A Fast and Generalizable Fourier Neural Operator-Based Surrogate for Melt-Pool Prediction in Laser Processing

arXiv:2602.06241 · cs.LG, cs.CE · Submitted 2026-03-31 · 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 "A Fast and Generalizable Fourier Neural Operator-Based Surrogate for Melt-Pool Prediction in Laser Processing".

Jane: The paper was written by Alix Benoit, Toni Ivas, Mateusz Papierz, Asel Sagingalieva, Alexey Melnikov et al. from Laboratory for Advanced Materials Processing, Swiss Federal Laboratories for Materials Science and Technology and Terra Quantum AG, Kornhausstrasse, St. Gallen, Switzerland..

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

Summary and Core Implications: Tom: So, moving beyond the scope, what did they summarize as the core achievement in "A Fast and Generalizable Fourier Neural Operator-Based Surrogate for Melt-Pool Prediction in Laser Processing"?

Jane: The summary tells us that this model maps process parameters—things like power and scan speed—to three-dimensional temperature fields and melt-pool boundaries across a broad process window. It covers everything from simple conduction modes to the complex keyhole regime.

Lu: That’s massive because, as you know, these two regimes are fundamentally different in terms how the heat flows and how much vapor is involved. To capture both in one unified model is a huge leap forward for AI representation.

Meng: The core implication for me is that this ability to predict across a wide parameter window means we can design control systems that operate smoothly across those vast ranges without sudden failures or poor performance.

Lalam: It seems like the summary highlights a shift from just having data to being able to achieve robust, real-time prediction, which facilitates a culture of precision in manufacturing.

Tom: It's not just predicting temperature; it' definitely about predicting the geometry—the melt pool boundaries—simultaneously.

Jane: Exactly, Tom. The model handles both the thermal state and the physical shape as a single output, which is something that previous models often had to address separately.

Lu: It’s a sophisticated way to encode physical intuition into a unified AI architecture, making the network far more robust than just inputting raw power and speed separately.

Meng: The practical result of this ability is that we can quickly identify optimal settings for any given process parameter combination without needing extensive physical testing.

Lalam: When we move toward such robust, fast surrogate modeling, we are facilitating a culture where the complexity of physics doesn's stop industrial progress.

Improvements and Methodology: Tom: Now, let's talk about the technical improvements they made in "A Fast and Generalizable Fourier Neural Operator-Based Surrogate for Melt-Pool Prediction in Laser Processing," especially since it handles both conduction and keyhole regimes.

Jane: The authors successfully reformulated the transient problem by moving it into a reference frame that moves with the laser, which is a huge conceptual step that simplifies the physics dramatically.

Lu: By doing this, they're essentially treating the complex heat source movement as a quasi-steady state problem, and they use temporal averaging to filter out those pesky high-frequency interface fluctuations. This allows for a steady AI training environment.

Meng: That concept of "quasi-steadiness" is crucial for engineering because it means we can apply established AI training paradigms even when dealing with inherently dynamic systems like melting metal. We're making the unmanageable manageable.

Lalam: It seems like the improved handling of dynamics allows us to build much more robust and predictable manufacturing processes, Lu's point about reliability resonates strongly with me in terms of global quality control.

Tom: And they use normalized enthalpy, H*, to ensure that this model can handle a vast range of process parameters without having to retrain for every new combination.

Jane: That’s the generalizability part—instead of mapping P and Vscan individually, they map based on this dimensionless parameter that captures the the physics across different regimes.

Lu: It’s a sophisticated way to encode physical intuition into a single normalized variable, making the network far more robust than just inputting raw power and speed.

Meng: This approach allows us to design control systems that operate smoothly across those vast parameter spaces without sudden performance drops, which is vital for industrial reliability.

Lalam: When we move toward such robust system design, we are facilitating a culture of precision and efficiency in global manufacturing processes.

