WorksMathematics
Localized Hybrid Analytical-Finite-Element Thermal Modeling for Laser Powder Bed Fusion
Abstract
Hybrid analytical–finite-element methods support scan-resolved, part-scale thermal simulation of laser powder bed fusion, but their analytical evaluation cost grows as scan-line contributions accumulate. This work develops a space–time localized hybrid formulation that restricts where and how long each contribution is evaluated analytically. Residual terms and boundary conditions derived from the heat equation allow the finite-element correction to account for the thermal effect of contributions removed from the analytical field. The formulation is assessed using a single-line benchmark against an independent full finite-element solution and multiline comparisons with the conventional hybrid formulation. Localization limits the number of analytical contributions retained while maintaining close agreement with the conventional solution. A comparison with uncompensated temporal truncation demonstrates the importance of the residual correction in retaining the thermal effect of removed contributions. In the 500-line case, analytical and total simulation times are reduced by factors of approximately 25.5 and 3.92, respectively, with a relative nodal 2-norm difference of 0.139%.
Introduction
Laser powder bed fusion (LPBF) is a metal additive manufacturing process in which a high-energy laser selectively melts successive layers of metal powder along a prescribed scan path (see Figure 1). Its ability to fabricate geometrically complex, high-accuracy components with high material efficiency has contributed to its widespread adoption [1]. LPBF is increasingly used to produce end-use components for aerospace, biomedical, automotive, and energy applications [2, 3].

Figure 1. Representative fin geometry for part-scale LPBF thermal simulations, showing the moving laser and localized thermal response. The schematic was generated with assistance from Claude Opus 5 (Anthropic) and reviewed by the authors.
LPBF involves complex thermal behavior arising from the highly localized moving heat source and the associated rapid heating and cooling rates, which produce steep temperature gradients within the material [4, 5, 6]. The resulting thermal history influences melt-pool behavior, microstructural evolution, residual-stress formation, and part distortion, thereby affecting the mechanical performance and structural integrity of the fabricated component [7, 8]. Experimental characterization of this thermal history is challenging because full-field temperature measurements are limited, subsurface and rapidly evolving thermal fields are difficult to capture, and extensive testing over large process-parameter spaces is costly [9, 10]. Consequently, predictive thermal simulations are essential for evaluating process conditions and understanding part-scale thermal behavior beyond what can be measured experimentally.
LPBF is a multiscale and multiphysics process in which different phenomena dominate at different spatial and temporal scales, leading to distinct modeling objectives and computational requirements [11, 12, 13, 14, 15]. At the microscale, models examine microstructural phenomena such as solidification, grain growth, and phase transformations, requiring very fine spatial and temporal resolution [7]. At the mesoscale, studies typically examine melt-pool thermodynamics and fluid flow along individual scan tracks, together with defect formation such as porosity evolution [16]. At the macroscopic or part scale, the objective shifts to predicting the accumulated thermal history over the full build and its thermomechanical consequences, including residual-stress accumulation and part distortion [15, 17, 18]. To remain computationally feasible, part-scale models employ simplified representations of localized melt-pool physics. The present work therefore focuses on part-scale transient heat conduction rather than explicitly resolving fluid flow and phase-change phenomena within the melt pool.
A wide range of approaches has been developed to predict thermal histories in LPBF, with different trade-offs between physical fidelity and computational efficiency. Fully numerical methods, including finite-element analysis (FEA), can accommodate complex geometries and nonlinear material behavior but may become prohibitively expensive at the part scale because the localized moving heat source requires fine spatial and temporal resolution [7]. Analytical moving heat-source solutions provide computationally efficient temperature predictions but are generally restricted to idealized domains, simplified material behavior, and a limited class of boundary conditions [19]. Hybrid analytical–numerical methods combine these complementary capabilities by representing the heat-source response analytically and using a numerical correction to account for the finite part geometry and prescribed boundary conditions [20, 21]. By capturing the steep near-source temperature gradients analytically, these methods can reduce the resolution required by the numerical mesh. Different formulations use different analytical source solutions and numerical methods to compute the correction.
The present work considers a hybrid formulation based on the transient Eagar–Tsai solution [19] and a finite-element correction. After a scan line ends, the deposited heat continues to diffuse and affect the temperature. As long as this effect is represented by the line’s analytical solution, that solution must be reevaluated at later time steps. Retaining all earlier lines therefore increases the analytical workload as scanning progresses. Controlling this cost requires deciding where and for how long each contribution is evaluated analytically, while accounting for its thermal effect when it is removed. Related approaches to managing analytical contributions, including those developed within other formulations, are reviewed in § 2.
To address this challenge, the present work introduces a space–time localized analytical–finite-element formulation. Each analytical scan-line field is multiplied by smooth spatial and temporal localization functions, restricting its explicit evaluation to a neighborhood of the scan segment and a limited time interval. Residual terms and boundary conditions derived from the heat equation account for the changing analytical weights through the finite-element correction. The resulting formulation reduces analytical evaluation in space and time while maintaining consistency with the governing heat equation.
The remainder of the paper is organized as follows. § 2 reviews numerical, analytical, and hybrid approaches for LPBF thermal simulation and strategies for managing analytical contributions. § 3 presents the analytical basis of the hybrid method and develops the proposed space–time localized formulation. § 4 presents the numerical validation and multiline studies. Finally, § 5 summarizes the principal findings and future directions.
Literature Review
This section reviews the developments that motivate the present work. It first examines part scale numerical thermal models and the strategies used to reduce their computational cost. It then discusses analytical moving-source solutions and their combination with numerical boundary corrections, including hybrid analytical–finite-element formulations for finite geometries. Finally, it considers the increasing cost of accumulated analytical contributions and the approaches developed to manage the retained scan history.
Part-Scale Numerical Thermal Modeling
Detailed LPBF models commonly use finite element, finite volume, or computational fluid dynamics formulations and may include material properties that vary with temperature, phase change, fluid flow in the melt pool, and thermomechanical coupling [7, 12]. Resolving the steep thermal gradients around the rapidly moving laser requires a spatial resolution comparable to the laser spot size and small time increments along the scan path. Extending this resolution to a part containing many tracks and layers results in very large computational requirements. Detailed multiphysics simulations are therefore generally limited to individual tracks, melt pools, or a small number of layers, while simulations of complete parts rely on simplified physical descriptions and specialized numerical strategies [22].
