Hybrid Fourier Neural Operator for Surrogate Modeling of Laser Processing with a Quantum-Circuit Mixer

This paper introduces HQ-LP-FNO, a hybrid quantum-classical Fourier Neural Operator that utilizes a compact variational quantum circuit mixer to reduce trainable parameters by 15.6% and improve prediction accuracy for three-dimensional laser processing surrogate modeling compared to classical baselines.

The Gist

The Big Picture: Predicting Laser Welding Without the Wait

Imagine you are a master chef trying to perfect a new recipe for laser-welding metal. You need to know exactly how the metal will melt, flow, and cool down under different settings (like laser power and speed).

To get this right, scientists usually use super-computers to run complex physics simulations. Think of these simulations like running a full-scale, slow-motion movie of the metal melting. They are incredibly accurate, but they take hours or even days to run. If you want to test 1,000 different settings, you'd be waiting for months.

The Goal: The researchers wanted to build a "shortcut" (a surrogate model) that acts like a crystal ball. It should predict the result of the laser welding instantly (in milliseconds) with high accuracy, so engineers can design better processes in real-time.

The Problem: The "Heavy" Shortcut

The team had already built a shortcut using a type of AI called a Fourier Neural Operator (FNO). Think of an FNO as a very smart translator that learns the "language" of heat and fluid flow.

However, this translator had a problem: it was too heavy.

• The Analogy: Imagine the translator is a librarian who has to read every single book in a massive library to answer one question. The more books (data modes) you add to the library, the more time and memory the librarian needs.
• In technical terms, the AI's "brain" (parameters) grew too large as the problem got more complex (3D space). This made it too slow and expensive to run on standard computers for real-time use.

The Solution: A Quantum "Smart Assistant"

The researchers introduced a new hybrid model called HQ-LP-FNO. They didn't just make the librarian faster; they replaced a specific part of the librarian's job with a Quantum Assistant.

Here is how they did it:

1. The "Mode-Shared" Trick:
   In the old model, the AI had to learn a unique rule for every single frequency of the wave (like learning a different dialect for every city).

   • The New Way: They realized they could use one single, compact rule to handle many frequencies at once.
   • The Analogy: Instead of hiring 100 different translators for 100 different cities, they hired one genius translator who can speak all 100 dialects fluently using a single, efficient method.

2. Enter the Quantum Circuit (The VQC):
   To be that "genius translator," they used a Variational Quantum Circuit (VQC).

   • The Analogy: Imagine a classical computer is a standard calculator. A quantum computer is like a magic kaleidoscope. When you look through it, the pieces (data) rearrange themselves in complex, interconnected patterns that a standard calculator would struggle to see without a massive amount of effort.
   • This quantum "kaleidoscope" is tiny and efficient. It doesn't need to grow bigger just because the problem gets bigger. It handles the complex mixing of data using fewer "ingredients" (parameters).

The Results: Smaller, Faster, and Smarter

They tested this new hybrid model on a dataset of laser welding simulations (specifically for Titanium alloy). Here is what happened:

• Shrunk the Brain: They reduced the number of "brain cells" (trainable parameters) in the AI by 15.6%. It's like taking a heavy backpack and removing 15% of the weight, making it much easier to carry.
• Improved Accuracy: Surprisingly, making it smaller made it smarter.
  • The error in predicting the temperature dropped significantly (from ~2.9% to ~2.6%).
  • The error in predicting the melted metal fraction dropped by 26%.
• The "Goldilocks" Zone: They tested different sizes of the quantum assistant. They found that a medium-sized quantum helper worked best. It wasn't too small to be useless, and not too big to get confused. It was just right.

Why This Matters

This paper proves two big things:

1. Hybrid is the Future: You don't need a full-blown quantum computer to get quantum benefits. You can plug a tiny quantum "module" into a standard AI to make it more efficient.
2. Efficiency is Key: By using a quantum circuit to share the workload across different frequencies, they solved the problem of "bloat" in 3D simulations.

The Bottom Line

The researchers built a lightweight, high-speed crystal ball for laser welding. By swapping a bulky, repetitive part of the AI with a sleek, quantum-powered "smart mixer," they created a tool that is smaller, faster, and more accurate than the previous version. This paves the way for real-time digital twins in manufacturing, where engineers can simulate and optimize laser processes instantly, rather than waiting days for a computer to catch up.

Technical Summary

1. Problem Statement

High-energy laser processing (e.g., welding, powder-bed fusion) involves complex, tightly coupled multiphysics phenomena including heat transfer, melt-pool convection, free-surface deformation, phase change, and keyhole dynamics. While high-fidelity multiphysics solvers (like FLOW-3D WELD) can simulate these processes accurately, their computational cost prohibits real-time control, uncertainty quantification, and systematic process-window exploration.

Data-driven surrogates, specifically Fourier Neural Operators (FNOs), offer a solution by learning mappings between function spaces. However, standard 3D FNOs face a critical scalability bottleneck:

• Parameter Explosion: The dense, mode-wise spectral channel mixing in FNO layers scales linearly with the number of retained Fourier modes ($N_m$).
• Deployment Limitation: This linear scaling inflates parameter counts, making it difficult to deploy compact, accurate surrogates for real-time applications or on edge devices.
• Research Gap: It remains unclear whether independent parameters are required at every frequency, or if a mode-shared mixer (parameters independent of mode count) can achieve comparable fidelity with fewer resources. Furthermore, no rigorous protocol existed to evaluate hybrid quantum-classical operators in this specific physics domain.

