Hybrid quantum recurrent neural network for remaining useful life prediction

arXiv:2504.20823 · cs.LG, quant-ph · Submitted 2026-08-17 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Hybrid quantum recurrent neural network for remaining useful life prediction of turbofan engines".

Jane: The paper was written by Olga Tsurkan, Aleksandra Konstantinova, Arsenii Senokosov, Asel Sagingalieva and Alexey Melnikov from Terra Quantum AG.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Jane: We also have Lu with us today — senior AI researcher at Tsinghua.

Tom: We also have Meng with us today — lead engineer at a mysterious AI startup.

Jane: We also have Lalam with us today — the in-house Large Language Model.

Tom: Alright, let's get started.

Title and Authors: Tom: Welcome back to the show, everyone. Today we're looking at a paper that's been making the rounds on arXiv — it's called "Hybrid quantum recurrent neural network for remaining useful life prediction." Jane, I have to say, that title is a mouthful, but it's got me genuinely excited.

Jane: It really is a mouthful, Tom, but let's break it down. "Remaining useful life" is exactly what it sounds like — for a jet engine, how many more flight cycles it can safely run before it fails. And the paper is about predicting that using a neural network that has a quantum component inside it.

Tom: Right, and the team behind it is from Terra Quantum AG in Switzerland. You've got Olga Tsurkan, Aleksandra Konstantinova, Aleksandr Sedykh, Arsenii Senokosov, Daniil Tarpanov, Matvei Anoshin, Asel Sagingalieva, and Alexey Melnikov as the corresponding author. That's a solid group of researchers.

Jane: And what they're doing is pretty clever. They're taking a classic recurrent neural network — the kind that's good at handling sequences of data like time series — and they're swapping out some of its internal math with quantum circuits. The idea is that the quantum part can pick up patterns that the classical part might miss.

Tom: Exactly. And the application is really concrete. We're talking about NASA's C-MAPSS dataset, which simulates turbofan engine degradation. The model takes sensor readings from an engine over a window of time and predicts how many cycles it has left. That's the kind of thing that could save airlines millions in maintenance costs.

Jane: And it's not just about money, Tom. It's about safety. If you can predict a failure before it happens, you can ground the engine, fix it, and avoid a catastrophic event. The paper is really about making that prediction more accurate.

Tom: And here's the kicker — their hybrid quantum model actually beat the purely classical version on the test data. We're talking about a roughly five percent improvement in error metrics. That's meaningful.

Jane: It is. And I love that they're not just throwing quantum at a problem for the sake of it. They're showing a real use case where it helps. That's the kind of work that moves the field forward.

Tom: Absolutely. And we're going to dig into how they actually built this thing and what the results really mean. Stay with us.

Summary of the Paper: Jane: So, Tom, we've talked about the title and the team. Now let's get into what the paper actually does. The full name again is "Hybrid quantum recurrent neural network for remaining useful life prediction." And the core idea is that they've built a hybrid quantum-classical model for predicting jet engine lifespan.

Tom: Right. And the architecture is what I find really interesting. They've taken a Long Short-Term Memory network — an LSTM — and replaced the linear transformations inside each of its gates with what they call Quantum Depth-Infused circuits, or QDI circuits. So each gate of the LSTM is now a small quantum circuit.

Jane: And for our listeners who aren't deep in the weeds, an LSTM has these gates that decide what information to remember and what to forget over time. By making those gates quantum, the network can theoretically explore a much richer set of patterns in the data.

Tom: Exactly. And the results are pretty compelling. On the NASA C-MAPSS FD001 dataset, their hybrid model achieved a root mean squared error of fifteen point four six. That's better than a random forest, better than a plain convolutional network, better than a multilayer perceptron. They beat those by roughly thirteen, sixteen, and eight percent respectively.

Jane: And importantly, they compared it against a classical LSTM with the same number of parameters. The quantum version won on the test set. That's the headline result — the quantum-enhanced model generalizes better even though it has fewer trainable parameters in some cases.

Tom: Yeah, that's the part that really got me. The classical LSTM had more parameters in one configuration, and the quantum model still did better. It suggests the quantum circuits are capturing something the classical ones aren't.

Jane: And the paper offers a possible explanation for that. Quantum circuits with angle-based encoding can be understood as truncated Fourier series. They can access higher-frequency components of the data. For time series like engine degradation, those high-frequency fluctuations might be exactly what you need to make a good prediction.

Tom: So it's not just that quantum is cool — it's that quantum gives the model a different kind of expressive power. And that shows up in the numbers.

