INL/PRT-25-88167 Multiscale And Machine Advanced Materials and Manufacturing Technologies Learning Modeling For Process-informed Microstructure Prediction In Additively Manufactured Materials Using MALAMUTE SEPTEMBER 2025 Sudipta Biswas, Somayajulu L. N. Dhulipala, Peter German, and Michael D. McMurtrey Idaho National Laboratory Purna Vindhya Kota John Hopkins University INL/PRT-25-88167 Multiscale And Machine Learning Modeling For Process-informed Microstructure Prediction In Additively Manufactured Materials Using MALAMUTE Sudipta Biswas, Somayajulu L. N. Dhulipala, Peter German, and Michael D. McMurtrey Idaho National Laboratory Purna Vindhya Kota John Hopkins University September 2025 Idaho National Laboratory Idaho Falls, Idaho 83415 http://www.inl.gov CONTENTS ABSTRACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii ACKNOWLEDGMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 2. High-fidelity Process Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 2.1. Governing equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 2.2. Pulsed heat source . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 2.3. Melt-pool dynamics prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 2.4. Comparison of the Pulsed vs. Continuous Energy Deposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 2.4.1. 2D Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 2.4.2. 3D Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 3. Uncertainty-Aware Machine Learning Model for Melt-Pool Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 3.1. Latent multi-output Gaussian process auto-regressive model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 3.2. CNN-based UNet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 3.3. Machine learning melt-pool predictions: UNet and L-MOGPAR. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 3.3.1. Temperature field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 3.3.2. Melt-pool morphology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 3.3.3. Uncertainty quantification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 4. Modeling Microstructural Evolution during the AM Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 4.1. Phase-field Model for Solute Segregation During Solidification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 4.1.1. Model formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 4.1.2. Phase-field model parameterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 4.2. Solidification Demonstration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 4.2.1. Model verification and parameter selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 4.2.2. Solute segregation during solidification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 4.3. Process to Microstructure Connections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 4.3.1. Solidification rate and morphology prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 4.3.2. Chemical segregation prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 5. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 viii REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 ix FIGURES Figure 1. Schematic of the multiscale modeling components for establishing PSPP correlations for AM. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 Figure 2. Beam position as function of time for pulsed and continuous energy deposition. . . . . . . . . . . . . 6 Figure 3. Simulation results for the process input (74.77 W , 1.0 m/s, 125 µm) sample at 303 µs. The velocity of the melt pool is visualized using unscaled arrows. The melt-pool deformation, which has a Gaussian distribution, is also visible at the leading edge of the laser beam.. . . . . 6 Figure 4. Melt-pool geometries for three representative input process parameters to show the influence of process parameters on the peak flow velocity of the melt pool. Results show that increasing the laser power results in a larger melt pool with a higher maximum melt-pool temperature. Conversely, increasing scanning speed reduces the melt-pool size because linear energy deposition is reduced. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 Figure 5. Melt pool geometry in the 2D model throughout the laser pass with a pulsed source (top) and a continuous source (bottom) of the same linear energy density. . . . . . . . . . . . . . . . . . . . . . . . . . 9 Figure 6. Meltpool geometry at the end of the simulation (t = 2 × 10−4 s) for continuous laser power delivery (bottom) and pulsed power deliver (top). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 Figure 7. Melt pool geometry in the 3D model throughout the laser pass with a pulsed source (top) and a continuous source (bottom) of the same linear energy density. . . . . . . . . . . . . . . . . . . . . . . . . . 10 Figure 8. Schematic of the UNet architecture. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 Figure 9. The temperature fields of a representative test data point of the MOOSE multiphysics model compared to the UNet and L-MOGPAR predictions at three time steps: (i) row 1 at t = 267.7 µs, (ii) row 2, at t = 444.4 µs, and (iii) row 3, at t = 800 µs. . . . . . . . . . . . . . . . . . . . . . 14 Figure 10. A comparison of the temperature profiles of the multiphysics simulation, L-MOGPAR, and UNet: (a) temperature history of the center point on the top surface, and the (b) temperature profile along the center line on the top surface at t = 444.4 µs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 Figure 11. Pixel-wise relative error maps corresponding to the predictions in Figure 9. . . . . . . . . . . . . . . . . . 16 Figure 12. Comparison of melt-pool boundaries predicted by the multiphysics simulation, L- MOGPAR, and UNet models for three time steps of a representative test point. Both machine learning models track the overall morphological evolution from initial melting through steady-state elongation to post-laser shrinkage. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 Figure 13. Quantitative comparison of melt-pool morphology evolution over time: the (a) melt-pool depth, (b) length, and (c) area of a representative test sample, and the (d) average melt-pool area evolution across all test samples.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 Figure 14. The spatially resolved standard deviation (first row), absolute error (second row), and mean predictions (third row) from L-MOGPAR at three representative time steps. The model predictions are less certain near the melt-pool boundaries and during peak heating. These regions coincide with a greater prediction error. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 Figure 15. Effect of microstructure on the propertie s and irradiation performance of the AM materials. Grain morphology, precipitation on dislocation cell walls influences the irradiation performance an creep properties. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 Figure 16. (a) Ternary phase diagram of the Fe-Cr-Ni system at 1600K. (b) Scheil calculation showing the solidification path and corresponding stable phases. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 Figure 17. Free energy calculation at for (a) solid at 1600 K and (b) liquid phase at 1740 K. For the free energy calculations, the equilibrium concentration of the species are considered same as the nominal concentration.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 Figure 18. Concentration variation with reduced temperature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 Figure 19. Fitted correlation for different parabolic coefficients: (a) ωeq s (b) µeq , and (c) µeq . . . . . . . . . . 24 Cr Ni x Figure 20. (a) Growth of a solid nuclei, and corresponding (b) Cr concentration and (c) Ni concentration distribution at simulation time t = 50. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 Figure 21. (a) Evolution of solid fraction over time. Variation in (b) Cr concentration and (c) Ni concentration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 Figure 22. Representative input locations around the melt pool for the three different time steps at which the phase-field simulations for solid growth were performed. These simulations were conducted for three input parameter sets and the three methods (i.e., multiphysics simulation, L-MOGPAR, and UNet) used to predict temperature fields. . . . . . . . . . . . . . . . . . . . . . 29 Figure 23. Dendritic structure formation for input 3 (154.16 µm, 1.17 m/s, 72.86 W) at spatial points (a) 3 (time 266.7 µs), (b) 3 (time 444.4 µs), and (c) 1 (time 800 µs). The definitions of the spatial points are presented in Figure 22. These were chosen from several combinations to be the representative spatial points and times for the results presented herein. . . . . . . . . . . . . . . . 30 Figure 24. Solid growth in the x- and y-direction for process inputs 2 (138.85 µm, 1.02 m/s, 68.27 W) and 3 (154.16 µm, 1.17 m/s, 72.86 W). These results are for a process modeling simulation time of 444.4 µs and spatial point 3 shown in Figure 22. Solid growth is compared with time predictions made using the process modeling inputs (i.e., temperature gradient and cooling rate) obtained using the multiphysics simulation, UNet, and L-MOGPAR models. The simulation time in this figure refers to the phase-field simulation time.. . . . . . . . . . . . . . . . . . 31 Figure 25. (a) Solid growth over time, (b) Interface location at simulation time t = 50, (c) Cr concentration, and (d) Ni concentration across the solid-liquid interface. Case I, Case II and Case III corresponds to process inputs 1 (138.97 μm, 1.41 m/s, 74.11 W), 2 (138.85 µm, 1.02 m/s, 68.27 W) and 3 (154.16 µm, 1.17 m/s, 72.86 W), respectively. These results are for the conditions obtained from a process modeling simulation time of 266.7 µs and spatial point 1 shown in Figure 22. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 xi TABLES Table 1. Values of input parameters in the training dataset. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 Table 2. Goodness-of-fit (R2 ) values of the temperature predictions and their spatial gradients. Generally, both models predict these quantities well, but UNet has an edge. . . . . . . . . . . . . . . . . . . 