2. Methodology

The authors propose HQ-LP-FNO (Hybrid Quantum Laser Processing Fourier Neural Operator), a hybrid architecture that replaces a configurable fraction of the dense spectral mixing blocks in a classical FNO with a Variational Quantum Circuit (VQC) mixer.

A. Data and Simulation Setup

• Dataset: Generated using FLOW-3D WELD 2025R1 for single-track Ti–6Al–4V processing.
• Physics: Coupled incompressible flow, heat transfer, phase change, and recoil pressure.
• Process Window: Covers conduction, lack-of-fusion, and keyhole regimes, sampled via a normalized enthalpy criterion ($H^*$).
• Preprocessing: Transient data is transformed into a quasi-steady moving frame (laser reference frame) with a sliding temporal average to isolate large-scale melt-pool structures.

B. Hybrid Architecture (HQ-LP-FNO)

The core innovation lies in the Partitioned Hybrid Spectral Convolution:

1. Channel Partitioning: For each retained Fourier mode, the output channels ($C$) are split into a classical subset ($C_c$) and a quantum subset ($C_q$).
2. Classical Branch: The remaining $C_c$ channels are processed by standard dense, mode-wise complex spectral weights (scaling with $N_m$).
3. Quantum Branch: The first $C_q$ channels are processed by a mode-shared VQC.
   • Input: Real and imaginary parts of the complex Fourier coefficients are concatenated and normalized.
   • Circuit Structure: A QFT–Mixer–Inverse-QFT structure. The mixer uses a depth-$d$ nearest-neighbor chain with odd-even pairing, utilizing IsingXY gates for entanglement and local $R_Z$ phases.
   • Parameter Sharing: The VQC parameters are shared across all retained Fourier modes and all samples, meaning the parameter count for this branch is independent of the Fourier mode budget ($N_m$).
4. Control Baseline (CM-LP-FNO): To isolate the effect of "mode-sharing" from "quantum-specific" inductive bias, the authors introduce a purely classical control. This replaces the VQC with a parameter-matched bottleneck MLP (Multi-Layer Perceptron) that shares parameters across modes in the exact same way.

C. Evaluation Protocol

• Metrics: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Intersection over Union (IoU) for temperature ($T$), metal fraction ($\alpha$), and liquid fraction ($f_l$).
• Noise Validation: A noisy-simulator study using backend-calibrated noise from IBM Torino was conducted to assess numerical stability under realistic hardware constraints.

3. Key Contributions

1. First Application: The first application of hybrid quantum-classical operator learning to high-energy laser processing, a domain characterized by sharp phase interfaces and nonlinear thermo-fluid dynamics.
2. Parameter-Efficient Spectral Mixing: Demonstration that replacing dense mode-wise mixing with a mode-shared VQC reduces trainable parameters by 15.6% while maintaining or improving accuracy.
3. Rigorous Ablation Study: The introduction of the CM-LP-FNO control allows for a precise attribution of performance gains, confirming that the mode-shared mixing principle (not just the quantum nature) is the primary driver of improvement.
4. Hardware Viability: Validation of the quantum mixer's numerical stability under calibrated noise and finite-shot sampling, providing a benchmark for near-term quantum hardware deployment.

4. Key Results

Evaluated on a held-out test set spanning conduction and keyhole regimes:

• Parameter Reduction: HQ-LP-FNO reduced trainable parameters from ~184.3M (classical baseline) to ~155.5M (15.6% reduction).
• Accuracy Improvements:
  • Temperature: Relative Mean Absolute Error (Rel. MAE) improved from 2.89% (classical) to 2.56% (HQ-LP-FNO).
  • Phase Fraction: Phase-fraction MAE reduced by 26%.
  • Trade-off: While the VQC improved mean error, it showed slightly higher RMSE and lower IoU for melt-related fields compared to the classical control (CM-LP-FNO), suggesting the VQC excels at global trends but may struggle with sharp interface tails under the current L1-dominated loss.
• Optimal Partitioning: A sweep over the quantum-channel width ($C_q$) revealed that a moderate allocation ($C_q=3$) yielded the best temperature metrics across all configurations, including the fully classical baseline. This suggests an optimal balance where the VQC regularizes the dense blocks without introducing excessive noise or underfitting.
• Noise Robustness: The quantum mixer remained numerically stable across the tested shot range (up to 20k shots) under IBM Torino noise models, with error decreasing monotonically up to ~5000 shots.

5. Significance and Conclusion

This work establishes a controlled evaluation protocol for hybrid quantum operator learning in physics-based settings. It demonstrates that Variational Quantum Circuits can serve as effective, parameter-efficient spectral mixers that decouple model complexity from the resolution of the Fourier grid.

The findings suggest that:

1. Mode-sharing is key: The structural constraint of sharing parameters across modes is a powerful inductive bias that improves generalization and reduces overfitting in high-dimensional PDE surrogates.
2. Hybrid is optimal: A moderate quantum allocation (neither fully classical nor fully quantum) often yields the best performance, balancing the expressive power of the VQC with the regularization of classical dense blocks.
3. Feasibility: The architecture is robust enough for near-term hardware deployment, provided shot noise is managed.

Future work will focus on optimizing training objectives for interface-sensitive metrics (like IoU), exploring hardware-efficient circuit ansatz, and validating these results on actual quantum processors.