Jane: Right. And we should mention that they used a piecewise linear degradation model for the RUL labels, which is a standard and sensible choice. They set an early RUL threshold at one hundred twenty-five cycles.

Tom: Good context. Now, let's talk about what this means for the field and where the improvements could go next.

Improvements Suggested by the Paper: Tom: So, Jane, we've covered the basics and the results. Now let's talk about what the paper suggests as next steps and improvements. And I want to bring in Lu and Meng for this one, because they'll have strong opinions.

Jane: Absolutely. Lu, you're the researcher — what do you see as the most promising direction coming out of "Hybrid quantum recurrent neural network for remaining useful life prediction"?

Lu: Thanks, Jane. I think the most exciting thing here is that the quantum layer is modular. You could drop it into other architectures, not just LSTM. The paper explicitly mentions integrating quantum modules into random forests or gradient boosting as feature encoders. And you could even think about transformer-based models like Chronos or TabPFN getting quantum-enhanced layers.

Tom: That's a big deal. So the QDI circuit isn't just for this one task — it's a building block.

Lu: Exactly. And the paper's own analysis supports that. They did a ZX calculus reduction and found the circuit is already parameter-efficient — no redundant weights. Then they looked at the Fisher Information Matrix and found the gradients are well-distributed across all parameters, which means it's trainable and not prone to barren plateaus. That's a strong signal for generalizability.

Meng: I'd push back a little there, Lu. The trainability analysis is nice, but the real question for me is whether this scales. The circuit has four qubits and eight parameters. That's tiny. If you want to handle more complex datasets, you'll need more qubits, and then you run into noise and decoherence on real hardware.

Jane: That's a fair point, Meng. The paper is working with a very small circuit. But that's also kind of the point — they're showing you can get a real improvement with a minimal quantum resource.

Meng: Sure, and I get that. But I also want to see how this performs on the other C-MAPSS subsets, like FD002 or FD004, which have multiple operating conditions and fault modes. That's where the model would really be tested.

Tom: Good point. The paper only uses FD001, which is the simplest subset. So the next step is clearly stress-testing this on harder data.

Lu: And I'd add that the paper suggests combining the quantum LSTM with attention mechanisms or joint architectures. Some of those advanced models in the comparison tables — like the Transformer plus TCNN — still beat the hybrid quantum model. So the path forward is not quantum alone, but quantum as a component in a larger pipeline.

Meng: Yeah, I can get behind that. If the quantum layer can be dropped into an existing state-of-the-art pipeline and squeeze out another few percent, that's practical value.

Jane: And that's really the takeaway — this paper is a proof of concept that quantum can help in a real industrial prediction task, and the improvements are about integration and scaling.

Conclusion: Tom: Alright, we're wrapping up our discussion of "Hybrid quantum recurrent neural network for remaining useful life prediction." Jane, what's the big picture for our listeners?

Jane: The big picture is that this paper shows a hybrid quantum-classical neural network can beat a purely classical one on a real-world predictive maintenance task. The quantum model achieved an RMSE of fifteen point four six on the NASA C-MAPSS dataset, outperforming several classical baselines and even a classical LSTM with more parameters.

Tom: And it's not just about the numbers. The paper does a thorough analysis of why the quantum circuit works — it's parameter-efficient, it's trainable, and it can access higher-frequency components in the data that classical models tend to miss.

Jane: Right. And while the model doesn't beat every advanced joint architecture out there, it proves that quantum layers can be a valuable addition to a larger machine learning pipeline. That's a meaningful step forward.

Tom: For the world, this could mean more reliable predictions of engine failures, safer flights, and lower maintenance costs. And for the field, it opens the door to more research on integrating quantum circuits into established forecasting models.

Jane: We want to thank Lu and Meng for joining us and adding their perspectives. And to our listeners, we're saying goodbye to this paper and getting ready to dive into the next one on our list.

Tom: Stay tuned, everyone. The conversation continues right after this short break.

Olga Tsurkan, Aleksandra Konstantinova, Arsenii Senokosov, Asel Sagingalieva, Alexey Melnikov

Terra Quantum AG

cs.LG, quant-ph

Submitted: 2026-08-17

Updated: 2026-08-18

Comments: 11 pages, 5 figures. 3 tables

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 54/100

Key concepts

Remaining Useful Life (RUL): This refers to predicting how many more flight cycles a jet engine can safely operate before it is expected to fail. Predicting RUL helps airlines schedule maintenance and prevent catastrophic events.
Hybrid Quantum Recurrent Neural Network: This model combines a Long Short-Term Memory (LSTM) network with quantum circuits. The quantum circuits replace specific linear transformations within the LSTM's gates to allow the model to capture complex patterns in time series data.
Quantum Depth-Infused (QDI) Circuits: These are small quantum circuits used within the gates of an LSTM. They enable the network to explore a richer set of data patterns than classical components, potentially accessing higher-frequency components in degradation signals.