17 Table 3. Non-dimensionalized simulation parameters for the phase-field solidification simulation . . . . 29 Table 4. Maximum absolute percentage errors (lower is better) for solid growth in the x- and y-direction across the phase-field simulation times. The average and maximum error values across the three process modeling times and three spatial locations are presented for the three process modeling inputs: 1 (138.97 µm, 1.41 m/s, 74.11 W), 2 (138.85 µm, 1.02 m/s, 68.27 W), and 3 (154.16 µm, 1.17 m/s, 72.86 W). . . . . . . . . . . . . . . . . . . . . . 31 xii xiii ACRONYMS 2D two dimension AM additive manufacturing AMMT Advanced Materials and Manufacturing Technologies CALPHAD CALculation of PHAse Diagrams CNN convolutional neural network FEM finite element method INL Idaho National Laboratory LPBF laser powder bed fusion ML machine learning MOGPAR multi-output Gaussian process auto-regressive MOOSE Multiphysics Object-Oriented Simulation Environment PSPP process-structure-property-performance SS stainless steel STM Stochastic Tools Module xiv xv Multiscale and Machine Learning Modeling for Process-informed Microstructure Prediction in Additively Manufactured Materials using MALAMUTE 1. INTRODUCTION In recent years, additive manufacturing (AM) techniques have gained significant interest across government agencies and nuclear industries due to their ability to produce complex structures in a shorter time, facilitating faster fabrication. These techniques have the potential to change the paradigm of new material and component designs for nuclear applications, such as producing new alloys and graded materials, as well as innovative component designs for advanced nuclear reactors [1]. The Advanced Materials and Manufacturing Technologies (AMMT) program under the Department of Energy Office of Nuclear Energy aims to accelerate the development and qualification of additively manufactured materials for nuclear applications. Several work packages under the program focus on how the additively manufactured products perform under extreme thermal and irradiation conditions. However, an understanding of the development of the initial microstructure is crucial to evaluate the performance of the material and eventually qualify it for advanced nuclear reactor deployment. The key challenge that AM techniques face is controlling process conditions to obtain desired microstructures and properties for targeted applications. AM also faces the issue of reproducibility due to microstructural and property variation across, as well as within, builds. For example, depending on the build direction, the texture of the additively manufactured sample could be very heterogeneous, leading to extremely anisotropic strength and deformation behavior [2–4]. Compared to traditional manufacturing techniques, qualifying additively manufactured component for extreme environments becomes challenging due to such morphology variations from build to build and even within a single build [5]. This is especially true for the development of novel materials for nuclear applications. It is therefore of utmost importance to understand the microstructural evolution during the manufacturing process and optimize the process parameters to design materials with the desired microstructures and properties. AM is a complex process involving the melting of powder particles and the eventual solidification to form a complex grain structure. The rapid cooling rate causes unique changes in the microstructure and the layer-wise manufacturing causes anisotropy in the build direction, thus changing the mechanical properties with respect to wrought material. Hence, understanding the AM process and its impact on the microstructural evolution of the material necessitates development of a multiscale multiphysics models, models that capture and correlate the intricacies of the AM process and the corresponding microstructure formation. The interconnection between different physics-based models is shown in Figure 1. 1 Figure 1. Schematic of the multiscale modeling components for establishing PSPP correlations for AM. Various experimental and computational efforts have been devoted to establishing a process-structure- property (PSP) relation for various AM processes. For example, the effects of processing parameters on solidification textures were evaluated, and an attempt was made to control texture during the electron beam melting process [4]. However, adopting a trial-and-error approach to experimentally evaluate the process parameters required for texture control would be costly and time-consuming. Modeling and simulation provide a cost-effective way to understand and correlate microstructural behavior to processing conditions. The melting and subsequent solidification that occurs during the AM process is a complex phenomenon that requires multiscale multiphysics analysis. Several macroscopic studies have been done to capture the heat and fluid flow during various AM processes as well as part-scale component modeling [6]. However, the physics-based microstructural analysis of the material undergoing the process has been limited [7]. Additionally, the multiscale multiphysics models that model the AM process and microstructures themselves can be computationally prohibitive to rapidly explore the vast PSP space and iterate toward optimal material and component designs. As such, surrogate modeling and machine learning that can reduce the computational expense of these multiscale multiphysics models have become popular alternatives. The Multiphysics Object-Oriented Simulation Environment (MOOSE) software developed at Idaho National Laboratory (INL), specifically the MOOSE Application Library for Advanced Manufacturing UTilitiEs (MALAMUTE) software, provides an ideal platform for developing the multiphysics multiscale model to explore the intricacies of the microstructural evolution during the AM processes within a single framework. This work package aims to develop a MOOSE-based multiscale model that can couple different physical aspects of the AM process in a modular way. Here, we utilize and extend the existing capabilities of MOOSE and MALAMUTE to capture different aspects of the AM process. The physics-based process models demonstrates the solidification and associated chemical segregation in 316 stainless steel (SS), modeled as a simplified Fe-Cr-Ni system. Furthermore, a machine learning (ML) framework is developed using MOOSE- based Stochastic Tools Module (STM) to correlate the process parameters to the final microstructure through a surrogate model representing the high fidelity process simulations[8]. This work package provides answers 2 to several programmatic questions regarding the segregtaion behavior observed in additively manufactured materials. It also delivers a computationally efficient way to connect the process and part-scale models to the microstructural evolution model. This work uses a high-fidelity multiphysics process model as compared to most heat-equation-centric models to evaluate the thermal history variation with various process conditions. We developed a latent multi-output Gaussian-process autoregressive (L-MOGPAR) emulator for the coupled temperature and melt-pool dynamics that yields calibrated uncertainty over space and time. We then pushed these predicted thermal histories through a grand-potential-based phase-field model to quantify how process-level predictions influence the solidification rate at the microstructure level. To our knowledge, recent AM machine learning studies reported fast field predictors with limited UQ and rarely connected process-level machine learning predictions to microstructure evolution; we closed this gap with an end-to-end pipeline. This report summarizes the efforts within fiscal year 2025 to quantify chemical micro-segregation around the solidification fronts and to connect the microstructural model to realistic process conditions using a reduced order model. The key highlights of the work are as follows: • Multiphysics process model: A high-fidelity, arbitrary Lagrangian-Eulerian-based finite element method with free-surface Navier–Stokes, enthalpy-based phase change, radiation, and surface forces (recoil, capillarity), which enables an evaluation beyond heat-equation-only surrogates. • Probabilistic process emulator: A latent multi-output Gaussian process (GP) autoregressive model for joint temperature and melt-pool fields with calibrated uncertainty and a Markov temporal structure. • Phase-field based solidifcation model: A phase-field-based solidification model is developed for 316SS, capturing the solidification rate and associated chemical micro-segregation. • Process-to-microstructure link: The first end-to-end demonstration (to our knowledge) of propagating machine learning predictions to a grand-potential phase-field model, which quantifies their impact on solidification metrics. The rest of the report is organized as follows. In Section 2, we describe the high-fidelity process model predicting the melt pool dynamics with varying process conditions. In Section 4 we present the phase-field model for capturing the solidification and associated micro-segregation in 316 SS. Section ?? presents the machine learning models necessary for deriving the reduced order models for AM thermal conditions. Finally, in Section 2 we demonstrate how different capabilities of MALAMUTE can be integrated to establish the process to microstructure connections. 3 2. HIGH-FIDELITY PROCESS MODEL To provide data for the machine learning models, either a high-fidelity model or detailed measurements are essential for different process parameters. The process parameters significantly influence the temperature and shape of the melt pool and subsequently the microstructure of the product itself. This parametric problem can be challenging to describe, especially with limited access to measurement data. For this reason, a comprehensive finite element numerical model has been developed in this work, incorporating multiphysics phenomena that describe the heat transfer and flow in melt pools, together with the melt pool deformation. This comprehensive model has been successfully applied to single pass laser melting simulations with different model parameters, ensuring state estimation with sufficient accuracy. 