Several reduction and acceleration strategies have been developed to make thermal simulation of complete parts tractable. Adaptive remeshing performed by layer maintains finer resolution near the currently deposited material while using a coarser representation of previously solidified regions [23]. Dugast et al. developed a GPU-based matrix-free finite-element solver that uses adaptive meshing in the build direction and excludes inactive elements from GPU memory transfer. They also examined layer lumping as a further cost-reduction strategy [24]. Refinement can also be updated as scanning progresses rather than maintained across the entire active layer. Olleak et al. combined scanwise adaptive remeshing with a GPU-based matrix-free finite-element solver, concentrating fine resolution around selected scan tracks [25]. Liu et al. combined a characteristic time-based heat input model with adaptive mesh remapping to simulate thermal histories at track-scale resolution while reducing computational cost [26]. These strategies reduce spatial or temporal discretization requirements or accelerate numerical calculations. For methods that use remeshing, these savings come with the additional work of managing mesh changes and transferring the temperature field between successive meshes.
Methods based on multiple resolution levels provide another way to separate the fine thermal response around the laser from the smoother temperature history across the part. In the GO-MELT framework, fine, intermediate, and coarse discretizations are coupled and evaluated using GPU acceleration [27]. Elahi et al. subsequently extended GO-MELT with temperature-dependent material properties and time-step subcycling between levels for part-scale LPBF thermal simulations [28]. When the localized moving-source response is resolved directly, fine numerical resolution is still required near the laser, whether through mesh refinement or coupled discretizations. An alternative is to represent this response analytically, reducing the need for the numerical mesh to resolve the laser length scale directly. Analytical and hybrid analytical–numerical methods are reviewed next.
Analytical and Hybrid Thermal Models
Analytical thermal models provide an alternative way to predict temperature fields generated by moving heat sources. Their development originates from classical welding theory, particularly the moving point and line source solutions introduced by Rosenthal [29]. Eagar and Tsai later considered a Gaussian distribution of heat over the surface of a semi-infinite domain [19]. This distribution accounts for the finite width of the applied heat flux and provides a more suitable description of a laser source than an ideal point source. Classical solutions generally assume linear heat conduction, constant material properties, and an infinite or semi-infinite domain. Their assumptions, capabilities, and limitations in welding and related laser processes are reviewed by Nasiri and Enzinger [30].
Analytical heat source models have subsequently been applied to LPBF to predict temperature histories, melt pool dimensions, and cooling rates under conduction-dominated conditions [31, 32]. Scan paths can be modeled by superposing source contributions at successive positions or by combining the contributions of moving sources along individual scan lines. Wang et al. used analytical source solutions to model different scan strategies and the temperature accumulated across multiple tracks [33]. Zagade et al. extended analytical modeling to scan strategies with curved laser trajectories, including circular and spiral paths [34]. Steuben et al. also enriched a moving-source analytical solution to account approximately for finite and complex part boundaries, temperature-dependent material properties, and material accretion [35].
A moving-source solution derived for a semi-infinite domain does not, by itself, satisfy the boundary conditions of a finite part, particularly when the laser approaches an external surface. A common hybrid approach combines an analytical source field with a numerical correction that enforces the prescribed boundary conditions on the actual part. The analytical field describes the response to the moving source in a simple reference domain. By representing the steep temperature gradients near the laser analytically, this approach allows a coarser numerical mesh away from part boundaries, although finer resolution may still be required near those boundaries.
Yang et al. introduced a semi-analytical formulation in which the laser path was represented by a sequence of instantaneous point sources [20]. The associated analytical temperature fields were combined with image fields and a smooth numerical complementary field solved using finite differences to impose the boundary conditions of the finite body. In later work, the discrete point-source representation of each scan vector was replaced by a continuous analytical line-source solution, and the boundary conditions were enforced through a finite-element correction with mesh refinement near surfaces [36]. This avoided the construction and geometric restrictions of image sources. Moran et al. developed a superposition-based finite-element approach combining a transient Rosenthal-type solution for a moving point heat source in a semi-infinite domain with a finite-element correction to enforce the part boundary conditions. The formulation was verified against a converged conventional finite-element solution [37]. The method was subsequently applied to scan-by-scan simulation of a realistic component while preserving the individual scan vectors and their timing [21].
Recent developments have focused on improving the treatment of boundaries and complex part geometries. Yang et al. introduced modulated image sources to reduce the numerical resolution required near curved adiabatic boundaries [38]. More recently, they combined the analytical source representation with an isogeometric discretization of the correction field, allowing complex part boundaries to be represented accurately with fewer numerical degrees of freedom [39].
Computational Cost and Management of Analytical Contributions
Evaluating analytical heat-source fields can contribute substantially to the computational cost of scan-resolved thermal simulations. The workload depends on the source representation, the spatial evaluation points, and the treatment of contributions from earlier scans. If all scan contributions are retained and reevaluated analytically, the work at each time step increases as scanning progresses. Approaches to controlling this cost include more efficient source evaluation, reuse of stored solutions, and numerical representation of earlier thermal contributions. These strategies are discussed below.
In the point-source formulation of Yang et al. [20], evaluating the source and image fields accounted for most of the reported computational time. Sufficiently diffused source contributions and their associated image fields were added to the numerical correction and removed from subsequent source evaluation, using a transfer time based on thermal diffusivity and grid spacing. The subsequent continuous line-source formulation reduced the analytical cost of evaluating each scan vector [36].
Wolfer et al. [40] used a different semi-analytical formulation in which the source contribution over each time interval was evaluated analytically and the previous temperature field was propagated through Gaussian convolution. Stored line solutions were reused, and analytical evaluation was restricted to the vicinity of the laser. This avoided explicit superposition of the complete scan history, but did not use a finite-element correction field.
Moran et al. [21] reduced the retained analytical history by projecting completed scan contributions onto the finite-element correction once their temperature fields had become sufficiently smooth to be represented by the coarse mesh. These contributions were then removed from subsequent analytical evaluation. The transfer was described as maintaining thermal continuity, although the paper does not detail the projection procedure or provide a quantitative criterion for selecting the transfer time.
