⚛️ quantum physics

On Embedding Design in Quantum Physics-Informed Neural Networks

This paper introduces a unified embedding framework for Quantum Physics-Informed Neural Networks (QPINNs) and proposes two novel methods, LQNN-TE-QPINN and AdaFreq-QPINN, which demonstrate that strategic embedding design significantly enhances approximation accuracy and optimization efficiency in solving partial differential equations compared to existing direct and fixed encoding approaches.

Original authors: Ban Q. Tran, Nahid Binandeh Dehaghani, Susan Mengel, Rafal Wisniewski, A. Pedro Aguiar

Published 2026-09-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ban Q. Tran, Nahid Binandeh Dehaghani, Susan Mengel, Rafal Wisniewski, A. Pedro Aguiar

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

Nature speaks in the language of change. From the swirling of a storm front to the flow of blood through a vein, the physical world is governed by equations that describe how quantities shift over space and time. For decades, scientists have relied on powerful computers to solve these equations, breaking complex problems into tiny, manageable pieces. However, as problems grow more intricate or involve too many variables, these traditional methods can become slow and expensive. In recent years, a new approach has emerged that treats these equations not as puzzles to be solved step-by-step, but as patterns for a machine to learn. This method, known as a physics-informed neural network, teaches a computer model to respect the laws of physics while it searches for a solution, allowing it to find answers without needing vast amounts of pre-existing data.

Now, researchers are taking this idea a step further by asking if the unique properties of quantum mechanics can make these learning machines even better. Quantum computers, which operate on the strange principles of subatomic particles, promise to handle certain types of calculations in ways classical computers cannot. The challenge is figuring out how to feed real-world data, like coordinates of space and time, into these fragile quantum systems. The way this data is translated into a quantum language is called an "embedding." It is a crucial first step: if the data is translated poorly, the quantum computer cannot learn the pattern, no matter how powerful it is. Until now, scientists have not fully understood how the design of this translation step affects the final result.

In a recent study, a team of researchers set out to map out the landscape of these embeddings and test which designs work best. They focused on a classic problem in fluid dynamics known as the Burgers equation, which describes how waves move and steepen, often forming sharp fronts similar to shockwaves. The team built a hybrid system where a classical computer handles the heavy lifting of optimization, while a quantum circuit acts as the core function approximator. They tested several different ways to translate the input data into the quantum system. Some methods used fixed, pre-set rules, much like a dictionary with a static list of definitions. Others allowed the translation rules to change and adapt as the computer learned.

The researchers introduced two new strategies to see if they could outperform the standard approaches. The first was a system that used a small, auxiliary quantum circuit to generate custom features for the data before it entered the main calculation. Think of this as a translator who doesn't just swap words but adds context and nuance based on the specific sentence being spoken. The second strategy took a fixed mathematical formula and made its scaling factors trainable, allowing the system to fine-tune the frequency of the data encoding as it learned. They tested these new methods against older, fixed designs and against purely classical learning models.

The results showed that the design of the embedding matters deeply. In one-dimensional tests, the new system that used the auxiliary quantum circuit to generate features achieved a relative error of 0.0916. This was significantly better than the standard approach, which had an error of 0.4501, and it also outperformed a sophisticated classical model that required roughly eleven times more adjustable parameters to reach a higher error rate of 0.1233. The system that simply adjusted the frequency of a fixed formula also improved upon the static versions, reaching the lowest error among the analytical methods with only four extra adjustable parameters. These findings suggest that giving the quantum system a way to adapt how it sees the data can lead to much more accurate solutions, even with a smaller number of trainable settings.

When the researchers moved to a more complex two-dimensional version of the problem, the picture became slightly more nuanced. The new system with the auxiliary circuit still achieved the lowest training objective, meaning it minimized the physics-based errors most effectively during the learning process. However, its final solution error was not clearly separated from another method that simply re-encodes the data repeatedly within the circuit. This indicates that while a more flexible embedding helps the system learn better, it does not automatically guarantee a perfect final answer in every scenario. The study also examined how these systems hold up under simulated noise, mimicking the imperfections found in real quantum hardware. The new system maintained the lowest absolute error in its predictions, though it was slightly more sensitive to noise than the simpler models, likely because the noise affected both its translation and calculation circuits.

Ultimately, this work demonstrates that the way data is prepared for a quantum computer is just as important as the computer itself. By treating the embedding as a flexible, learnable component rather than a fixed rule, researchers can significantly improve the accuracy of quantum models solving physical problems. The study confirms that there is no single "best" design for every situation; the optimal choice depends on the specific problem and the resources available. As quantum technology matures, understanding these subtle design choices will be essential for unlocking the full potential of quantum machines in solving the complex equations that govern our world.

Technical Summary: Embedding Design in Quantum Physics-Informed Neural Networks

Problem Statement
Quantum Physics-Informed Neural Networks (QPINNs) solve partial differential equations (PDEs) by training parameterized quantum circuits (PQCs) to minimize physics-based residuals. A critical yet under-explored component of QPINNs is the embedding (or encoding) mechanism, which maps continuous input coordinates (e.g., space and time) into quantum states before variational processing. The choice of embedding dictates the feature representation available to the variational circuit, thereby influencing the model's approximation properties, trainability, and convergence behavior. While classical Physics-Informed Neural Networks (PINNs) have explored adaptive activation functions to improve approximation, the relationship between embedding design and the representation of PDE solutions in the quantum setting remains insufficiently understood, particularly regarding how different feature representations handle nonlinearities and steep gradients.