2.1. Governing Equations The melt pool dynamics simulations in this work use the open source capabilities available in the Navier- Stokes module of MOOSE [9, 10] which rely on the physics formulation specified in [11]. The Navier-Stokes module provides algorithms to resolve melt pool dynamics using the the Navier-Stokes equations employing a stabilized continuous finite element method (FEM). In this work, we assume that the working fluids are incompressible. This leads to the following formulation of the momentum conservation equation: ∂ρ~u + ∇ · (ρ~u ⊗ ~u) = ∇ · −pI + µ ∇~u + ∇~uT ,   (1) ∂t where ~u and p~ denote the velocity and pressure fields, respectively. Furthermore, ρ and µ are the density and the dynamic viscosity of the fluid. Two surface effects are considered that act as source and sink terms in the momentum equation: (i) the force exerted by the vapor recoil and (ii) the surface tension. The surface force of the vapor recoil can be expressed as: f~vr = pvr ~n, (2) where pvr is a temperature-dependent recoil pressure and ~n is the surface normal. The expression for recoil pressure is taken from [12]:   Lv M (T −Tv ) pvr = 0.54p0 e Rg T T v , (3) where p0 is the atmospheric pressure, Lv is the latent heat of vaporization (Lv = 7.456 × 106 J/kg in this case), M is the molar mass (M = 5.6 × 10−2 kg/mol in this case), Tv is the vaporization temperature (Tv = 3080 K win this case) and Rg is the ideal gas constant. Similarly, the surface tension can be expressed as [11]: f~st = −2σH~n, (4) where σ is a temperature-dependent surface tension and H is the local mean surface curvature. The more complex terms from the surface tension expression in [11] have been ommited in this work. The mass and energy conservation equations can be expressed as: ∇ · ~u = 0 . (5) ∂ρh + ∇ · (~uρh) = ∇ · (k∇T ), (6) ∂t where T is the temperature field, h = f (T ) is the specific enthalpy and k is the thermal conductivity. During laser melting processes the heat loss through radiation can be considerable. This accounted for using a surface heat sink that describes the following flux: qrad = σb (T 4 − Tenv 4 ), (7) 4 where σb is the Stefan-Boltzmann constant,  is the absorptivity, while Tenv is the environmental temperature. The laser heat source is also applied a surface source with a Gaussian source profile: 2ηP − 2r2 qlaser = e R , (8) πR2 where r is the distance from the moving laser centroid, R is the effective beam radius, while ηP denotes the effective power which is the product of the total laser power P and the effective absorption rate η. The system of coupled nonlinear partial differential equations are solved using the FEM and Newton’s method. The surface deformation (due to the surface forces) is accounted for by a Arbitrary Lagrangian Eulerian (ALE) scheme following the process in [11]. In this method, the mesh is displaced at the boundary with additional smoothing steps to reposition the mesh nodes. The node structure in the mesh does not change, so it supplies a constant topology for the machine learning models to train and predict. 2.2. Pulsed Heat Source In laser powder bed fusion (LPBF) and laser welding, process settings typically use either continuous- wave (CW) or pulsed-wave (PW) emission [13]. CW delivery favors stable, high-throughput energy input that often promotes higher track stability at the cost of larger cumulative heat input and a higher propensity for keyhole/overheating defects. On the other hand, PW delivery lowers average heat input and can sharpen feature fidelity or thermal control in thin sections. The choice of power delivery mode influences melt pool geometry, stability, and flow and therefore the local thermal histories that set solidification rates, grain morphology, and microstructural features. These ultimately influence macroscopic properties such as strength, fatigue, and anisotropy. Given prior comparative studies showing that power delivery mode measurably alters melt pool dimensions and melting efficiency, we developed numerical models within MALAMUTE to investigate the effect of power delivery mode on melt pool characteristics. As already discussed, we model the laser heat source with a Gaussian distribution (see Eq. (8)), where r = k~r − ~r0 (t)k measures the distance from the instantaneous beam centroid ~r0 (t). The distinction between continuous and pulsed scanning can be described through the motion of ~r0 (t). For continuous scanning, the beam center traverses the path smoothly, ~r0 (t) = ~r0start + ~v t, (9) moving linearly at velocity ~v . For pulsed operation, ~r0 (t) advances in a piecewise-constant, step-wise manner between discrete spatial locations, ~r0 (t) = ~r0,i , t ∈ [ti , ti+1 ), (10) remaining fixed during each interval and jumping at pulse transitions. The jump can be described by parameters such as the exposure time and the point distance [14]. This difference is depicted in Figure 2. An important parameter is the linear energy density which is kept constant in this work when we compare the two power delivery systems. In case of the pulsed source the linear energy density can be computed as: P ·T , (11) D where P is the power, T is the exposure time at a location and D is the distance between the two locations on a line. 5 175 Continuous Pulsed 150 Beam center position (µm) 125 100 75 50 25 0 0 50 100 150 200 250 Time (µs) Figure 2. Beam position as function of time for pulsed and continuous energy deposition. 2.3. Melt-pool Dynamics Prediction We considered a single-pass laser simulation on a 316L stainless steel plate (100 µm thick, 850 µm long). The baseline thermal field predictions were obtained using a high-fidelity finite element multiphysics solver implemented in the MOOSE framework, based on the formulation in Noble et al. [11]. The plate dimensions were chosen because they would allow the melt pool to develop fully. The simulation setup is presented in Figure 3. Figure 3. Simulation results for the process input (74.77 W , 1.0 m/s, 125 µm) sample at 303 µs. The velocity of the melt pool is visualized using unscaled arrows. The melt-pool deformation, which has a Gaussian distribution, is also visible at the leading edge of the laser beam. This model showcases relevant phenomena for many AM techniques such as direct energy deposition or LPBF. The model incorporates the temperature- and phase-dependent material properties of 316L stainless steel, laser-material interaction via a Gaussian heat source, vapor pressure recoil, radiative heat losses, and surface tension. The deformation of the melt-pool surface is simulated using an ALE approach involving the fully coupled solution of the energy, momentum, and mass conservation equations. The full-field spatiotemporal temperature evolution is obtained on the fly by treating the laser heat source as a time- dependent heat flux boundary condition on the top surface of the domain. For a more detailed description of 6 the computational model, we refer the reader to Biswas et al. [15]. The simulated temperature fields reveal classical melt-pool signatures consistent with LPBF processes, showing rapid thermal gradients near the melt-pool boundary, highly localized temperature peaks exceeding 2500 K, and strong anisotropy along the laser travel direction. These features are comparable in scale and topology to the experimental and numerical results reported by Lamb et al. [16], who examined the melt- pool geometries and solidification behavior in Ti-6Al-4V under simulated AM conditions. The melt-pool elongation and boundary instability are captured well in our simulation outputs, especially under high laser power and low scan velocity conditions. Time-dependent snapshots from the simulation further exhibit melt-pool oscillations and convective front instability, which are particularly evident at intermediate laser velocities [17]. The model serves as a data generator for the machine learning workflow. To investigate the solution manifold, three process parameters were selected: effective laser beam radius, effective laser power, and laser velocity. These parameters are known to vary significantly, both across different configurations of a single additive-manufacturing machine and between different machines. Figure 4 demonstrates the influence of process parameters on the peak flow velocity of the melt pool. The results show that increasing the laser power results in a larger melt pool with a higher maximum melt-pool temperature. Conversely, increasing scanning speed reduces the melt-pool size because the linear energy deposition is reduced. On the other hand, depending on the laser power, changing the effective laser radius can both increase and decrease the melt-pool size. In total 498 simulations were performed for different values of the process parameters using the distributions presented in Table 1. Table 1. Values of input parameters in the training dataset. Parameter Value 8/31/25, 9:10 PM op_0.svg8/31/25, 9:11 PM op_167.svg Effective laser power(ηP ) [W] U(60, 75) Laser scanning speed (v) [m/s] U(1.0, 1.5) Effective laser beam radius(R) [µm] U(125, 155) 8/31/25, 9:10 PM op_15.svg (a) (63.1442 W, 1.2530 m/s, 139.7014 µm) (b) (72.5904 W, 1.0465 m/s, 135.3907 µm) (c) (72.5935 W, 1.0465 m/s, 135.3907 µm) Figure 4. Melt-pool geometries for three representative input process parameters to show the influence of process parameters on the peak flow velocity of the melt pool. Results show that increasing the laser power results in a larger melt pool with a higher maximum melt-pool temperature. Conversely, increasing scanning speed reduces the melt-pool size because linear energy deposition is reduced. 7 The output temperature maps were collected to serve as the ground truth for training and evaluating the machine learning models. We note that that the computational cost of simulating the full thermal trajectory over a 51 × 162 spatial grid for one process parameter configuration exceeds 20 minutes to 2 hours of CPU time on average depending upon the process inputs (with a considerable variance based on whether the melt pool reaches evaporation temperatures or not). This underlines the need for machine learning models that can enable the rapid exploration of process-structure relationships across the design space. 