The present work develops a different treatment based on smooth spatial and temporal localization functions. Rather than adding or projecting completed analytical contributions onto a numerical field at discrete transfer events, the formulation gradually reduces their analytical weights and derives the corresponding residual terms and boundary conditions from the heat equation. Spatial localization also restricts where each retained contribution is evaluated. The numerical studies examine the resulting agreement and computational savings, while a comparison with uncompensated temporal truncation demonstrates the importance of accounting for removed contributions.
Theory and Methodology
Consider a finite LPBF domain with boundary , where . Under the assumptions of homogeneous and isotropic material properties, constant thermophysical parameters, and heat transfer by conduction, the temperature field satisfies
where , , and denote the density, specific heat capacity, and thermal conductivity, respectively, denotes the heat-source distribution associated with the moving laser, and is the Laplacian. For use in the hybrid formulation, define the linear heat operator
Equation (1) can then be written as
The governing equation is supplemented by the boundary and initial conditions
where is the outward unit normal, is the prescribed temperature on , is the prescribed outward heat flux on , and is the initial temperature field. In the present study, convective and radiative heat losses are neglected, so the free Neumann boundaries are treated as insulated, . The laser heat input is represented through and is therefore not applied again through the Neumann boundary condition.
The present study further neglects phase change and melt-pool fluid flow since the focus is on part-scale thermal modeling. Together with the preceding assumptions, these simplifications yield a linear thermal problem, allowing the temperature contributions produced by individual laser exposures to be superposed.
The following subsections present the analytical basis and numerical formulation of the proposed method. First, the Eagar–Tsai solution is introduced for evaluating the temperature field generated by a moving laser. Next, the localized analytical–finite-element formulation is developed by introducing spatial and temporal localization into the conventional hybrid decomposition. Finally, the weak form and its finite-element discretization are derived for numerical implementation.
Throughout this section, denotes a spatial point. Boldface symbols denote vectors and matrices, whereas scalar quantities are written in standard font.
Analytical solution for a moving Gaussian heat source
The analytical component of the present formulation is based on the classical Eagar–Tsai solution for a moving heat source in a homogeneous semi-infinite solid [19]. The analytical domain is bounded by a planar upper surface and extends infinitely in the depth and lateral directions. The formulation assumes homogeneous and isotropic material properties, constant thermophysical parameters, heat transfer by conduction, and negligible convective and radiative heat losses.
The laser is represented by a Gaussian heat-source distribution acting on the upper surface. Let denote the nominal laser power and the absorptivity. The total absorbed power is
For a source centered at the origin of the surface coordinate system, the corresponding Gaussian surface power density is
where is the in-plane radial distance from the source center and is the standard deviation of the Gaussian distribution. The distribution is normalized so that its integral over the upper surface equals . This planar Gaussian source model is used in the Eagar–Tsai solution and subsequent analytical LPBF thermal models [19, 33, 40].
Following Wolfer et al. [40], the source term in Eq. (1) can be written as
where , is the in-plane laser position, is the coordinate of the upper surface, and is the Dirac delta distribution. (9) represents a planar Gaussian source acting only on the upper surface rather than throughout the material volume.
Consider now a source that begins moving at time from the origin with constant velocity in the positive -direction. The corresponding transient Eagar–Tsai solution gives the temperature rise above the initial uniform temperature:
where
is the thermal diffusivity.
In the semi-infinite domain, the Eagar–Tsai temperature rise satisfies the heat equation with the Gaussian surface source defined in Eq. (9):
In this representation, the laser input is included in , and the Neumann condition on the planar surface is homogeneous.
(10) gives the temperature field generated by a source that begins at and remains active up to the observation time . This expression is next used to construct the contribution of a finite scan line with arbitrary position and orientation.
Let scan line start at time from the point and travel with constant speed in the unit direction . Since the scan line lies on the planar upper surface, let denote the in-plane rotation that aligns the scan direction with the local -axis, so that
The local coordinates of a global observation point are defined by
In this coordinate system, the scan direction coincides with the positive -axis, while remains the depth coordinate. At time , the beam center therefore has the local coordinates
In the local coordinate system, the contribution of a source beginning at and remaining active is denoted by and is obtained from Eq. (10) as
For a finite scan line, the laser is active only over the interval
where is the exposure duration. The endpoint of the scan line is
Using the same rotation matrix , define a second local coordinate system with its origin at :
Now consider a fictitious source that starts at and remains active for . Its temperature field is denoted by and is given by
For ,
Using Eq. (20),
with and . Therefore, the kernels of and are identical over the interval , and their difference removes the fictitious continuation of the laser after the physical scan line has ended.
For , the temperature contribution of the finite scan line is therefore
Equivalently, a form that applies both while the line is being scanned and after it has been completed is obtained by restricting the source integration to the physical exposure interval. For ,
Before the line starts, for . The contribution remains nonzero after the scan line is completed, as the deposited heat continues to diffuse through the domain.
The quantity in Eq. (23) represents the temperature contribution generated by scan line , and the analytical temperature field associated with a complete scan path is obtained by superposing the individual line contributions.
In addition to temperature, the spatial temperature gradient is required in the hybrid formulation. Equation (23) is evaluated in the local coordinates , whereas the hybrid formulation requires derivatives with respect to the global coordinates . From Eq. (14), the coordinate Jacobian is . Therefore, by the chain rule,
Thus, the derivatives of Eq. (23) are first evaluated in the local coordinate system and then transformed by to obtain the temperature-gradient components in the global coordinate system.
Conventional hybrid analytical–finite-element formulation
Hybrid analytical–finite-element formulations combine the ability of analytical moving-source solutions to capture steep temperature gradients near the laser with the geometric flexibility of finite-element analysis [21, 36, 38]. The total temperature field is decomposed as
where represents the laser-induced temperature field in a semi-infinite reference domain and is a finite-element correction that enforces the boundary conditions of the finite part. Figure 2 illustrates the decomposition and the different spatial scales represented by the two fields.

Figure 2. Schematic of the conventional hybrid decomposition. (a) Total temperature field in the finite part; the fine grid illustrates the spatial scale of the near-source response. (b) Analytical moving-source field in the semi-infinite reference domain. (c) Finite-element correction field computed on a coarser mesh in the finite part. The schematic was generated with assistance from Claude Opus 5 (Anthropic) and reviewed by the authors.
