🔬 materials science

Efficient spectral Galerkin framework for nonlinear transient heat transfer in finite domains

This paper presents a spectral Galerkin framework that reformulates the nonlinear transient heat equation into a set of modal ordinary differential equations, enabling highly accurate, GPU-accelerated simulations of complex thermal processes like laser-metal interactions with significantly lower computational costs than traditional finite element methods.

Original authors: Théo Andrieux, Andreas Ntinos, Manas V. Upadhyay

Published 2026-09-22

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Théo Andrieux, Andreas Ntinos, Manas V. Upadhyay

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

Heat is a restless traveler. It moves through solids, liquids, and gases, carrying energy from hot spots to cool ones, driven by the simple urge to balance out. In the world of manufacturing, particularly when building metal parts layer by layer with powerful lasers, managing this heat is a matter of life or death for the final product. If the metal gets too hot, it can warp, crack, or melt away entirely; if it stays too cool, the layers fail to fuse. To predict exactly how the temperature will rise and fall, engineers rely on computer simulations. For decades, the most trusted method has been to chop the metal part into millions of tiny, irregular pieces and solve the heat equations for each one. This approach is incredibly accurate, but it is also painfully slow. It can take days of computing time to simulate a single laser pass, making it nearly impossible to test thousands of different scanning patterns to find the perfect one.

A team of researchers at the École Polytechnique in France has found a way to speed up this process by a factor of two hundred and twenty-seven without losing the accuracy needed to trust the results. They developed a new mathematical framework that treats the heat problem differently. Instead of breaking the metal into a messy jigsaw puzzle of tiny pieces, they view the entire part as a whole, vibrating like a drumhead. By using a specific set of smooth, wave-like patterns to describe the temperature, they can calculate how heat moves through the metal with far fewer steps. This method allows them to include complex, real-world effects—like the metal melting, the sudden release of hidden energy during phase changes, and the cooling effect of evaporation—while running simulations in minutes rather than days.

The challenge of simulating laser heating lies in the extreme nonlinearity of the physics involved. When a laser hits metal, the material properties change drastically as the temperature rises. The metal expands, its ability to conduct heat shifts, and if it gets hot enough, it melts and even evaporates. These changes happen simultaneously and depend entirely on the current temperature, creating a feedback loop that is notoriously difficult for computers to solve. Traditional methods handle this by repeatedly solving massive systems of equations, a process that requires immense computing power. The new approach, called the Spectral Galerkin framework, sidesteps this bottleneck. It starts by separating the problem into two parts: a simple, predictable baseline that describes how heat moves in a uniform material, and a set of "forcing" terms that account for all the messy, changing details like melting and evaporation.

The researchers realized that the simple baseline part could be solved exactly using a set of mathematical waves, similar to the standing waves you see on a plucked guitar string. Because these waves are mathematically independent of one another, the computer does not need to solve a giant, tangled system of equations. Instead, it can update each wave pattern individually and instantly. The complex, changing parts of the problem are then added back in through a rapid, iterative process. The computer guesses the temperature, calculates the messy effects, updates the waves, and repeats this cycle until the answer stabilizes. This structure is perfectly suited for modern graphics processors, the same chips found in gaming computers, which can perform millions of these independent calculations at the same time.

To test their method, the team simulated a laser beam scanning across a block of stainless steel, a process common in 3D printing. They included the full complexity of the physics: the laser heating the metal, the metal melting into a liquid pool, the latent heat released as it solidifies, and the heat lost to the air as the surface evaporates. They compared their results against a high-fidelity simulation using the traditional, slow method. The new framework reproduced the temperature fields with less than one percent error, capturing the steep temperature gradients near the laser and the shape of the molten pool with remarkable precision. While the traditional method took nearly 189 hours to run on a single processor core, the new method solved the same problem in just 30 hours on the same hardware. When run on a powerful graphics card, the time dropped to less than an hour.