2.4. Comparison Of The Pulsed Vs. Continuous Energy Deposition We developed two numerical examples to compare pulsed and continuous energy deposition: a two- dimensional case and a three-dimensional case. The two-dimensional example incorporated the complete set of physical processes described in Section 2.1, capturing the full dynamics of the system. In contrast, the three-dimensional example employed a simplified conduction-only model, neglecting the melt pool fluid dynamics and surface deformation effects in order to reduce computational complexity while still illustrating the key differences between pulsed and continuous deposition. 2.4.1. 2D Example For the two-dimensional case, we prepared a model that incorporates the full set of physical processes described previously. The material properties were taken from the earlier study [15], and the computational domain extended [−80, 240] µm in length with a depth of 100 µm. The effective laser power was set to 28 W with an effective radius of 70 µm, and the position of the laser centroid is shown in Figure 2. The maximum simulated time was 240 µs, using adaptive time-stepping with time increments in the range of 10−9 − 10−8 s. For the continuous energy deposition case, the laser scanning speed was fixed at 0.75 m s . The two-dimensional domain was split into 300 intervals in the x direction with 80 intervals in the y direction, yielding 24,000 elements in total. First order Lagrange element have been used for every variable. A Streamwise-Upwind Petrov Galerkin (SUPG) [18] stabilization was used for the Navier-Stokes equations. Figure 5 showcases the difference of melt pool shape and position during a laser scan. We see a noticable difference in the meltpool shapes. The meltpool with pulsed source is 11.6%-13.1% deeper at the presented time steps. This is in agreement with the observations mad in [13]. However, when the pulsed source transitions from one position to the other, the meltpool can shrink. 2.4.2. 3D Example To extend the comparison into three dimensions, we constructed a conduction-only model that neglects surface deformation while retaining the dominant heat transfer mechanisms. The material properties were taken from a previous study [15]. The effective laser power was set to 30 W, a scanning speed (for continuous deposition) to 0.75 ms , and a Gaussian effective radius of 70 µm. The laser position follows that shown in Figure 2 for the pulsed source. The computational domain was discretized with a mesh of 300×120×60 elements in the x–y–z directions, spanning dimensions of [−80, 240] µm in x ×[−80, 80] µm in y, and [0, −70] µm in z. This setup provided sufficient resolution to capture the conduction-driven thermal field while maintaining computational feasibility. The simulation was carried out for a maximum time of 240 µs altogether. Similarly to the 2D case, first order Lagrange element have been used for every variable. The compared results at t = 200 µs are presented in Figure 7. We see obvious differences in the melt pool geometry. In these figures, the meltpool is cropped out using a threshold filter to better visualize the melt pool geometry. To further investigate this, the vertical cross section at the centerline is presented in Figure 7. We see that the pulsed deposition due to it staying in the same position for an extended period of time results in overall 8 t = 8 × 10−5 s t = 1.6 × 10−4 s t = 2.4 × 10−4 s Figure 5. Melt pool geometry in the 2D model throughout the laser pass with a pulsed source (top) and a continuous source (bottom) of the same linear energy density. Figure 6. Meltpool geometry at the end of the simulation (t = 2 × 10−4 s) for continuous laser power delivery (bottom) and pulsed power deliver (top). 9 larger and more uneven melt pools. This is in agreement with the literature [13]. The rigorous validation of this methodology is part of future studies. t = 8 × 10−5 s t = 1.6 × 10−4 s t = 2.4 × 10−4 s Figure 7. Melt pool geometry in the 3D model throughout the laser pass with a pulsed source (top) and a continuous source (bottom) of the same linear energy density. 10 3. UNCERTAINTY-AWARE MACHINE LEARNING MODEL FOR MELT-POOL DYNAMICS In this section we provide the probabilistic machine learning for predicting the temperature and meltpool dynamics using a latent multi-output Gaussian process auto-regressive (L-MOGPAR) model. The training data set is generated using the arbitrary Lagragian-Eulerian (ALE) method with free-surface tracking in the multiphysics object-oriented simulation environment (MOOSE) [9, 19], presented in Section 2.3. An L-MOGPAR models the covariances between both the inputs (i.e., process parameters) and outputs (i.e., temperature fields) and captures time dependencies using a Markov property [20–22]. In addition to predicting the dynamics, L-MOGPAR provides uncertainty estimates over the fields. Complimentarily, a UNet is developed, which is adept at processing image data, to predict the temperature and meltpool dynamics in a deterministic fashion. The performance of both the model to predict the melt pool geometry and temperature is compared. 3.1. Latent Multi-output Gaussian Process Auto-regressive Model A probabilistic nonlinear model, L-MOGPAR, was used to predict the underlying time-dependent behavior of the AM process. The spatial covariance of the temperature field from the laser melt-pool simulations is high dimensional, nonlinear, and time varying. Hence, we seek a low-dimensional representation (latent space) of this large temperature field using principal component analysis (PCA) [23]. To keep the time-dependent behavior of the AM process intact, the dimensionality reduction is performed on each time snapshot of the temperature field for a given combination of process parameters. There are three controllable process parameters: laser power, scanning speed, and laser beam radius, which are discussed further in Section 2.3. Using PCA, the high-dimensional 51 × 162 spatial grid temperature field is first reduced in dimensionality by identifying the orthogonal basis (eigenfields) that captures the dominant spatial patterns of temperature variation. We retain the top r = 25 principal components at each time step, which explain approximately 99% of the variance in the training data. Each full temperature field is thus represented by a vector of PCA coefficients, ζ t ∈ Rr . We employ a multi-output Gaussian process in the latent dimension (L-MOGP) to model the relationship between process inputs and the PCA-reduced outputs, and extend it in time via an autoregressive (AR) formulation. GPs provide a Bayesian non-parametric approach to regression, and L-MOGPAR generalizes this to vector-valued outputs that vary in time by defining a joint covariance across all output dimensions [20]. The principal component (PC) scores representing the spatial correlations of the evolving temperature field at time t can be expressed using a GP surrogate as ζ t ∼ GP(µt (x), k t (x, x0 )), (12) where x is the input process parameter feature vector, and µt (·) and k t are the mean and covariance kernel at time t. A covariance kernel that captures the time-dependent evolution of the covariance across the inputs and outputs is employed based on the squared-exponential kernel with input process parameters x augmented by ζ t−1 , . . . , ζ 1 as 0 0 k t (x, x0 ) = k t ({x, ζ t−1 , ..., ζ 1 }, {x0 , ζ t−1 , ..., ζ 1 }) D t−1 Pr c 0c 2 2 1 X (xd − x0d )2 1 X p=1 (ζp − ζp ) (13) = σ exp(− − ), 2 ld2 2 c=1 lc2 d=1 where the input covariance is now a function of the inputs x and the outputs of the previous time steps ζ t−1 , ..., ζ 1 , and ld and lc represent the length scale for the inputs and the previous outputs, respectively. The sensitivity to the previous states is governed by the length scales lc . In addition, simultaneous predictions 11 made for all the PC scores considered in the L-MOGPAR are considered through the linear model of co- regionalization (LMC) [20–22]. A set of Q basis kernels, {kqt (x, x0 )}Q q=1 , is introduced to model the r covariates jointly. The kernel for the cross-covariance of any pair of PCs is modeled by a weighted, structured linear mixture of the Q basis kernels. The full joint kernel is expressed as Q X t K = Bq ⊗ kqt (x, x0 ). (14) q=1 This covariance kernel enables improved performance through encoding correlations between output dimensions and through the linear superposition of multiple latent GPs. Properties of the time series observations, such as periodicity and short-term dependencies, are captured in the Q basis kernels. For k t (x, x0 ) to be symmetric positive, Bq has to be positive definite to produce a Mercer kernel. So, Bq is parametrized as a rank 1 matrix given by Bq = aq a|q + diag(λq ), (15) where aq ∈ Rr×1 , λ q ∈ Rr×1 . In multi-output GP models, the linear correlations between the output variables are fixed across the input space. Thus, the L-MOGPAR models the joint probability of the time-dependent multiple outputs in the latent dimension ζ t , ..., ζ T by factorizing output at time t to depend upon the previous k time steps, where k is the Markov order of the autoregressive model [24–27]. For this study, k = 1 is considered, and the conditional joint probability of the outputs can be expressed as [28] T Y p(ζ 1 , . . . , ζ T | x) = p(ζ 1 | x) p(ζ t | x, ζ t−1 ), (16) t=2 where T is the total number of time steps. The first factor is a standard MOGP, while each conditional p(ζ t | x, ζ t−1 ) is itself modeled by a GP that takes the previous output as part of its input. Unlike a linear autoregressive model, this GP�based formulation captures nonlinear, input�dependent dependencies between consecutive states. The outputs of the model are the temperature fields Y1 , ..., YT , which are learned through the PCA of the outputs trajectories ζ 1 , ..., ζ T . At each discrete time step, the L-MOGPAR receives as input (x, ζt−1 ), thus learning the evolution operator directly from data without a separate time coordinate. The surrogate was trained on N = 473 simulated trajectories, each comprising T = 10 snapshots over 800µs. In addition, we set Q = 15 in Equation (14). The trainable hyperparameters are ξ = {ld }rd=1 , lc , {aq }Q Q  q=1 , {λq }q=1 , σ . To introduce regularization into the training, a log-normal prior f (ξξ ) is introduced to the loss function. All hyperparameters are learned by maximizing the summed log marginal likelihood of the GP chain over time and the prior over the trainable hyperparameters as T X 1 1 1 L= − log | K̂t | − (ζ t )| (K̂t )−1 ζ t − Ns log(2π) − log(f (ξξ )), (17) 2 2 2 t=1 where K̂t = Kt + σζ2 I. An Adam optimizer [29] with learning rate 5 × 10−3 was used for this study. For a new data point, x? , in the test dataset the GP first predicts p(ζ1 | x? ). The mean and covariance of the predicted outputs r at time t are given by µt∗ = Kt∗ (Kt )−1 ζ t (18) Σt∗ = Kt∗∗ − (Kt )−1 (Kt∗ )| , (19) t PQ t t PQ t where K∗ = q=1 Bq ⊗ kq (x, x∗ ) and K∗∗ = q=1 Bq ⊗ kq (x∗ , x∗ ). Mapping these PCA statistics back through the eigenbasis delivers full�field temperature predictions with credible intervals. The optimized length�scales ld reveal the relative influence of each process parameter, and the predictive variance quantifies confidence in the test predictions. 