For a scan path containing multiple lines, the analytical field is obtained by superposition:
where is the set of scan lines whose exposure has begun by time . Let denote the laser-source contribution associated with line . Each line field satisfies in the semi-infinite reference domain. Since the heat operator is linear, the analytical field satisfies
Substituting Eq. (25) into the governing equation and using Eq. (27) gives
The correction field is solved in the finite domain . Its boundary and initial conditions follow from Eq. (25) and the prescribed conditions for . On ,
On ,
Initially,
Thus, the analytical field represents the laser-source response, while the finite-element field provides the correction required by the boundary and initial conditions of the finite part.
Because the analytical field captures the steep temperature gradients near the moving laser, the finite-element mesh needs to resolve only the smoother correction field. This permits a coarser mesh than a fully numerical model that directly resolves the laser-generated field, reducing the number of finite-element unknowns. For explicit time integration, a coarser mesh also relaxes the diffusion stability restriction, which typically scales as , where is a characteristic element size [41]. For implicit schemes without this stability restriction, the smoother correction can still permit larger time steps, provided its evolution and the changing boundary conditions are adequately resolved.
However, these computational advantages do not eliminate the growing cost of analytical evaluation. Equation (26) includes every scan line whose exposure has begun. After a line is completed, its analytical field continues to evolve as the deposited heat diffuses. As long as this contribution remains in the analytical representation, it must be reevaluated at subsequent time steps wherever or its gradient is required, including at points far from the completed line. The analytical workload therefore grows with the accumulated scan history. The localized formulation developed next restricts these evaluations in space and time and derives the correction terms needed to preserve the thermal effect of the removed analytical contributions.
Proposed Formulation
The proposed formulation restricts each analytical scan-line contribution in space and time. This subsection first defines the localization functions and then derives the corresponding correction terms for the finite-element field.
Spatial and temporal localization functions
The conventional decomposition in Eq. (25) is generalized as
where the localization function associated with scan line is written as
Here, controls the spatial extent over which the contribution of scan line is explicitly represented, whereas controls its temporal retention.
The localization functions are chosen to vary smoothly between their fully active and inactive states. For the spatial localization associated with scan line , let denote the shortest distance from to the corresponding finite scan segment. For a segment with start point and end point ,
The spatial localization function is defined as
where is the radius of the fully analytical core and is the width of the spatial transition region.
The temporal localization is defined relative to the completion time of each scan-line contribution. Let denote the completion time of line , the subsequent retention time, and the duration of the transfer to the finite-element field. Then
In Eqs. (35) and (36), is a smooth, monotonically decreasing transition function with , , and . The spatial regions and temporal stages of localization are shown schematically in Figure 3.
Figure 3. Spatial and temporal localization of an analytical scan-line contribution. (a) The spatial localization function satisfies within the analytical core of radius , decreases smoothly over the transition width , and vanishes outside the localized region surrounding the scan segment. (b) The temporal localization function remains equal to one during the scan and for the retention interval following the line completion time , and then decreases smoothly to zero over the transfer duration . The schematic was generated with assistance from Claude Opus 5 (Anthropic) and reviewed by the authors.
The localized formulation is controlled by four parameters: the analytical-core radius , the spatial transition width , the post-completion retention time , and the temporal transfer duration . During active scanning and throughout the retention interval, . Spatially, the analytical contribution is fully retained for , smoothly reduced for , and omitted from the explicit analytical representation outside this region. The conventional analytical field is recovered when for every scan-line contribution.
Localization residual and finite-element correction
To determine the governing equation for the correction field, Eq. (32) is substituted into Eq. (3), giving
Because and , the time derivative of one localized analytical contribution is
Furthermore,
Therefore,
Since , this becomes
Substituting this expression into Eq. (37) gives
When the laser is not scanning line , , so the source-localization term is zero. During exposure, . The term is then also zero inside the spatial core, where . The only remaining case is active exposure at points where . The beam center lies on the scan segment, so these points are at least from the beam center. From (8), the laser power density there is at most times its value at the beam center. For the parameter values used in this work, µm. The selected core radius mm gives a factor of . Even the smallest core radius examined, mm, gives only . Thus, where the source-localization term is not exactly zero, the Gaussian source it contains is extremely small. Dropping this term gives
The right-hand side of (43) represents the correction introduced by localization of the analytical field. For each scan-line contribution, it can be separated into temporal and spatial components,
where
and
The complete localization residual is
The temporal term accounts for the gradual removal of an analytical contribution from the explicit analytical representation. The spatial term accounts for the transition between regions in which that analytical contribution is fully represented and regions in which its effect is represented through the finite-element correction field.
The boundary conditions for follow from the decomposition in (32). On the Dirichlet boundary,
On the Neumann boundary,
Together, the localization residual and the modified boundary conditions allow the finite-element field to account for contributions removed from the explicit analytical representation.
The analytical evaluation can therefore remain concentrated near recent scan lines, limiting its spatial workload and the amount of scan history retained as scanning progresses.
To obtain the weak form, define the test space
Multiplying Eq. (43) by , integrating over , and integrating the diffusion term by parts gives
The Laplacian of the spatial localization function in Eq. (46) is also integrated by parts. Using gives
The boundary integral in Eq. (51) cancels the term containing when the Neumann condition in Eq. (49) is substituted into Eq. (50). The resulting weak form, valid for every , is
Thus, the weak form accounts for the spatial Laplacian through first derivatives and does not require its explicit evaluation.
The correction field is approximated as , where are the finite-element shape functions. Taking , with on , gives
where contains the nodal values of the correction field, and
The Dirichlet condition is imposed using Eq. (48). Algorithm 1 summarizes the resulting time-stepping procedure.
Algorithm 1 Localized hybrid thermal simulation
Require: Scan path, finite-element mesh, and localization parameters , , ,
Initialize from the prescribed initial temperature.
for each time step do
Select lines that have started and satisfy .
Initialize the analytical field and load vectors.
for each selected line do
Compute and .
Evaluate , , , and where needed.
Add to the analytical field.
Assemble the line’s contributions to and using Eq. (54).
end for
Include the prescribed flux contribution in using Eq. (54).
Form the finite-element system in Eq. (53) and impose Eq. (48).
Advance and reconstruct using Eq. (32).
end for
ChatGPT 6 (OpenAI) was used to assist in developing the Python simulation code. The implementation was tested and validated by the authors.