The true value of this speed becomes apparent when the researchers applied it to a real-world design problem. They revisited a study on how to scan a large metal plate to ensure it cools down evenly, a critical factor in preventing warping. The original study used a simplified model that ignored melting and evaporation to compare different scanning patterns. The new, faster simulations allowed the researchers to run the same comparison but with the full, complex physics included. The results were striking. When evaporation and the hidden energy of melting were accounted for, the measure of temperature uniformity dropped by more than half compared to the original predictions. This suggests that the simplified models used in the past may have significantly overestimated how uneven the heat distribution would be.

The study also revealed that the specific pattern used to move the laser matters less than previously thought when the full physics are considered. While the ranking of the best and worst scanning strategies remained the same, the actual temperature differences were much smaller than the older models predicted. This implies that engineers might have been over-designing their processes to compensate for heat issues that are less severe than assumed, or conversely, that they have been missing subtle thermal effects that only appear when the full complexity of the material's behavior is captured. The ability to run these high-fidelity simulations quickly opens the door to testing thousands of variations, optimizing processes that were previously too computationally expensive to explore in detail.

This work does not claim to solve every problem in heat transfer. The method works best on regular, box-shaped domains and currently does not simulate the fluid flow inside the molten pool, which would require a different set of equations. However, for the vast majority of thermal problems in manufacturing where the shape is fixed and the focus is on temperature evolution, the framework offers a powerful new tool. It bridges the gap between the speed of simple, approximate calculations and the accuracy of slow, detailed simulations. By making high-fidelity modeling practical for everyday engineering use, it allows scientists and engineers to see the thermal behavior of their designs with a clarity that was previously out of reach, ensuring that the next generation of metal parts is built with a deeper understanding of the heat that shapes them.

Technical Summary: Efficient Spectral Galerkin Framework for Nonlinear Transient Heat Transfer

Problem Statement

Accurately modeling the temporal evolution of heterogeneous temperature fields in finite domains is challenging due to strong nonlinearities arising from temperature-dependent thermophysical properties, latent heats of transformation, and localized heat sources or sinks. While high-fidelity numerical approaches like Finite Element (FE) methods can resolve these complex behaviors, they require the repeated solution of large matrix problems, leading to prohibitive computational costs when fine spatial and temporal discretizations are necessary. Conversely, analytical and semi-analytical solutions are computationally efficient but are typically restricted to linear operators with constant coefficients, simple boundary conditions, or infinite/semi-infinite domains. There remains a significant gap between computationally intensive high-fidelity simulations and rapid analytical solutions for nonlinear thermal evolution in finite domains.

Methodology: The Spectral Galerkin (SG) Framework

The authors propose a unified Spectral Galerkin (SG) framework that bridges this gap by reformulating the fully nonlinear transient heat equation into a linear reference problem with constant thermophysical properties and a set of residual forcing terms.

1. Decomposition and Reference Problem:
The governing heat equation and Neumann boundary conditions are decomposed into:

A linear isotropic reference problem with constant density ( $\rho$ ), specific heat ( $c$ ), and thermal conductivity ( $k$ ).
Residual forcing terms ( $R(T)$ and $R_s(T)$ ) that account for all nonlinearities, including temperature-dependent property fluctuations, anisotropic conductivity, latent heats of fusion/vaporization, and nonlinear surface fluxes (e.g., evaporation, radiation).

2. Spectral Basis and Galerkin Projection:
The linear reference operator, subject to homogeneous Neumann boundary conditions on a cuboid domain, defines a classical eigenvalue problem. The eigenfunctions are analytically derived using the method of separation of variables, resulting in a complete, three-dimensional orthonormal trigonometric basis (cosine series).

The temperature field is expanded on this basis.
A Galerkin projection transforms the spatially discretized partial differential equation (PDE) into a system of modal ordinary differential equations (ODEs) in time.
Crucially, the reference operator is diagonal in this modal basis, meaning the modes are uncoupled for a fixed nonlinear forcing.