12 3.2. CNN-based UNet A convolutional neural network (CNN)-based UNet was built to learn the time-evolving temperature field from the AM process. The UNet is composed of an encoder-decoder-type architecture with additional skip connections. The encoder extracts the small feature map from the input image by convolution and pooling layers, and the decoder expands it to the original image size by convolution and upsampling layers. The input vector contains the effective laser input parameters, and the feature map extracted from the inputs is then passed through several upsampling and convolution layers to increase the spatial resolution and decrease the number of channels. The encoder consists of five convolution blocks (ConvBlocks) with output channels 64, 128, 256, 512, and 1,024, each followed by max-pooling layers to downsample the feature maps. Each ConvBlock includes two convolution layers with instance normalization [30] and rectified linear unit (ReLU) [31] activation. The instance normalization layer is applied to each channel of the data, which corresponds to normalizing the output data at each time step. Downsampling is implemented as strided convolutions (motivated by representational bottleneck [32]), and upsampling is implemented as convolution transposed. The bottleneck layer connects the encoder and decoder, doubling the number of channels from 1,024 to 2,048. To transmit high-resolution information in the input images, the network has skip connections between each convolution layer of the encoder and the corresponding upsampling layer of the decoder. Each skip connection simply concatenates all channels at the layer of the encoder with one of the decoders. The decoder path includes upsampling layers and ConvBlocks, with output channels 1024, 512, 256, 128, and 64, and it uses skip connections to concatenate corresponding encoder feature maps and retain high-resolution features. The final output layer is a 2D convolution layer that reduces the number of channels to the number of discrete time steps, producing the segmentation map, and bilinear interpolation ensures the output dimensions match the original input size. A schematic of the UNet is shown in Figure 8. 64 64 10 128 128 256 256 512 512 1024 3x3 2D convolution Instance normalization 2x2 max pooling ReLU skip connection 3x3 2D convolution Instance normalization 2x2 transpose 3x3 2D convolution ReLU convolution Figure 8. Schematic of the UNet architecture. The UNet model was trained over 10,000 epochs, with an initial learning rate of 5e − 4 to optimize the parameters of the network. A scheduler [33] dynamically adjusts the learning rate based on the validation loss to improve convergence. The UNet model, implemented using PyTorch, was trained using an L1 loss criterion and optimized with the Adam optimizer [29]. 13 3.3. Machine Learning Melt-pool Predictions: UNet And L-MOGPAR To reduce the computational burden while preserving fidelity, two machine learning models were trained with spatiotemporal temperature fields generated with the MOOSE multiphysics model discussed in the previous section: a probabilistic L-MOGPAR model and a deterministic UNet model. In this section we compare three aspects of these models’ simulation results: temperature field, melt-pool morphology, and solidification. For all three comparisons, the UNet and L-MOGPAR predictions were consistent with those of the MOOSE multiphysics model. 3.3.1. Temperature field Both the UNet and L-MOGPAR models successfully replicated the spatiotemporal evolution of the melt-pool temperature fields compared to the MOOSE multiphysics simulations, though their accuracy and output characteristics differed slightly. Figure 9 compares the temperature fields of the MOOSE multiphysics model of a representative test data point to the UNet and L-MOGPAR predictions at three time steps: (1) row 1 at t = 267.7µs, (2) row 2 at t = 444.4µs, and (3) row 3 at t = 800µs. The UNet’s predictions closely match the MOOSE multiphysics model temperature fields, capturing the high-temperature melt-pool region and its evolution. In contrast, the L-MOGPAR model produces relatively smoother mean predictions and slightly diffuses the sharp thermal gradients. The primary discrepancy between the machine learning models and the MOOSE multiphysics model is observed in the melt-pool area, particularly under high energy density regimes. The performance of the machine learning models is evaluated at the center point of the domain at (425, 49) µm with time in Figure 10a and along the center line at the top surface at t = 444.4 µs in Figure 10b. These figures highlight that the UNet model’s predictions for both melt-pool depth and length closely align with those of the MOOSE multiphysics model, while the predictions of the L-MOGPAR model are not much different than those of the UNet model. Figure 9. The temperature fields of a representative test data point of the MOOSE multiphysics model compared to the UNet and L-MOGPAR predictions at three time steps: (i) row 1 at t = 267.7 µs, (ii) row 2, at t = 444.4 µs, and (iii) row 3, at t = 800 µs. To evaluate the localized error distribution in the temperature maps, these figures show the relative error (RE) for each pixel using: ij ij |Tpred − Ttrue | RE = ij , (20) max(Ttrue ) 14 (a) (b) Figure 10. A comparison of the temperature profiles of the multiphysics simulation, L-MOGPAR, and UNet: (a) temperature history of the center point on the top surface, and the (b) temperature profile along the center line on the top surface at t = 444.4 µs. ij ij where Tpred and Ttrue are the true and predicted pixel temperature fields at pixel (i, j). Rather than normalizing with respect to each true pixel value in this expression, the normalization is done with respect to the maximum of the true pixel values, to highlight the significance of the error within the temperature map’s dynamic range. Figure 11 compares the relative errors of the machine learning predictions with respect to the outputs of the multiphysics model. The temperature maps predicted by the UNet model are very similar to the true maps, with REs in the range of 0%–5%, 0%–6%, and 0%–5% for a test point at t = 267.7µs, t = 444.4µs, and t = 800µs, respectively. The mean RE values for the UNet predictions are 0.3%, 0.54%, and 0.4% at those three times. In comparison, the L-MOGPAR produced mean RE values of 0.78%, 1.48%, and 0.85% at the three times with a maximum RE of approximately 15% at t = 444.4 µs. 3.3.2. Melt-pool morphology It is important to analyze melt-pool morphology since it reflects the interactions between LPBF process parameters and material solidification. Figure 12 presents the melt-pool boundaries for a representative test point at three time steps. Both machine learning models track the overall morphological evolution from initial melting through steady-state elongation to post-laser shrinkage. To visualize the local differences, the quantified melt-pool boundaries are identified in Figure 13 at t = 267.7 µs, t = 444.4 µs, and t = 800 µs. The melt pool marches in the positive x-axis direction, transforms into a long and shallow shape, and shrinks after the laser has passed the material domain. Both machine learning models produce morphologies that are quantitatively similar to that of the MOOSE multiphysics model in terms of depth (Figure 13a), length (Figure 13b), and area (Figure 13c) for a representative test sample. The L-MOGPAR model lags slightly in its tracking of rapid morphological changes, particularly near the end of the simulation, whereas UNet produces more accurate morphological change predictions. Across all the test points, both machine learning models produce accurate melt-pool area evolution predictions, on average, as seen from Figure 13d. Table 2 presents the goodness-of-fit (R2 ) values of the temperature predictions and their spatial gradients across all the test points. While both models predict these quantities well, UNet has an edge since its R2 values are consistently greater than 0.9 and close to 0.99 for the temperature predictions. 15 Figure 11. Pixel-wise relative error maps corresponding to the predictions in Figure 9. Figure 12. Comparison of melt-pool boundaries predicted by the multiphysics simulation, L-MOGPAR, and UNet models for three time steps of a representative test point. Both machine learning models track the overall morphological evolution from initial melting through steady-state elongation to post-laser shrinkage. 