Results and Super-Resolution: Tom: Let's talk about the results now. The performance metrics in "A Fast and Generalizable Fourier Neural Operator-Based Surrogate for Melt-Pool Prediction in Laser Processing" are very impressive, especially the speed difference. It’s a truly game-changing performance metric, right?

Jane: It is staggering. We’re talking about up to one hundred thousand times faster than traditional Finite Volume simulation, which allows us to see if the AI can even keep up with real-time process control requirements.

Lu: I find the one percent error rate for temperature prediction to be exceptionally good given the complexity of thermal gradients and phase changes involved in the keyhole regime. That shows incredible fidelity in the model's learning capacity.

Meng: And that level of accuracy, combined with an IoU score for melt-pool segmentation over zero point nine, means that if we use this model to identify where the metal pool is, it is extremely accurate for quality control purposes.

Lalam: This high level of accuracy translates directly into a reduction in scrap and material waste on the factory floor, leading to a more sustainable and efficient culture.

Tom: The results also show that the FNO has this "super-resolution" capability when we train it on coarser data but testing it on finer grids.

Jane: It’s like the AI learns the underlying physics at a certain resolution, and then it can predict what happens even if you ask for much higher detail than what was shown in the training data.

Lu: But, Meng pointed out, that's only as good as the mesh convergence of that training data—it doesn's magically create physics that weren't there to begin with. We are limited by our initial understanding of the physical system.

Meng: That limits our expectations for super-resolution; we can only refine what we know is physically possible from a coarser input grid, which is a practical constraint on my end.

Lalam: Even with those limitations, the ability to strive for higher resolution is itself pushing us toward a culture of relentless optimization and high standards in production.

Conclusion and Wrap-up: Tom: So, as we wrap up our discussion on "A Fast and Generalizable Fourier Neural Operator-Based Surrogate for Melt-Pool Prediction in Laser Processing," we’ve seen a model that is both extremely fast and capable of handling complex welding dynamics.

Jane: It truly is a powerful, fast, and efficient surrogate modeling framework that can provide near-instantaneous predictions across the entire process window.

Lu: I think the path forward into parameter-efficient spectral mixing using quantum circuits, which they mentioned in the conclusion, is where the future of this work really lies for me.

Meng: I'm just focused on how practical this is; it allows us to build digital twins that are actionable and reliable enough to implement these industrial process controls right now.

Lalam: We are facilitating a culture where speed meets precision in manufacturing, ensuring that the next generation of industry is both powerful and highly efficient.

Tom: It’s a model that allows us to understand the physics deeply while making it incredibly practical for real-time control, which is exactly what we hoped for.

Jane: It's great to hear all your perspectives on this exciting leap in AI application within industrial processes, everyone.

Lu: I think the future of this work is definitely in leveraging these operators to really push the boundaries of how we model complex systems globally.

Meng: For me, ensuring a reliable and fast prediction is absolutely vital for making sure that's what it sounds like in a real-world application.

Lalam: This work on "A Fast and Generalizable Fourier Neural Operator-Based Surrogate for Melt-Pool Prediction in Laser Processing" has certainly inspired us to think about the future of efficiency itself.

Tom: Well, listeners, that’s all the time we have for this segment today. Thanks to everyone!

Alix Benoit, Toni Ivas, Mateusz Papierz, Asel Sagingalieva, Alexey Melnikov, Elia Iseli

Laboratory for Advanced Materials Processing, Swiss Federal Laboratories for Materials Science and Technology · Terra Quantum AG, Kornhausstrasse, St. Gallen, Switzerland.

cs.LG, cs.CE

Submitted: 2026-03-31

Updated: 2026-08-20

Code: https://github.com/NVIDIA/physicsnemo

Importance score: 90/100

The gist: The paper presents the Laser Processing Fourier Neural Operator (LP-FNO), which is described as "a Fourier Neural Operator (FNO) based surrogate model that learns the parametric solution operator of

Key concepts

Melt-Pool Prediction: The model predicts both the thermal state (temperature fields) and the physical shape (melt-pool boundaries) simultaneously. It maps process parameters like power and scan speed to these outputs, enabling quick identification of optimal settings for laser processing.
Fourier Neural Operator (FNO): This is the core AI architecture used in the paper. It functions as a surrogate model, providing near-instantaneous predictions of complex physical systems, such as heat flow in melting metal. Its speed makes it suitable for real-time industrial process control.
Generalizability: The model uses a dimensionless parameter called normalized enthalpy ($H^*$). This approach allows the AI to handle vast ranges of process conditions and different physical regimes without needing extensive retraining for every new combination of input parameters.