Results and discussion
The proposed localized hybrid formulation is evaluated on a fin geometry (see Figure 1) through four studies. First, the conventional and localized hybrid solutions are compared against an independent full finite-element solution obtained using Abaqus for a single-line benchmark. Next, a comparison with uncompensated temporal truncation examines the importance of accounting for removed analytical contributions. The effect of localization parameters on accuracy and computational cost is then examined. Finally, the analytical history and computational savings are analyzed for a 500-line case. The last three studies use the conventional hybrid formulation as the reference.
The fin geometry (see Figure 1) is assigned homogeneous and isotropic material properties with constant mass density, thermal conductivity, and specific heat capacity. The initial temperature is set to 32 °C throughout the domain, and the bottom surface is maintained at this temperature during the simulation. The remaining external surfaces are treated as adiabatic. The laser power, absorptivity, beam size, and scan speed are also kept fixed throughout the numerical studies. Their values are listed in Table 1.
| Category | Parameter | Value |
| Material | Mass density | 8440 kg m |
| Material | Thermal conductivity | 11.4 W m K |
| Material | Specific heat capacity | 443 J kg K |
| Laser and scan | Laser power | 300 W |
| Laser and scan | Absorptivity | 1.0 |
| Laser and scan | Gaussian source width | 50 m |
| Laser and scan | Scan speed | 2 m s |
Table 1. Physical parameters common to all numerical studies.
All simulations were performed on an Intel Core i9-14900KF CPU with 32 GB of RAM. The difference between discrete temperature fields is reported using the relative 2-norm difference,
where and are the discrete temperature vectors associated with the two solutions being compared, and is the reference solution. The choice of solution and reference solution is specified for each comparison below.
Numerical validation against full finite-element analysis
To independently assess the temperature predictions of the hybrid formulations, a single-scan-line problem is compared with a full finite-element solution obtained using Abaqus. The setup is illustrated in Figure 4. A single scan line is applied to the upper surface of the fin. The laser travels along a 29 mm line at 2 m s, corresponding to a scanning duration of 0.0145 s, followed by 0.2 s of cooling. The hybrid models represent the laser through the Eagar–Tsai solution, while Abaqus applies the corresponding moving Gaussian surface flux directly.

Figure 4. Single-line problem used for validation of the hybrid formulations against the full finite-element solution. (a) Validation geometry and scan path. (b) Representative Abaqus temperature field during scanning, displayed over 30–2000 °C, with higher temperatures shown in the upper contour color. Panel (a) was generated with assistance from Claude Opus 5 (Anthropic) and reviewed by the authors.
The Abaqus model uses linear tetrahedral heat-transfer elements with a prescribed size of 50 m near the upper surface, where the Gaussian heat flux is applied, and 250 m away from the surface. The resulting model contains 1,373,437 nodes and 7,171,717 elements. Fixed time increments of s during scanning and s during cooling are used. The hybrid simulations use a coarser linear tetrahedral mesh with prescribed element sizes of near the upper surface and away from it. This mesh contains 49,641 nodes and 238,833 elements, and the hybrid time-step size is .
Spatial and temporal refinement checks were performed for the Abaqus model. Refining the prescribed element sizes from to changed the temperature field by 0.605% in the relative 2-norm, with a root-mean-square error (RMSE) of 0.198 °C. On the mesh, halving the scanning and cooling time increments to the values stated above changed the field by 0.0463%, with an RMSE of 0.015 °C. These results show small differences under the tested refinements, with greater sensitivity to spatial than temporal discretization. The solution is therefore used as the finest tested Abaqus numerical reference.
Since the Abaqus and hybrid models use different finite-element meshes, their temperature fields are compared at 14,000 common spatial sampling points and 70 common saved times. The Abaqus temperature is obtained by finite-element interpolation at the sampling points. For the hybrid solutions, the analytical temperature is evaluated directly at the same points and combined with the interpolated finite-element correction. In addition to the relative 2-norm difference defined in Eq. (55), the root-mean-square error (RMSE) is reported over the same temperature samples.
As shown in Table 2, the conventional and localized hybrid formulations show close agreement with the full finite-element solution. Their relative 2-norm differences are 0.330% and 0.332%, respectively, with RMSE values of approximately 0.108 °C.
| Comparison | Relative 2-norm difference (%) | RMSE (°C) |
| Conventional hybrid vs. Abaqus | 0.330 | 0.108 |
| Localized hybrid vs. Abaqus | 0.332 | 0.108 |
| Localized vs. conventional hybrid | 0.0258 | 0.008 |
Table 2. Temperature-field comparison for the single-line validation problem.
The difference between the localized and conventional hybrid solutions is considerably smaller. The relative 2-norm difference is 0.0258%, with an RMSE of 0.008 . These results indicate that localization introduces only a small additional difference relative to the conventional hybrid formulation, while both hybrid formulations remain in close agreement with the independent full finite-element solution.
The Abaqus simulation required approximately 88,848 s (24.7 h) of wall-clock time, compared with 1576.2 s for the conventional hybrid simulation and 1281.6 s for the localized hybrid simulation. The full finite-element simulation required substantially more computational time. This reflects the cost of directly resolving the localized moving heat source with the finer spatial and temporal discretization required by the full finite-element model.
Comparison with temporal truncation
This study and the studies in Sections 4.3 and 4.4 use the parallel scan-line arrangement and finite-element mesh shown in Figure 5, with the numerical settings listed in Table 3. All three studies use 500 scan lines and include 1 s of post-scan cooling. Each line extends in the -direction from mm to mm, leaving 1 mm margins at both ends. The lines are uniformly spaced in the -direction between mm and mm, leaving 0.5 mm margins at the corresponding edges. Together, they span a mm mm region of the mm mm upper surface.

Figure 5. Fin geometry, finite-element mesh, and scan arrangement for the multiline studies. (a) Linear tetrahedral mesh with finer resolution near the upper surface and coarser resolution toward the bottom. (b) Schematic scan-line arrangement, scan direction, and thermal boundary conditions. Panel (b) was generated with assistance from Claude Opus 5 (Anthropic) and reviewed by the authors.
| Category | Parameter | Value |
| Scan path | Line length | 8 mm |
| Hatch spacing | 58 µm | |
| Delay between lines | 0.5 ms | |
| Finite-element model | Prescribed element-size range | 200–1000 µm |
| Number of nodes | 42470 | |
| Number of tetrahedral elements | 202094 | |
| Time integration | Time-step size | 0.5 ms |
Table 3. Numerical settings shared by the multiline hybrid studies.