16 Melt pool depth Melt pool length (a) Melt pool depth (b) Melt pool length Melt pool area (c) Melt pool area (d) Mean melt pool area Figure 13. Quantitative comparison of melt-pool morphology evolution over time: the (a) melt-pool depth, (b) length, and (c) area of a representative test sample, and the (d) average melt-pool area evolution across all test samples. Table 2. Goodness-of-fit (R2 ) values of the temperature predictions and their spatial gradients. Generally, both models predict these quantities well, but UNet has an edge. ∂T ∂T Model T ∂x ∂y UNet 0.9861 0.9183 0.9384 L-MOGPAR 0.9311 0.7456 0.8504 3.3.3. Uncertainty quantification In addition to generating deterministic predictions of the temperature fields, quantifying the predictive uncertainty of the model predictions to evaluate their reliability is also important. Uncertainty-aware models can be very useful for optimization and control tasks, as they allow researchers to quantify confidence in different scenarios and guarantee the robustness of the calculations [34]. L-MOGPAR inherently provides uncertainty estimates over the mean predictions of the spatiotemporal temperatures. This means that the model provides a position-resolved uncertainty for any mean value in the feature space, unlike the R2 value, which only rates the performance accuracy of the GP model overall. Figure 14 shows the predicted standard deviation (first row), absolute error in the prediction (second row), and mean predicted temperature field (third 17 row) at three representative time steps for a test sample. The standard deviation maps reveal that L-MOGPAR is less confident in regions where and at time steps when the melt pool is evolving rapidly—notably around the melt-pool boundaries and during the peak heating period. This result makes sense because models struggle to learn complex regimes involving nonlinear physics at the melt-pool front and solidification boundary. The L-MOGPAR model reflects this complexity by assigning such regions a larger standard deviation. Conversely, in regions far from the melt pool in the unheated/cooled substrate, the L-MOGPAR uncertainty is low. Importantly, these uncertainty estimates correlate with the error distribution—where L-MOGPAR shows large errors, it also reports high uncertainty. Whereas the computation of the error estimates requires the ground truth solution, the standard deviations provided by L-MOGPAR does not require it. The UNet model does not provide uncertainty as it outputs a single point estimate of the temperature field. Figure 14. The spatially resolved standard deviation (first row), absolute error (second row), and mean predictions (third row) from L-MOGPAR at three representative time steps. The model predictions are less certain near the melt-pool boundaries and during peak heating. These regions coincide with a greater prediction error. 18 4. MODELING MICROSTRUCTURAL EVOLUTION DURING THE AM PROCESS Microstructural evolution during the AM process is a complex phenomenon encompassing different length scales. At the grain level, the morphologies of the grains vary from build to build, as well as within the same build, depending on the process conditions and build directions. At the subgrain level, various solidification structures and dislocation cells with different shapes are formed owing to such process variations. Recently, a MOOSE-based model of alloy solidification for polycrystals has been developed that can predict the geometry of solidified structures based on the variation in the thermal gradient and cooling rate [35]. We extend the model to incorporate 316 SS thermodynamics and properties to capture the micro-segregation that occurs during solidification. We evaluate how process conditions influence the formation of various dislocation cell structures within the subgrains of the additively manufactured material. This is aims to quantitatively predict the subgrain features that are important for evaluating the properties and performance of the AM parts (Figure 15 Figure 15. Effect of microstructure on the propertie s and irradiation performance of the AM materials. Grain morphology, precipitation on dislocation cell walls influences the irradiation performance an creep properties. In this section, we present the solidification model and investigate the impact of accurately predicting the process dynamics using machine learning on the microstructure evolution, with an emphasis on solidification rate, the shape of the solidified structure and chemical microsegregation. 4.1. Phase-field Model For Solute Segregation During Solidification In this section, we present the phase-field model to capture the solidification and associated solute segregation observed during the laser powder bed fusion AM method. Here, we utilize a grand-potential- based multi-phase, multi-order parameter, multi-component phase-field model to capture the solidification process. The model was initially developed and implemented in MOOSE to capture alloy solidification [36? , 37]. Considering the non-equilibrium rapid solidification that occurs during the AM process, we extend the model to incorporate the CALculation of PHAse Diagrams (CALPHAD) method to obtain the relevant phase diagrams and the thermodynamic properties for 316 SS [38, 39]. The details about the model formulation and its extension to connect to thermodynamic database, along with verification and preliminary demonstration, are provided below. 4.1.1. Model formulation The phase-field model represents the microstructure with solid and liquid phases consisting of K possible chemical species. For computational simplification, 316 SS has been modeled as a ternary alloy with 17.5 wt.% chromium (Cr) and 12.5 wt.% nickel (Ni), with the balance being iron (Fe). Multiple phases and grains can exist within the solid phase; as such, the solid can be represented by multiple order parameters 19 corresponding to different solid phases or grains within those phases. There are Ns possible solid phases and grain orientations such that they are individually represented by a set of non-conserved order parameters: ~ηs = (ηs1 , ηs2 , . . . , ηsNs ). Within phase or grain i, ηsi = 1 and all other order parameters are 0. The interface between phase or grain i and j is represented by the smooth variation of order parameters ηsi from 1 to 0, and ηsj from 0 to 1. Finally, the liquid phase is represented by a single order parameter ηl . In addition to the phases and crystallographic orientations, the chemical composition of the material system is represented with the help of the number density of the chemical components. Here, the number density ρ of the solute species Cr and Ni are tracked at each position. Here, Fe is considered the dependent component such that its atomic fraction is defined as: cFe = 1 − cCr − cNi , where cI represents the atomic fraction of different components in the material. The evolution of the order parameters (ηl , ηsi ) are represented by Allen-Cahn equations derived from the grand potential functional such that ∂ηl δΩ = −L (21) ∂t δηl ∂ηsi δΩ = −L (22) ∂t δηsi Here, L is the Allen-Cahn mobility, and Ω is the total grand potential functional of the system: Z Ω= (ωbulk + ωgrad + ωchem ) dV , (23) V where ωbulk is the bulk free energy density term, ωgrad adds the gradient energy contribution, and ωchem incorporates the appropriate grand potential densities for each phase such that:      4 2  X ns  4 2  ns ns X ns η η ηsi ηsi γls X X γsisj 2 2  1  ωbulk = m  l − l + − +  ηl2 2 ηsi + η η + , (24) 4 2 4 2 2 2 si sj 4 i=1 i i i6=j Ns ! κ X ωgrad = |∇ηl |2 + |∇ηsi |2 , (25) 2 i=1 and ωchem = hs ωs + hl ωl . (26) Here, s and l represent the solid and liquid phases, i and j are the indices for the phases/grains within the solid, Ns is the total number of grains, m is the free energy barrier coefficient, and γls and γsisj are a set of constants that enable adjustment of the interfacial energy between solid and liquid phases as well as between the ith and j th grains of the solid phase, respectively. κ is the gradient energy coefficient, ωs and ωl are the local grand potential densities for the solid and liquid phases, and hs and hl are the switching functions used to interpolate the grand potential densities between the phases defined as Pns 2 η hs = 2 i=1 Pns si 2 , (27) ηl + i=1 ηsi and η2 hl = Pl ns 2 , (28) ηl2 + i=1 ηsi for the solid and liquid phases, respectively [40]. The grand potential density for the phase ωs/l is given by ωs/l = fs/l − ρA µA − ρB µB − . . . − ρK−1 µK−1 , (29) 20 where fs/l is the Helmholtz free energy density of the phase considered and µA is the chemical potential between species A and the solvent. In this work, we define the grand potential densities of the phases from the CALPHAD free energy of the Fe-Cr-Ni system. The details parameterization is provided in 4.1.2. The evolution of the solute species ( Cr and Ni ) is represented by the generalized diffusion equation, K−1 ∂ρA X =∇· MAI ∇µI , (30) ∂t I=1 where MAI is a mobility coefficient. For the grand-potential-based model, however, the chemical potential of each species is used as a field variable for simulating the system [36]. Hence, the evolution equation is represented as N X nα ∂µA X ∂ρA ∂ηαi χAA = ∇ · (MAA ∇µA ) − , (31) ∂t ∂ηαi ∂t α=1 i=1 where χAA is the susceptibility. During alloy solidification, the solute diffusivity in the solid is usually several orders slower than in the liquid. This results in spurious solute trapping in the diffuse interface model, becoming pronounced with upscaling of the interface thickness. To mitigate this, an antitrapping current was added to the model, counterbalancing the solute trapping by expelling the solute from the freezing material [36]. This restores the local equilibrium at the interface, with the help of the additional solute current expressed as h i ∂η αi j~at = −at W ρlI (µI ) − ρsI (µI ) n̂ , (32) ∂t ~ where n̂ = ∇η αi ~ αi | is the unit normal to the solid-liquid interface, W is the interface thickness, and at is |∇η the antritrapping interpolating coefficient. Furthermore, for non-equilibrium solidification conditions the antritrapping coefficient is modified further to incorporate controlled trapping. The modified coefficient is expressed as, 1 at → avt = √ (1 − A(1 − ηs2 )), (33) 2 2 where A is a trapping parameter introduced to set a finite amount of trapping. The modified mass conservation equation for a specific solute species is represented as ∂ρA   = −∇. j~ρA + j~at ∂t " #  ~  ∇η ∂ηαi αi = ∇. MAA ∇µA + aW ρlA − ρsA . (34) ~ αi | ∂t |∇η Thus, Eq. 31 is modified as nα N X ! ∂µA h ~ i ∇η αi∂ηαi X ∂ρA ∂ηαi χAA =∇· MAA ∇µA + aW ρlA − ρsA − . (35) ∂t ~ |∇ηαi | ∂t ∂ηαi ∂t α=1 i=1 Furthermore for non-isothermal conditions, the temperature dependence needs to be accounted for [41]. Therefore, Eq. 35 is further modified as N Xnα ! ∂µA h ~ αi ∂ηαi i ∇η X ∂ρA ∂ηs,l ∂ρA ∂T χAA = ∇ · MAA ∇µA + aW ρlA − ρsA − − . (36) ∂t ~ αi | ∂t |∇η ∂ηs,l ∂t ∂T ∂t α=1 i=1 21 The evolution equation for the temperature is expressed as α N n ∂T XX ∂S ∂ηαi ∂S ∂µ cp = ∇ · k∇T − T − , (37) ∂t ∂ηαi ∂t ∂T ∂t α=1 i=1 where k is the thermal conductivity, cp is the heat capacity, and S is the generalized entropy density defined as ∂Ω S=− . (38) ∂T In this work, we use a dimensionless reduced temperature, which is a scaled value with respect to the melting temperature, defined as T − Tm Tr = , (39) Tm where Tm is the melting temperature and Tr is the reduced temperature. However, for directional solidification, a thermal profile can be applied on the simulation domain based on the process conditions. The temperature profile can be specified using the function Tr = Tr0 + G [(~r − r~0 ) − V t] , (40) where ~r is the position vector representing spatial location, Tr0 is the initial reference temperature at ~r = r~0 , G is the temperature gradient, and V is the velocity. Thus, the cooling rate is defined as, Q = G · V . 