Methodology
The authors introduce a unified embedding framework that formulates the embedding as a functional transformation shaping the feature representation. Within this framework, they propose and evaluate two novel embedding strategies against a taxonomy of existing methods:

Linear Quantum Neural Network Trainable Embedding (LQNN-TE-QPINN): This approach utilizes an auxiliary trainable quantum circuit to generate data-dependent features. Unlike previous multiplicative combinations, LQNN-TE adds these features linearly to the rescaled input coordinates before encoding them into the main variational circuit. The encoding angle for wire $i$ is defined as $\gamma_i = \tilde{\xi}_{j(i)} + b_i(\tilde{\xi}; \theta_{emb})$ , where $b_i$ is the expectation value from the auxiliary circuit.
Adaptive-Frequency QPINN (AdaFreq-QPINN): This method retains the analytical arccos-Chebyshev encoding structure but promotes the fixed frequency-scaling factors ( $\lambda_k$ ) to trainable parameters. This allows the encoding frequencies to adapt jointly with the variational circuit during optimization, analogous to adaptive activation functions in classical PINNs.

The study evaluates these methods on one- and two-dimensional Burgers equations, a benchmark nonlinear PDE characterized by convective and diffusive dynamics. The evaluation compares the proposed embeddings against:

Direct coordinate encoding (baseline).
Fixed analytical frequency embeddings: TowerCheb, Binary, Golomb, Turnpike, and Hamming (using predefined integer sequences for scaling factors).
Alternative hybrid architectures: Classical feed-forward trainable embedding (FNN-TE), Quantum-assisted PINN (QA-PINN), and Data Re-uploading QPINN (Reupload-QPINN).
Classical baselines: Standard PINN and Locally Adaptive Activation Function PINN (LAAF-PINN).

Experiments were conducted on classical hardware simulating quantum circuits (using PennyLane and PyTorch) under various conditions, including circuit sizing (width and depth), training budgets, simulated hardware noise channels (depolarizing, amplitude/phase damping, readout bit-flip), and finite measurement shots.

Key Contributions

Unified Framework: A taxonomy of embedding strategies (direct, fixed analytical, adaptive-frequency, and trainable quantum) that enables systematic comparison under a common formulation.
Novel Architectures: The introduction of LQNN-TE-QPINN (linear combination of auxiliary quantum features) and AdaFreq-QPINN (trainable frequency scales in analytical encoding).
Comprehensive Empirical Study: A rigorous evaluation across 1D and 2D Burgers equations, analyzing the impact of embedding design on approximation accuracy, optimization dynamics, and robustness to noise.
Analysis of Trade-offs: Investigation into how embedding adaptivity interacts with circuit size, training budget, and noise, revealing that no single embedding is uniformly optimal.

Results

One-Dimensional Performance:
- LQNN-TE-QPINN achieved the lowest relative $L_2$ error of 0.0916 with 720 trainable parameters. This significantly outperformed the direct-encoding QPINN (0.4501) and the classical LAAF-PINN (0.1233), despite the classical model having approximately 11 times more parameters.
- AdaFreq-QPINN achieved the lowest error among analytical embeddings (0.3262) with only four additional trainable parameters compared to fixed embeddings, demonstrating that adapting frequency scales improves performance over fixed choices.
- The trainable quantum embedding (LQNN-TE) continued to improve over a larger fraction of the training budget compared to fixed embeddings, which often plateaued early.
- Noise and Stability: Under simulated noise, LQNN-TE-QPINN maintained the lowest absolute derivative error across all tested channels. However, it exhibited the highest relative degradation compared to its noiseless baseline, likely because noise affects both its auxiliary and main circuits.
Two-Dimensional Performance:
- LQNN-TE-QPINN attained the lowest physics-informed training objective (0.298 at 200 epochs) but did not clearly separate from the Reupload-QPINN in terms of solution error (40.58 vs. 39.11) within the tested budget.
- The study highlighted a divergence between the training objective (loss) and the solution error, noting that minimizing the loss does not always guarantee the lowest solution error, particularly when comparing different embedding families.
General Observations:
- Circuit Width vs. Depth: Increasing the width of the main variational circuit had a stronger impact on reducing solution error than increasing depth.
- Embedding Adaptivity: While greater adaptivity (LQNN-TE) can yield superior accuracy, it introduces more trainable parameters and a more complex optimization landscape. Adaptive-frequency embeddings (AdaFreq) offer a lightweight alternative that improves upon fixed analytical embeddings with minimal parameter overhead.

Significance and Claims
The paper claims that embedding design substantially influences the approximation and optimization behavior of QPINNs. The results demonstrate that:

Trainable quantum embeddings can achieve high accuracy with fewer parameters than classical counterparts, provided the embedding depth is sufficient to supply a rich feature map.
Adapting encoding frequencies (AdaFreq) is a low-cost mechanism to improve performance over fixed analytical embeddings.
There is no universally optimal embedding; the relative benefits of fixed, adaptive, and trainable embeddings depend on the target problem, circuit architecture, and training configuration.
A critical distinction exists between the training objective and the solution error; optimizing one does not guarantee the other, necessitating separate reporting of both metrics.

The authors modestly note limitations, including the fact that the two-dimensional comparison did not definitively separate LQNN-TE-QPINN from Reupload-QPINN due to budget constraints and simulation non-determinism. Furthermore, while LQNN-TE-QPINN achieved the lowest absolute error under noise, its higher relative degradation suggests it is not intrinsically more noise-robust, a weakness attributed to the noise affecting both the embedding and solver circuits. Future work is directed toward longer training budgets, noise mitigation strategies, and generalization to unseen points.

More like this