For comparisons between the localized and conventional hybrid solutions, the relative 2-norm difference in Eq. (55) is evaluated using nodal temperatures over all saved times, with the conventional solution as the reference.
Computational performance is quantified using the analytical speedup
and the overall speedup
where and are the times spent evaluating analytical temperatures and gradients, and and are the total computational times of the conventional and localized simulations, respectively. Analytical work is also measured by the number of line–point evaluations, where each evaluation corresponds to the contribution of one scan line at one spatial point. The count is accumulated over all time steps and includes mesh nodes and the boundary and volume integration points required by the hybrid formulation.
The first multiline study compares the localized formulation with uncompensated temporal truncation for the 500-line problem. The comparison examines whether discarding older analytical contributions without accounting for their removal can maintain agreement with the conventional hybrid solution. The localized formulation uses a spatial core radius and transition width of 1 mm each, and a temporal retention age and transfer duration of 50 ms each. The sensitivity to these parameters and the reason for choosing these values are examined in § 4.3.
In the temporal-truncation formulation, each line is evaluated throughout the computational domain until a cutoff interval s after its completion. This interval matches the selected localized formulation, which uses 50 ms of retention followed by 50 ms of gradual transfer. The retained analytical field is
The finite-element correction is advanced using Eq. (28), with the boundary conditions in Eqs. (29) and (30) evaluated using . No localization residual or compensating transfer is applied when a line is removed. No spatial localization is used. Thus, both methods use the same removal deadline, but localization accounts for the gradual reduction of each analytical contribution.
A completed scan line continues to affect the temperature field as its heat diffuses through the domain. Deleting its analytical contribution without transferring the removed field to the finite-element correction therefore loses part of the thermal history. Updating the boundary correction does not replace the missing contribution throughout the interior of the domain. Repeated deletions can therefore accumulate into a substantial discrepancy from the conventional solution. In the localized formulation, the temporal residual in Eq. (45) accounts for the decreasing analytical weight, together with the spatial residual and boundary conditions derived in § 3.
Figure 6 shows the relative -norm difference at each saved time. Before the first deletion at s, temporal truncation agrees with the conventional solution to numerical precision. The difference then increases as scan lines are removed, reaching approximately 47.2% at s and 54.9% at scan completion, s. The localized solution remains close to the conventional solution throughout the simulation.

Figure 6. Relative nodal 2-norm difference from the conventional hybrid solution at each saved time for the 500-line case. The vertical lines indicate the first truncation event and scan completion.
This demonstrates the importance of compensating for removed analytical contributions when limiting the retained scan history.
Effect of localization parameters
This study examines how the spatial and temporal localization parameters in Algorithm 1 affect accuracy and computational cost. The localization parameters are varied while the scan path, material properties, mesh, and numerical integration settings are held fixed.
A common scale factor is applied to the four localization parameters. At , the spatial core radius and transition width are both 1 mm, while the temporal retention age and transfer duration are both 50 ms. Each parameter is multiplied by , with
Increasing enlarges the spatial region and extends the time interval over which analytical contributions are evaluated. Larger spatial regions allow each contribution to diffuse further before it is represented through the finite-element correction. Longer temporal transition intervals also reduce the rate at which its analytical weight changes, as reflected in the temporal residual in Eq. (45). For a fixed mesh and time step, these changes are expected to improve agreement with the conventional solution, while requiring more analytical evaluation. All localized results are compared with the same conventional hybrid reference.
Figure 7 shows how accuracy and analytical speedup change with the localization scale. The relative nodal 2-norm difference, evaluated over all saved times, decreases monotonically from 0.226% at to 0.0797% at . Over the same range, the number of analytical line–point evaluations increases from to , and the analytical speedup decreases from 94.8 to 8.37. These results are consistent with the expected tradeoff: larger localization scales give closer agreement at the cost of more analytical evaluation. Substantial analytical savings remain throughout the tested range.

Figure 7. Influence of the localization scale in the 500-line case: (a) relative nodal 2-norm difference from the conventional hybrid solution over all saved times and (b) analytical speedup relative to the conventional formulation.
The setting is used in the temporal-truncation comparison in § 4.2 and the computational-savings study in § 4.4. Although provides a much larger analytical speedup, its overall speedup (when also accounting for FEA) increases only from 3.92 to 4.59. Thus, provides closer agreement at a modest additional computational cost.
Analytical history and computational savings
The same 500-line case used in § 4.2 is examined here to show how localization limits the analytical history and reduces computational cost. The localization parameters are fixed at the values discussed in § 4.3.
Figure 8 compares the number of scan lines that have begun with the number that remain explicitly included in the localized analytical field. In the conventional hybrid formulation, every started line remains in the analytical superposition throughout the simulation. The number of analytical contributions therefore increases to 500 by the end of scanning. In the localized formulation, completed lines are gradually removed from the analytical field, and the number of retained analytical lines never exceeds 24.

Figure 8. Number of started scan lines and explicitly retained analytical scan lines for the localized formulation in the 500-line case.
Table 4 compares the time spent in the main computational components of the two formulations. Localization reduces the analytical evaluation time from 13059.6 s to 512.1 s, giving an analytical speedup of 25.5. The finite-element solution times remain similar, at 3308.9 s and 3209.0 s, respectively.
| Component | Conventional (s) | Localized (s) |
| Analytical evaluation | 13059.6 | 512.1 |
| Finite-element solution | 3308.9 | 3209.0 |
| Localization and assembly | 4.8 | 280.0 |
Table 4. Time spent in the main computational components for the 500-line case. Localization and assembly includes localization operations, localization-residual assembly, and boundary assembly.
Localization introduces additional work to construct the localization functions and assemble the residual terms. The combined localization and assembly time is 280.0 s, compared with 4.8 s for boundary assembly in the conventional formulation. The total simulation time decreases from 16375.5 s to 4176.4 s, giving an overall speedup of 3.92. Over all saved times and mesh nodes, the relative nodal -norm difference is 0.139%.