4.1.2. Phase-field model parameterization To parameterize the model for 316 SS, thermodynamic calculations are carried out assuming a ternary alloy with nominal composition of 17.0 wt.% Cr and 12.0 wt.% Ni and the balance being Fe. This corresponds to the mole fraction of 0.181, 0.113, and 0.705 for Cr, Ni, and Fe, respectively. The phase diagram of the Fe-Cr-Ni system at 1600 K is presented in Figure 19a. A Scheil calculation was also performed to identify the stable solid phases and solidification path under non-equilibrium conditions. The free energy of the Fe-Cr-Ni ternary system is constructed following the Redlich-Kister model (RK) used by Miettinen [39]. The thermodynamic model uses the mixing parameters proposed by Miettinen [39] and the pure component parameters from Dinsdale [38]. The free energy calculation was verified with the single point equilibrium calculation performed in ThermoCalc. The grand potential density requires a one-to-one relationship between the composition and the chemical potential within a phase. As such, using the regular solution or RK thermodynamic model is not a viable option. An approximation can be made using a parabolic free energy density to model each phase [42]. An example of a free energy and its parabolic approximation fitting is shown in Figure 17a and 17b. Considering that the phase fraction for the BCC ferrite phase in only 5%, we only consider the liquid to FCC-A1, i.e. austenite, transformation in this work. The model can be extended to include the BCC ferrite transformation in the future. The parabolic free energy for the solid and liquid phase is defined using a second order Taylor expansion such that [42] K−1 1 X s,l 1 χIJ µI − µeq µJ − µeq µI − µeq  eq ω s,l = ωeq s,l   − I J − I cI (41) 2Va Va I,J=1 ∂ω s,l 1 s,l 1 s,l 1 ρs,l χAA µA − µeq χAI µI − µeq + ceq   A =− = A + I (42) ∂µA Va 2Va Va I N   X s,l eq  1 s,l eq  eq cA = Va ρA = hβ χAA µA − µA + χAI µI − µI + cI (43) 2 β=1 22 h 2 s,l i−1 Here, the χs,l IJ is the susceptibility defined as ∂ G ∂cI ∂cJ , with Gs,l is the CALPHAD free energy of the solid and liquid phases. The equilibrium concentrations and derivatives of the free energy are obtained from the CALPHAD free energy for fitting the parabolic grand potential. (a) (b) Figure 16. (a) Ternary phase diagram of the Fe-Cr-Ni system at 1600K. (b) Scheil calculation showing the solidification path and corresponding stable phases. CALPHAD Data CALPHAD Data Parabolic Fitting Parabolic Fitting Gibbs Free Energy Gibbs Free Energy −0.94 −1.08 (eV/atom) (eV/atom) −0.96 −1.10 −0.98 −1.12 −1.00 −1.14 −1.02 −1.16 −1.04 −1.18 −1.06 −1.20 0.4 0.4 0.3 0.3 n n io io 0.0 0.2 0.0 0.2 at at 0.1 0.1 r r nt nt 0.2 0.1 0.2 0.1 ce ce Cr C Cr C on on once 0.3 0.0 o 0.3 0.0 ic ic ntra 0.4 ncen 0.4 trat N N t ion ion (a) (b) Figure 17. Free energy calculation at for (a) solid at 1600 K and (b) liquid phase at 1740 K. For the free energy calculations, the equilibrium concentration of the species are considered same as the nominal concentration. The equilibrium concentration of the species in each of the phases are obtained from the ThermoCalc single point equilibrium calculations. The variation of the equilibrium concentration of Cr and Ni in the solid FCC phase as a function of temperature is provided in Figure 18. The fitted correlations presented in 18 are only valid between the solidus (1660 K) and liquidus (1740 K) temperatures. Beyond that, the equilibrium concentration is same as the nominal composition. The free energy coefficients corresponding to Equation 41 23 are calculated as a function of temperatures. These coefficients are then fitted to a closed form function with respect to reduced temperatures. Sample fitting plots are shown in Figure 19. It is observed that a higher (a+bT +cT 2 ) order polynomial or rational fitting function of the form f = (1+dT +eT 2 +f T 3 ) is able to appropriately capture the functional form for these coefficients, specifically between the solidus and liquids temperatures. 0.185 0.122 0.180 0.120 Cr Concentration, csCr Ni Concenration, csN i 0.175 ThermoCalc Data 0.118 Fitted curve Nominal Composition 0.170 0.116 0.165 0.114 0.160 0.112 −0.05 −0.04 −0.03 −0.02 −0.01 0.00 −0.05 −0.04 −0.03 −0.02 −0.01 0.00 Reduced Temperature, Reduced Temperature, (T r = (T − Tm)/Tm) (T r = (T − Tm)/Tm) Figure 18. Concentration variation with reduced temperature −0.3 Original data 0.04 Data Original data Fitted curve Rational Function fit Fitted curve −0.4 −0.22 −0.5 0.03 −0.24 −0.6 −0.7 0.02 Ni Cr ωeq µeq µeq −0.26 s −0.8 −0.9 0.01 −0.28 −1.0 −0.30 −1.1 0.00 −0.5 −0.4 −0.3 −0.2 −0.1 0.0 −0.5 −0.4 −0.3 −0.2 −0.1 0.0 −0.5 −0.4 −0.3 −0.2 −0.1 0.0 Reduced Temperature Reduced Temperature Reduced Temperature (a) (b) (c) s (b) µeq , and (c) µeq Figure 19. Fitted correlation for different parabolic coefficients: (a) ωeq Cr Ni Interfacial properties play a key role in relating the model parameters to the physical properties and also dictate the microstructural evolution. In this model, anisotropic interfacial properties are introduced by considering the dependence on the misorientation between adjacent grains and the inclination of the interface. Here, the interface inclination is determined based on the interface normal calculated from the gradients of the transitioning adjacent order parameters, ∇ηsi − ∇ηlj nsilj = , (44) |∇ηsi − ∇ηlj | where nsl indicates the outward unit normal to the interface between the phases represented by the variables ηsi and ηlj . In this case, it indicates the solid-liquid interface. In two dimensions, the inclination angle relative 24 to the x-axis (φsilj ) is calculated from the normal as nsilj y   φsilj = arctan . (45) nsilj x To introduce interfacial energy anisotropy, the gradient energy coefficient is computed based on the inclination and reference orientation κsl = κ0 [1 + δ cos {q(φsl − φ0 )}]2 , (46) where κsl is the interfacial coefficient between the solid and liquid phases, δ is the strength of anisotropy, q is the mode number, and φ0 is the reference angle of the grain. The average interfacial coefficient, kappa0 , is calculated as [35] 3 κ0 = σl, (47) 4 where σ is the interfacial energy and l is the interface thickness. The bulk energy coefficient, m, is computed as σ m=6 . (48) l 4.2. Solidification Demonstration 4.2.1. Model verification and parameter selection For verification of the model, we simulate solidification and the corresponding solute segregation with the growth of the planar interface. A simplified domain ranging from y = −20 to y = 20 and x = −1 to x = 1 was initialized with a planar solid-liquid interface. Assuming a length scale of 10µm, this represents a 400µm × 20µm domain. The interface was initialized with equilibrium conditions such that    1 x ηs = 1 − tanh √ , (49) 2 2 and    1 x ηl = 1 + tanh √ . (50) 2 2 The stability of the solid-liquid interface under equilibrium conditions was evaluated for T = Tm . The solute concentration within the solid and liquid was set at the equilibrium condition to avoid any solute transport. In the absence of any undercooling, the interface remains stable and maintains the equilibrium configuration, as there is no driving force for the solidification to occur. Next, we consider the growth of the solid under an isothermal cooling condition. With undercooling, when T < Tm , the planar interface moves at a steady pace over time as the solid area fraction increases due to solidification. Additionally, the initial supersaturation of solute is considered within the liquid to facilitate the solidification. Phase-field is a diffuse interface model and solidification characteristics highly depend on the interface width and corresponding mesh discretizations. In FY 2024, the effect of the interface width on solidification was studied. The interface width was varied from 200 nm to 80 nm and the steady-state growth rates for the solids are compared. Based on this analysis, an interface with of 100 nm was deemed sufficient to capture the appropriate solidification behavior in a computationally efficient manner. Reducing the interface width further only increased the computational cost without adding significant improvement to the results. Furthermore, the value of the Allen-Chan mobility, L, was varied to study the kinetics of solidification. From this analysis, L = 1 was selected, beyond which the solidification rate remains unaffected. This ensures that the solidification in the material is primarily driven by the thermal and solute diffusion, rather than the interface kinetics. Additionally, it was demonstrated that the steady-state solute concentration does not vary 25 significantly with varying interface width and mobility. This indicates that the model is not influenced by the choice of interface width or other model parameters and can capture the appropriate physical behavior. Details about this analysis can be found in Ref. [43]. Furthermore, considering that the solute diffusivity in solids is slower than liquid, the diffusivity of the overall of the system is defined as, D = hl DL + hs DS . For computational simplification, we assume the same diffusivity values for both of the components and vary them within the solid and liquid phases individually. Based on the parametric study done in Ref. [43], the non-dimensionalized diffusivity values were set as 1.0 and 1 × 10−3 . 