Terminology

Remaining Useful Life — 23× in this paper
Hybrid Quantum Recurrent Neural Network — 11× in this paper
Long Short-Term Memory — 11× in this paper
Fisher Information — 9× in this paper · explained in 3 episodes
Quantum Circuits — 9× in this paper · explained in 2 episodes
Transformer — 8× in this paper · explained in 2 episodes
Predictive Maintenance — 7× in this paper · explained in 2 episodes
Quantum Depth-Infused — 6× in this paper
Fisher Information Matrix — 5× in this paper
Generalization — 4× in this paper · explained in 13 episodes
Core — 3× in this paper · explained in 2 episodes
Entanglement — 3× in this paper · explained in 4 episodes
Hilbert Space — 3× in this paper · explained in 2 episodes
Redundancy — 3× in this paper · explained in 2 episodes
Adam — 2× in this paper · explained in 2 episodes
Attention Mechanism — 2× in this paper · explained in 3 episodes
Benchmark — 2× in this paper · explained in 2 episodes
Interpretability — 2× in this paper · explained in 15 episodes
Time Series Forecasting — 2× in this paper · explained in 3 episodes
ZX-calculus — 2× in this paper · explained in 2 episodes
Federated Learning — 1× in this paper · explained in 15 episodes
Generalizability — 1× in this paper · explained in 6 episodes
Hybrid Approach — 1× in this paper · explained in 4 episodes
Hybrid Model — 1× in this paper · explained in 2 episodes
Parameter Efficiency — 1× in this paper · explained in 3 episodes
Robustness — 1× in this paper · explained in 5 episodes
Score Function — 1× in this paper · explained in 4 episodes
Temporal Convolutional Neural Networks — 1× in this paper

Summary

Summary

This paper introduces a Hybrid Quantum Recurrent Neural Network (HQRNN) framework for predicting the Remaining Useful Life (RUL) of jet engines, evaluated on the NASA Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) dataset, specifically the FD001 subset. The core innovation is the replacement of linear transformations in each gate of a classical Long Short-Term Memory (LSTM) network with Quantum Depth-Infused (QDI) circuits, creating a quantum-enhanced LSTM (QLSTM).

The dataset comprises run-to-failure data for 100 engines, with 14 sensor channels retained after removing constant features. The authors use a piecewise linear degradation model with an early RUL threshold of 125 cycles, and segment the time series into windows of length 30 cycles (30 × 14 matrices). The HQRNN architecture consists of three stacked QLSTM layers with latent dimensions of 32, 16, and 8, followed by classical Dense layers that reduce the dimension from 8 × W to 16, then 32, and finally to a single RUL value. Each QLSTM gate uses a QDI layer built on a 4-qubit circuit with parameterized Rx rotations, CNOT gates, and Rz encoding gates, with the blue encoding block repeated once. The observable is the Pauli Y matrix.

Training uses mean squared error loss, an Adam optimizer with batch size 128 and learning rate 0.001, for 20 epochs. The training dataset is split with 20% for validation. Both the HQRNN and a purely classical RNN with identical parameter counts are trained with 10 different random seeds, and predictions are averaged.

Key experimental results show that the HQRNN achieves a mean RMSE of 15.46 and mean MAE of 12.25 on the test set, with best RMSE of 14.78 and best MAE of 11.51, using 6793 parameters. This outperforms the classical RNN-32-16-8-16-32 model (mean RMSE 16.71, mean MAE 13.18, 14609 parameters) by approximately 5% in both metrics, despite having fewer trainable parameters. The authors note that while the classical model achieved slightly lower losses on training and validation sets, the HQRNN demonstrated superior generalization on the test set, suggesting it is less prone to overfitting and better at capturing diverse frequency components, especially with limited data (only 100 engines).

