degrades sharply in OOD regimes with spurious structures.
GRF-Helmholtz (Stochastic High-Frequency): In this frequency-dominated regime, FNO achieves the best errors (reflecting its inductive bias). However, UFO remains competitive and substantially outperforms DeepONet, particularly in extreme wave number cases ( $k=120$ ), demonstrating a more balanced realization mechanism across heterogeneous regimes.
Discretization Decoupling: Experiments show that UFO's error remains stable as input resolution decreases (from $100\times100$ to $55\times55$ ) and as output query resolution increases (up to $550\times550$ ). In contrast, FNO deteriorates rapidly with increased output resolution, confirming its grid dependence.
Ablation Study: Removing the adaptive phase modulation $\alpha$ (replacing it with a fixed separable readout) leads to substantial degradation in performance, confirming that the jointly conditioned interaction is essential for the framework's success.

5. Significance and Claims

The paper claims that UFO establishes cross-domain, phase-modulated realization as a powerful framework for discretization-decoupled neural operator learning.

Beyond Single-Domain Constraints: UFO reframes operator learning by separating input-domain encoding, solution-domain representation, and adaptive coupling. This allows the model to operate effectively across heterogeneous regimes (discontinuous, irregular, nonlinear, stochastic) without being specialized to a single domain.
Practical Flexibility: The ability to handle varying input resolutions and query arbitrary output resolutions addresses a critical practical limitation where dense measurements are expensive but high-resolution predictions are needed.
Structural Coherence: The framework ensures that predictions remain physically coherent and structurally consistent under distribution shifts, a property often lacking in single-domain approaches.

The authors note that this flexibility is not "free"; if the target operator relies heavily on localized input features, overly sparse observations will still degrade prediction quality. However, UFO is discretization-flexible, meaning it does not require a fixed grid, provided the input observations resolve the structures determining the operator response.

The work positions UFO not merely as a new architecture, but as a conceptual shift toward thinking about operator realization beyond domain constraints, opening avenues for multi-domain, time-dependent, and physics-informed extensions in complex scientific systems.