Terminology

Fourier Neural Operator — 20× in this paper
Fidelity — 14× in this paper · explained in 2 episodes
Accuracy — 8× in this paper · explained in 1 episode
Core — 7× in this paper · explained in 2 episodes
Machine Learning — 6× in this paper
Digital Twin — 5× in this paper · explained in 4 episodes
Generalization — 5× in this paper · explained in 13 episodes
DeepONet — 4× in this paper · explained in 2 episodes
Quantum Circuits — 3× in this paper · explained in 2 episodes
Activation Function — 2× in this paper · explained in 1 episode
Uncertainty Quantification — 2× in this paper · explained in 8 episodes
Volume of Fluid — 2× in this paper
Direct Energy Deposition — 1× in this paper
Entanglement — 1× in this paper · explained in 3 episodes
Finite Volume Method — 1× in this paper
Generalizability — 1× in this paper · explained in 6 episodes
Inverse Problem — 1× in this paper · explained in 2 episodes
Latent Representation — 1× in this paper · explained in 2 episodes
Loss Function — 1× in this paper · explained in 2 episodes
Quantum Physics — 1× in this paper · explained in 10 episodes
alpha — 1× in this paper · explained in 1 episode

Summary

The paper presents the Laser Processing Fourier Neural Operator (LP-FNO), which is described as a Fourier Neural Operator (FNO) based surrogate model that learns the parametric solution operator of various laser processes from multiphysics simulations generated with FLOW-3D WELD®. This approach addresses the limitation that "high-fidelity simulations of laser welding capture complex thermo-fluid phenomena, including phase change, free-surface deformation, and keyhole dynamics, however their computational cost limits large-scale process exploration and realtime use."

The core innovation involves a method for handling the transient nature of the problem: "Through a novel approach of reformulating the transient problem in the moving laser frame and applying temporal averaging, the system results in a quasi-steady state setting suitable for operator learning, even in the keyhole welding regime." This allows for efficient operator learning across different welding regimes.

The LP-FNO is designed to map process parameters to three-dimensional temperature fields and melt-pool boundaries. The model covers a broad process window spanning conduction and keyhole regimes using the non-dimensional normalized enthalpy formulation. The performance metrics achieved by this model are highly accurate: The model achieves temperature prediction errors on the order of 1% and intersection-over-union scores for melt-pool segmentation over 0.9.

The efficiency of the LP-FNO is a major highlight. It "provides an efficient surrogate modeling framework for laser welding, enabling prediction of full three-dimensional fields and phase interfaces over wide parameter ranges in just tens of milliseconds, up to a hundred thousand times faster than traditional Finite Volume multi-physics software."

Methodologically, the simulations were performed using the commercial software FLOW-3D WELD® 2025R1. The physical model incorporates complex dynamics, such as modeling evaporation through the added recoil pressure source term in the Navier–Stokes equations and utilizing a Volume of Fluid (VoF) method to track surface tension and incorporate phase change.

To ensure generalization across welding regimes, the study used a parameter sampling strategy based on normalized enthalpy (H*), which allows for an equally spaced grid in H* and P, covering everything from lack of fusion (LoF) to keyhole.

The FNO architecture is employed to approximate the solution operator G, which maps input parameters A directly to the solution field U. The FNO utilizes a structure where the Fourier layer... corresponds to a global convolution with a translation-invariant kernel.