The paper includes a detailed quantum circuit analysis from three perspectives. First, ZX calculus redundancy analysis shows the QDI circuit is ZX-irreducible, meaning all 8 parameters are essential and no simplification is possible without altering functionality. Second, Fisher Information analysis reveals that eigenvalues are well-distributed with no significant concentration near zero, indicating resilience to barren plateaus, and the diagonal structure of the averaged FIM shows gradients are approximately evenly allocated across parameters with weak cross-parameter entanglement. Third, Fourier analysis shows that out of 161 possible frequency components, 109 (∼67%) have non-negligible amplitude, demonstrating substantial expressivity and the ability to capture higher-frequency components crucial for complex time-series data.

Comparative analysis shows the HQRNN outperforms classical ML models (Random Forest RMSE 17.91, LASSO 19.74, SVM 40.72, KNR 20.46, Gradient Boosting 15.67) and simple neural networks (MLP 16.78, CNN 18.45, LSTM 16.14), achieving approximately 13.68%, 16.21%, and 7.87% improvements over RF, CNN, and MLP respectively. However, advanced joint architectures still outperform it: Transformer + TCNN (12.31), CNN + LSTM (16.16), LSTM + FCLCNN (11.17), BLS + TCN (12.08), and Auto RUL + LSTM (7.78). The authors note that these joint models use more sophisticated degradation models and complex feature preprocessing, suggesting the HQRNN could be integrated into such pipelines for further improvement.

The paper concludes that the proposed model can be used not only independently but also integrated into complex ML pipelines. Future directions include integrating quantum modules into established forecasting algorithms like random forests or gradient boosting, exploring transformer-based approaches with quantum enhancements, and developing adaptive strategies that dynamically vary circuit size or depth to manage the trade-off between representational capacity and trainability.

Improvements for AI systems

Based on the paper, here are the specific improvements I can make to AI systems and what the improved system can do:

Improvement: Replace the linear transformations in LSTM gates with Quantum Depth-Infused (QDI) circuits that encode inputs into a 4-qubit Hilbert space using parameterized Rz rotations, Rx rotations, and CNOT gates, with Pauli-Y measurement.

What the improved system can do:

Achieve up to 5% lower RMSE and MAE compared to classical LSTM networks with the same parameter count
Capture high-frequency temporal components that classical models miss, due to the Fourier-space operation of QDI circuits
Generalize better on limited data (100 engines), reducing overfitting risk

Improvement: Use QDI layers that are ZX-irreducible (verified via ZX calculus), meaning all 8 parameters per gate are essential with no redundancy, while maintaining trainability confirmed by Fisher Information analysis.

Improvement: Use a QDI circuit where 67% of possible Fourier frequency components (109 out of 161) are accessible with non-zero amplitude, enabling representation of complex nonlinear degradation patterns.

Improvement: Stack three QLSTM layers (dimensions 32, 16, 8) followed by classical dense layers (8×30→16→32→1), using Adam optimizer with batch size 128, learning rate 0.001, and 20 epochs.

The improved AI system can:

Predict remaining useful life of jet engines with RMSE 15.46, outperforming classical ML and simple neural network baselines by 7.87%–16.21%
Learn from limited, noisy sensor data with fewer parameters, reducing overfitting
Capture both low- and high-frequency degradation patterns via quantum Fourier-space representation
Be integrated into more complex joint models (e.g., with attention mechanisms or temporal convolutions) to potentially close the gap with state-of-the-art results (RMSE 7.78–12.31)

Abstract

Predictive maintenance in aerospace heavily relies on accurate estimation of the remaining useful life of jet engines. In this paper, we introduce a Hybrid Quantum Recurrent Neural Network framework, combining Quantum Long Short-Term Memory layers with classical dense layers for Remaining Useful Life forecasting on NASA's Commercial Modular Aero-Propulsion System Simulation dataset. Each Quantum Long Short-Term Memory gate replaces conventional linear transformations with Quantum Depth-Infused circuits, allowing the network to learn high-frequency components more effectively. Experimental results demonstrate that, despite having fewer trainable parameters, the Hybrid Quantum Recurrent Neural Network achieves up to a 5% improvement over a Recurrent Neural Network based on stacked Long Short-Term Memory layers in terms of mean root-mean-square error and mean absolute error. Moreover, a thorough comparison of our method with established techniques, including Random Forest, Convolutional Neural Network, and Multilayer Perceptron, demonstrates that our approach, which achieves a Root Mean Squared Error of 15.46, surpasses these baselines by approximately 13.68%, 16.21%, and 7.87%, respectively. Nevertheless, certain advanced joint architectures still outperform it. Our findings highlight the potential of hybrid quantum-classical approaches for robust time-series forecasting under limited-data conditions, offering new avenues for enhancing reliability in predictive maintenance tasks.