Conclusions
This work developed a space–time localized hybrid analytical–finite-element formulation for part-scale thermal simulation of laser powder bed fusion. The formulation addresses the growing analytical workload of conventional hybrid methods as scan-line contributions accumulate. Spatial and temporal localization restrict where and how long each contribution is evaluated analytically. Residual terms and boundary conditions derived from the heat equation allow the finite-element correction to account for the thermal effect of contributions removed from the analytical field.
The proposed formulation maintained close agreement with an independent full finite-element solution for the single-line benchmark. Both the conventional and localized hybrid formulations reproduced the reference temperature field closely, supporting the use of the localized formulation in the subsequent multiline comparisons.
Limiting analytical history requires accounting for the thermal effect of removed contributions. Uncompensated temporal truncation produced a large discrepancy from the conventional hybrid solution, whereas localization maintained close agreement. These results demonstrate the importance of the residual correction in reducing the explicit analytical history while retaining its thermal effect.
The extent of localization controls the tradeoff between agreement and computational cost. Larger spatial regions and longer temporal intervals improved agreement with the conventional solution, while smaller settings provided greater analytical savings. Substantial savings were obtained throughout the tested parameter range, allowing the parameters to be chosen according to the desired balance between agreement and computational effort.
With the selected parameters, the retained analytical history never exceeded 24 line contributions in the 500-line case, while the conventional formulation retained all 500 contributions by the end of scanning. This reduction gave an analytical speedup of 25.5 and an overall speedup of 3.92, with a relative nodal 2-norm difference of 0.139%. The additional localization and assembly work was small compared with the analytical evaluation time saved. The finite-element solution time remained nearly unchanged and became the dominant computational cost, limiting the overall speedup. This provides a basis for applying the proposed method to larger part models (with a much larger number of scan lines), with greater analytical savings expected as the conventional history grows.
Possible future directions include multilayer builds with evolving geometry, more varied scan paths, and adaptive selection of the localization parameters. Adjusting the spatial regions and transfer intervals to the thermal field, mesh, and time step could provide further savings while controlling the difference from the conventional solution. Extending the formulation to temperature-dependent material properties and phase change would broaden its physical scope, while experimental validation would help assess its predictions under practical LPBF conditions.
Acknowledgments
The authors acknowledge support from the US Office of Naval Research under Grant N000142412391.
Conflicts of Interest
The authors declare no conflicts of interest.
Data Availability Statement
The code developed in this work can be made available upon request.
Use of AI Tools
The authors used generative AI assistants, specifically, ChatGPT 6 (OpenAI) and Claude Opus 5 (Anthropic), during the preparation of this manuscript. The tools were used for code development, drafting of text, and generating schematic illustrations. All content was reviewed, verified, and edited by the authors, who take full responsibility for the accuracy and integrity of the work.
References
- [1]A. Busachi, J. Er koyuncu, P. Colegrove, F. Martina, C. Watts, R. Drake, A review of additive manufacturing technology and cost estimation techniques for the defence sector, CIRP Journal of Manufacturing Science and Technology 19 (2017) 117–128.DOI
- [2]L. J. Kumar, C. Krishnadas Nair, Current trends of additive manufacturing in the aerospace industry, in: Advances in 3D printing & additive manufacturing technologies, Springer, 2016, pp. 39–54.DOI
- [3]T. Pereira, J. V. Kennedy, J. Potgieter, A comparison of traditional manufacturing vs additive manufacturing, the best method for the job, Procedia manufacturing 30 (2019) 11–18.DOI
- [4]P. Promoppatum, S.-C. Yao, P. C. Pistorius, A. D. Rollett, A comprehensive comparison of the analytical and numerical prediction of the thermal history and solidification microstructure of inconel 718 products made by laser powder-bed fusion, Engineering 3 (5) (2017) 685–694.DOI
- [5]S. C. Subedi, A. Shahba, M. Thevamaran, D. J. Thoma, K. Suresh, Towards the optimal design of support structures for laser powder bed fusion-based metal additive manufacturing via thermal equivalent static loads, Additive Manufacturing 57 (2022) 102956.DOI
- [6]S. C. Subedi, D. J. Thoma, K. Suresh, Deformation constrained support-structure optimization for laser powder bed fusion, Additive Manufacturing 89 (2024) 104294. doi:10.1016/j.addma.2024.104294.DOI
- [7]B. Soundararajan, D. Sofa, D. Barletta, M. Poletto, Review on modeling techniques for powder bed fusion processes based on physical principles, Additive Manufacturing 47 (2021) 102336.DOI
- [8]H. Gong, K. Rafi, H. Gu, T. Starr, B. Stucker, Analysis of defect generation in ti–6al–4v parts made using powder bed fusion additive manufacturing processes, Additive Manufacturing 1 (2014) 87–98.DOI
- [9]P. Scheel, E. Mazza, E. Hosseini, Adaptive local–global multiscale approach for thermal simulation of the selective laser melting process, Additive Manufacturing 36 (2020) 101518.DOI
- [10]X. Ji, Y. Wang, S. Y. Liang, Analytical modeling of temperature evolution in laser powder bed fusion considering the size and shape of the build part, Journal of Materials Processing Technology 301 (2022) 117452.DOI
- [11]Y. Bresson, A. Tongne, M. Baili, L. Arnaud, Global-to-local simulation of the thermal history in the laser powder bed fusion process based on a multiscale finite element approach, The International Journal of Advanced Manufacturing Technology 127 (9) (2023) 4727–4744.
- [12]E. De Leon, A. Riensche, B. D. Bevans, C. Billings, Z. Siddique, Y. Liu, A review of modeling, simulation, and process qualification of additively manufactured metal components via the laser powder bed fusion method, Journal of Manufacturing and Materials Processing 9 (1) (2025) 22.DOI
- [13]W. King, A. T. Anderson, R. M. Ferencz, N. E. Hodge, C. Kamath, S. A. Khairallah, Overview of modelling and simulation of metal powder bed fusion process at lawrence livermore national laboratory, Materials Science and Technology 31 (8) (2015) 957–968.DOI
- [14]A. Love, O. A. Valdez Pastrana, S. Behseresht, Y. H. Park, Advancing metal additive manufacturing: A review of numerical methods in ded, waam, and pbf, Metrology 5 (2) (2025) 30.DOI
- [15]M. Bayat, W. Dong, J. Thorborg, A. C. To, J. H. Hattel, A review of multi-scale and multi-physics simulations of metal additive manufacturing processes with focus on modeling strategies, Additive Manufacturing 47 (2021) 102278.DOI
- [16]M. Xia, D. Gu, G. Yu, D. Dai, H. Chen, Q. Shi, Porosity evolution and its thermo-dynamic mechanism of randomly packed powder-bed during selective laser melting of inconel 718 alloy, International Journal of Machine Tools and Manufacture 116 (2017) 96–106.