3. Time Integration and Iteration:
The resulting modal ODEs are integrated using an Exponential Time Differencing (ETD1) scheme. Since the nonlinear forcing terms depend on the unknown temperature field at the current time step, the method employs a Picard fixed-point iteration:

The nonlinear forcing is evaluated in physical space using the temperature field from the previous iteration.
This forcing is projected back onto the spectral basis.
The modal coefficients are updated analytically using precomputed exponential operators.
A relaxation factor is applied to ensure convergence of the fixed-point loop.

4. Implementation:
The structured-grid nature of the SG method allows for efficient implementation using Discrete Cosine Transforms (DCT) to switch between physical and spectral spaces. This structure is highly amenable to massive parallelization on Graphics Processing Units (GPUs).

Key Contributions

Unified Framework: The first combination of linear reference operators, spectral expansion, and exponential time integration to solve fully nonlinear, transient, anisotropic heat transfer problems in finite domains.
Computational Efficiency: The method eliminates the global algebraic solves required by implicit FE formulations, replacing them with mode-wise multiplications and spectral transforms.
GPU Acceleration: The structured grid formulation enables efficient GPU parallelization, achieving orders of magnitude speedup over CPU-based solvers.
Open Source: The authors provide the source code and worked examples under an Apache 2.0 license.

Results and Performance

The framework was validated and benchmarked against analytical solutions and high-fidelity FE simulations using a rapid laser-metal interaction case (316L stainless steel) involving melting and evaporation.

1. Accuracy:

Linear Regime: The SG solver matched the Eagar-Tsai semi-analytical solution with a relative $L_2$ error of less than 0.1%.
Nonlinear Regime: Compared to a high-fidelity nonlinear FE solver, the SG method reproduced temperature fields with a relative $L_2$ error of 0.67%. The peak temperature difference was less than 1% (31.3 K).

2. Computational Speed:

CPU Performance: On a single-core CPU, the SG solver was 6.2 times faster than the FE solver, despite using approximately one order of magnitude more degrees of freedom (25 million vs. 2.5 million).
GPU Performance: Leveraging GPU parallelization, the SG solver achieved a 227-fold speedup relative to the single-core FE baseline (0.83 hours vs. 189 hours for the same nonlinear simulation).

3. Application to Part-Scale Scanning:
The method was applied to a part-scale laser scanning study (Ramani et al., 2022) to evaluate thermal uniformity metrics.

When accounting for latent heat and evaporation (nonlinearities omitted in the original study), the proposed SG method revealed that the thermal uniformity metric was more than halved compared to the linear baseline.
While the ranking of scan strategies remained consistent, the quantitative energy balance was significantly altered, demonstrating the necessity of resolving nonlinear effects for accurate process optimization.

Significance and Limitations

The paper claims that the SG framework successfully bridges the gap between the computational efficiency of analytical approaches and the fidelity of high-resolution FE methods. It enables the retention of complex nonlinear physics (temperature-dependent properties, phase changes, evaporation) in problems where high-fidelity FE simulations are currently impractical due to cost.

Limitations noted by the authors:

Convergence Direction: Convergence is slowest in the direction normal to the surface with non-homogeneous fluxes (vertical direction) because the trigonometric basis cannot directly represent the non-zero surface flux; many modes are required to resolve the surface temperature.
Melt Pool Dynamics: The current framework does not resolve melt pool convection, Marangoni flow, or keyhole effects, as it does not solve the Navier-Stokes equations.
Domain Evolution: Simulating material addition (e.g., in additive manufacturing) requires projecting solutions onto the eigenbasis of an enlarged domain at each step, which is not yet fully automated in the current implementation.
Iteration Convergence: The Picard iteration converges linearly and may require many steps for strong nonlinearities; the authors suggest Anderson acceleration as a potential future improvement.

In conclusion, the SG method offers a highly efficient alternative for nonlinear transient heat transfer, providing a pathway to high-fidelity simulations of rapid thermal processes with significantly reduced computational time.