To handle the time dependency, the data was transformed into a moving laser reference frame, where time steps were defined such that the laser position coincides with grid points at every time step in the moving reference frame. Furthermore, to justify treating the melt pool as quasi-static, "a temporal averaging procedure is applied over a sliding window... This demonstrates that, once high-frequency interface oscillations are filtered out, the temperature field and melt-pool geometry exhibit only weak residual temporal dependence after the laser has traveled a sufficient distance."

The LP-FNO also incorporates measures to maintain physical integrity. A smooth masking procedure was applied to suppress fields in the surrounding gas phase using a smooth hyperbolic tangent gate, ensuring that the temperature field is then smoothly blended toward a fixed value and preventing spurious sharp gradients at the keyhole interface.

Regarding performance, the model demonstrates high accuracy across various metrics:

For 10 µm mesh, LP-FNO inference takes approximately 10 ms (0.01 s).
The average absolute temperature error is reported at around 3%.

The intersection over union (IoU) for the metal–gas interface (alpha) is near perfect, with a mean of 0.9992.

A critical feature of the FNO is its resolution-invariant nature. The paper demonstrates that a LP-FNO model trained on coarse-resolution data can be evaluated on a much finer mesh, yielding accurate super-resolved predictions in meshconverged conduction regimes. However, the limitations regarding this capability are noted: "The visible discrepancy between coarse and fine FLOW-3D solutions indicates that mesh convergence has not yet been achieved at 10 µm resolution in the keyhole regime. This shows that while useful, FNO super-resolution can only be as good as its training data fidelity."

In conclusion, the LP-FNO provides a powerful, fast, and efficient surrogate modeling framework for laser welding, capable of delivering near-instantaneous predictions of full three-dimensional fields over broad parameter ranges, making it highly suitable for applications like rapid process optimization, uncertainty quantification, and real-time digital twins.

Improvements for AI systems

As a fastidious researcher, I have analyzed the methodology of the LP-FNO (Laser Processing Fourier Neural Operator) to identify specific architectural enhancements and applications that move beyond its current state-of-the-art application in laser welding, thereby significantly improving broader AI systems designed for complex physical simulation.

The core strength of this paper—the ability to map a parameter space (A) to an operator (G) rather than just learning the solution for a is fixed—is the foundation of its value. The improvements outlined below focus on generalizing this paradigm, enhancing robustness, and addressing inherent limitations in current applications.

Improvement: Formalizing and applying the Moving Reference Frame + Temporal Averaging technique (Equations 11 & 12) to dynamic, high-speed processes that are not inherently quasi-steady (e.g., chemical kinetics, high-velocity fluid mixing, or rapid material crystallization).

What the Improved System Can Do:

Eliminate the Need for Fixed Time Steps: The system can simulate transient phenomena by treating the travel distance or process duration as the primary evolution parameter, rather than fixed physical time.
Real-Time Dynamic Prediction: It allows for real-time process control and optimization in systems where traditional FVM/FEM would require prohibitive computational resources, enabling instantaneous prediction of steady-state structures even when the underlying physics are inherently unsteady (e)g., predicting the stable interface shape of a rapid chemical reaction mixture.

Feature Original LP-FNO Capability Improved System Capability

:---:---:---

Core Function Map Process Parameters to Solution Fields (Operator Learning) Map Process Parameters to Solution Fields (Operator Learning) + Physics Constraints.

Dynamic Range Quasi-steady approximation via time averaging in a specific process (laser welding). Real-time prediction for any transient, high-speed physical system by treating evolution as a parameter.

Fidelity/Accuracy Excellent accuracy in the conduction regime; limited super-resolution fidelity in keyhole regimes. Guaranteed accuracy across all regimes via multi-scale progressive training; true super-resolution.

Generalization Generalizes across parameter space without retraining. Generalizes robustly, even when extrapolating into unknown parameter spaces (aided by physics enforcement).

Computational Scaling High cost for high complexity/high mesh resolution due to dense spectral mixing. Highly scalable and memory-efficient, allowing complex multi-physics simulation at low computational cost.