- [17]C. Li, C. Fu, Y. Guo, F. Fang, A multiscale modeling approach for fast prediction of part distortion in selective laser melting, Journal of materials processing technology 229 (2016) 703–712.DOI
- [18]B. C. Prabhune, K. Suresh, A fast matrix-free elasto-plastic solver for predicting residual stresses in additive manufacturing, Computer-Aided Design 123 (2020) 102829.DOI
- [19]T. Eagar, N. Tsai, et al., Temperature fields produced by traveling distributed heat sources, Welding journal 62 (12) (1983) 346–355.
- [20]Y. Yang, M. Knol, F. Van Keulen, C. Ayas, A semi-analytical thermal modelling approach for selective laser melting, Additive manufacturing 21 (2018) 284–297.DOI
- [21]T. Moran, D. Warner, N. Phan, Scan-by-scan part-scale thermal modelling for defect prediction in metal additive manufacturing, Additive Manufacturing 37 (2021) 101667.DOI
- [22]Z. Zhang, Y. Wang, P. Ge, T. Wu, A review on modelling and simulation of laser additive manufacturing: Heat transfer, microstructure evolutions and mechanical properties, Coatings 12 (9) (2022) 1277.
- [23]A. Olleak, Z. Xi, Part-scale finite element modeling of the selective laser melting process with layer-wise adaptive remeshing for thermal history and porosity prediction, Journal of Manufacturing Science and Engineering 142 (12) (2020) 121006.DOI
- [24]F. Dugast, P. Apostolou, A. Fernandez, W. Dong, Q. Chen, S. Strayer, R. Wicker, A. C. To, Part-scale thermal process modeling for laser powder bed fusion with matrix-free method and gpu computing, Additive Manufacturing 37 (2021) 101732.DOI
- [25]A. Olleak, F. Dugast, P. Bharadwaj, S. Strayer, S. Hinnebusch, S. N arra, A. C. To, Enabling part-scale scanwise process simulation for predicting melt pool variation in lpbf by combining gpu-based matrix-free fem and adaptive remeshing, Additive Manufacturing Letters 3 (2022) 100051.DOI
- [26]M. Liu, L. N. Chiu, D. Liu, A. Huang, C. Davies, X. Wu, W. Yan, A macroscale thermal simulation strategy with track-scale resolution for laser powder bed fusion, Materials & Design 226 (2023) 111622.DOI
- [27]J. P. Leonor, G. J. Wagner, Go-melt: Gpu-optimized multilevel execution of lpbf thermal simulations, Computer Methods in Applied Mechanics and Engineering 426 (2024) 116977.DOI
- [28]S. Elahi, J. Leonor, R. Wu, G. Wagner, Efficient part-scale thermal modeling of laser powder bed fusion via a multilevel finite element framework, Additive Manufacturing (2025) 104897.
- [29]D. Rosenthal, The theory of moving sources of heat and its application to metal treatments, Transactions of the American Society of Mechanical Engineers 68 (8) (1946) 849–865.DOI
- [30]M. B. Nasiri, N. Enzinger, Powerful analytical solution to heat flow problem in welding, International Journal of Thermal Sciences 135 (2019) 601–612.
- [31]P. Zagade, B. Gautham, A. De, T. DebRoy, Analytical estimation of fusion zone dimensions and cooling rates in part scale laser powder bed fusion, Additive Manufacturing 46 (2021) 102222.DOI
- [32]P. Honarmandi, R. Seede, L. Xue, D. Shoukr, P. Morcos, B. Zhang, C. Zhang, A. Elwany, I. Karaman, R. Arroyave, A rigorous test and improvement of the eagar-tsai model for melt pool characteristics in laser powder bed fusion additive manufacturing, Additive Manufacturing 47 (2021) 102300.DOI
- [33]Y. Wang, X. Ji, S. Y. Liang, Analytical modeling of temperature distribution in laser powder bed fusion with different scan strategies, Optics & Laser Technology 157 (2023) 108708. doi:10.1016/j.optlastec.2022.108708.DOI
- [34]P. R. Zagade, B. Gautham, A. De, T. DebRoy, Analytical modelling of scanning strategy effect on temperature field and melt track dimensions in laser powder bed fusion, Additive Manufacturing 82 (2024) 104046.DOI
- [35]J. C. Steuben, A. J. Birnbaum, J. G. Michopoulos, A. P. Iliopoulos, Enriched analytical solutions for additive manufacturing modeling and simulation, Additive Manufacturing 25 (2019) 437–447.
- [36]Y. Yang, F. van Keulen, C. Ayas, A computationally efficient thermal model for selective laser melting, Additive Manufacturing 31 (2020) 100955.DOI
- [37]T. Moran, P. Li, D. Warner, N. Phan, Utility of superposition-based finite element approach for part-scale thermal simulation in additive manufacturing, Additive Manufacturing 21 (2018) 215–219.DOI
- [38]Y. Yang, Y. Ji, M. Möller, C. Ayas, Computational efficient process simulation of geometrically complex parts in metal additive manufacturing, International Journal of Heat and Mass Transfer 248 (2025) 127059.DOI
- [39]Y. Yang, Y. Ji, M. Möller, C. Ayas, Efficient thermal simulation in metal additive manufacturing via semi-analytical isogeometric analysis, Computer Methods in Applied Mechanics and Engineering 457 (2026) 118992.arxiv.org/abs/2511.20432
- [40]A. J. Wolfer, J. Aires, K. Wheeler, J.-P. Delplanque, A. Rubenchik, A. Anderson, S. Khairallah, Fast solution strategy for transient heat conduction for arbitrary scan paths in additive manufacturing, Additive Manufacturing 30 (2019) 100898.DOI
- [41]G. Strang, Computational science and engineering, Vol. 791, Wellesley-Cambridge Press Wellesley, MA, 2007.