4.2.2. Solute segregation during solidification To demonstrate the solute segregation, we simulate the growth of a solid nuclei in undercooled liquid field in two dimension (2D). For the first scenario, we set Tr = −2.0. At the initial conditions, the solute concentrations within the solid is set at the nominal composition, while a supersaturation is introduced within the liquid to facilitate solidification. The initial Cr and Ni concentration within the liquid are set at 0.2 atomic fraction. The generated solidification microstructure and corresponding solute concentration is provided in Figure 20. The corresponding solute tracking is presented in Figure 21. In this case, higher Ni segregation is observed compared to Cr segregation. This is primarily due to the higher level of deviation from the nominal composition at the initial condition. Note that although the solid demonstrates anisotropic growth, no dendritic structure formation has been observed in this case. Inducing surface undulation or high magnitude of interfacial noise is needed for dendrite formation. This will be explored in the future. 4.3. Process To Microstructure Connections In this section, we present the process to microstructure connection via surrogate model. To this end, we conduct microstructural simulations with the phase-field model presented in Section 4.1 in conjunction with the processing model. In Section 4.3.1, we compare the effect of process conditions on solidification using the temperature gradients and cooling rates obtained from the two surrogates, namely, L-MOGPAR and UNet, with the predictions from the multiphysics simulation. To accomplish this, we used the MOOSE stochastic tools module [44] to sample the temperature gradients and cooling rates obtained from the three models (i.e., L-MOGPAR, UNet, and multiphysics simulation). This module then launches phase-field simulations (using the phase-field module of MOOSE) in parallel corresponding to the temperature inputs using the MultiApp feature of the MOOSE [45]. 4.3.1. Solidification rate and morphology prediction In order to study the impact of the L-MOGPAR and UNet temperature field predictions at the material mesoscale, we performed phase-field simulations of solid formation due to the AM process. We performed these simulations for all three temperature prediction methods (i.e., multiphysics simulation, L-MOGPAR, and UNet) for three different input parameter sets, three time instances, and three locations around the melt pool. The three process parameter sets are input 1 (138.97 µm, 1.41 m/s, 74.11 W), input 2 (138.85 µm, 1.02 m/s, 68.27 W), and input 3 (154.16 µm, 1.17 m/s, 72.86 W). The three time instances used are 266.7 µs, 444.4 µs, and 800 µs. Figure 22 shows the melt-pool locations for the phase-field simulations for the three time steps of interest. In total, we performed 81 phase-field simulations of solid growth across the different prediction methods, time steps, spatial locations, and input parameters. The goal was to compare the solidification dynamics corresponding to the temperature fields predicted by the multiphysics simulation, L-MOGPAR, and UNet models. The phase-field solidification model itself was not approximated in any fashion using machine learning. Rather, the inputs for evaluating the phase-field model, such as the temperature gradients and cooling rate, were obtained from the three methods (i.e., multiphysics simulation, L-MOGPAR, and UNet) to predict temperature fields. 26 (a) (b) (c) Figure 20. (a) Growth of a solid nuclei, and corresponding (b) Cr concentration and (c) Ni concentration distribution at simulation time t = 50. 27 0.25 25 0.20 20 Solid Fraction (%) Cr Concentration 0.15 15 0.10 10 0.05 5 0.00 0 0 200 400 600 800 1000 0 2 4 6 8 10 Simulation Time Position (a) (b) 0.30 0.25 Ni Concentration 0.20 0.15 0.10 0.05 0.00 0 2 4 6 8 10 Position (c) Figure 21. (a) Evolution of solid fraction over time. Variation in (b) Cr concentration and (c) Ni concentration 28 Point 1 Point 3 Point 1 Point 3 Point 1 Point 2 Point 2 Point 2 Point 3 Figure 22. Representative input locations around the melt pool for the three different time steps at which the phase-field simulations for solid growth were performed. These simulations were conducted for three input parameter sets and the three methods (i.e., multiphysics simulation, L-MOGPAR, and UNet) used to predict temperature fields. Due to the large number of combinations for which the phase-field simulations were performed, only representative results are presented here. Figure 23 presents the dendritic structure formation due to solidification for input 3 (154.16 µm, 1.17 m/s, 72.86 W) at spatial locations 3, 3, and 1 and for times 266.7µs, 444.4µs, and 800µs, respectively (also refer to Figure 22). The corresponding thermal gradient (G) and velocity (V) ranges are -74.86–49.79 K/m and 0.3–1.15 m/s based on the multiphysics simulation. Other non-dimensionalized parameters for the phase-field simulations are provided in Table 3. A simplified parameters are used here to eliminate the dependency on the phase-field solve while comparing the outputs from different temperature fields. More details about the phase-field model can be found in [46]. Table 3. Non-dimensionalized simulation parameters for the phase-field solidification simulation Parameters Value Molar volume (Vm ) 1.0 Curvature of the parabola for solid (ks ) 10.0 Curvature of the parabola for liquid (kl ) 10.0 Concentration at minimum solid energy (cs,min ) 0.1 Concentration at minimum liquid energy (cl,min ) 0.9 Interfacial energy coefficient (γsl ) 1.5 Bulk energy coefficient (m) 10.0 Average gradient energy coefficient (κ0 ) 0.02 Diffusion coefficient (D) 1.0 Mobility (L) 33.33 Figures 23(a) and 23(b) show greater solid growth in the x- and y-directions, respectively. In contrast, Figure 23(c) shows equiaxed solid growth. Generally, the UNet model for temperature prediction is resulting in a very good accuracy when predicting the solid growth compared to the multiphysics simulation. The L- MOGPAR model performs well in the x-direction skewed and equiaxed solid growth cases but underpredicts the growth in the y-direction skewed case (i.e., Figure 23(b)). This is due to its poorer prediction of the temperature gradient. Figures 24a and 24b present the solid growth in the x- and y-direction for process inputs 2 (138.85 µm, 1.02 m/s, 68.27 W) and 3 (154.16 µm, 1.17 m/s, 72.86 W), respectively, for the process modeling simulation time of 444.4 µs and at spatial point 3 shown in Figure 22. The solid growth is compared using the process modeling outputs (i.e., temperature gradient and cooling rate) obtained from the multiphysics simulation, UNet, and L-MOGPAR models. Note that the simulation time in this figure refers to the phase-field simulation time. It is seen that for both directions, the solid growth results made using the UNet predictions of the process conditions match closely with those obtained from the multiphysics simulations. Although the solid growth predictions for the L-MOGPAR case match well in the x-direction 29 Simulation MOGPAR UNet (a) (b) (c) Figure 23. Dendritic structure formation for input 3 (154.16 µm, 1.17 m/s, 72.86 W) at spatial points (a) 3 (time 266.7 µs), (b) 3 (time 444.4 µs), and (c) 1 (time 800 µs). The definitions of the spatial points are presented in Figure 22. These were chosen from several combinations to be the representative spatial points and times for the results presented herein. case (Figure 24a), they deviate in the y-direction case (Figure 24b). To provide a more comprehensive picture, Table 4 presents the maximum absolute percentage errors across the phase-field simulation time. The average and maximum values of these error values across the three process modeling times and three spatial locations are presented for the three process modeling inputs. In the x-direction, both models have low errors, but UNet has the lower average error across the inputs ( 1.48% vs. 1.64% for L-MOGPAR). L-MOGPAR does achieve a lower worst-case error on inputs 1 and 2 (3.529% and 2.131% vs. 5.12% and 2.476%), but it underperforms for input 3 (3.874% vs. 2.256%). In the y-direction, UNet clearly and consistently outperforms L-MOGPAR. Its averages (2.494%, 1.048%, 3.027%) and maxima (5.196%, 1.752%, 8.869%) are all much lower than L-MOGPAR’s (2.55%, 2.731%, 5.889% averages; 5.814%, 8.738%, 18.438% maxima). Both models achieve single-digit average errors, so either is a viable option for phase-field simulations. Previous researchers have explored using machine learning for the additive manufacturing of metals (for e.g., see [47–49]). But the surrogate modeling of full-field temperature trajectories with CNN-based UNets and GPs, and the performance of surrogate models for predicting solidification and microstructure formation, have not been studied before. It is valuable to predict the thermal response of the AM substrate and its effect on directional solidification, which impacts the final grain morphology of the AM materials. Analyzing melt-pool dynamics is a challenging task considering its multiphysics and multiscale nature. The developed coupled model in this work yielded detailed insight into the underlying mechanisms by simulating melt-pool energy transport, deformation, and solidification. However, this came at the price of increased computational resources. Other works neglected similar melt-pool resolution by considering a heat-conduction-only model [50, 51], which can reduce the accuracy in energy gradient and temperature history that is essential for high- fidelity solidification modeling. Therefore, in this work, we relied on data-driven workflows to accelerate parametric studies. The L-MOGPAR and UNet surrogate models offer a computationally-efficient and accurate analysis of the thermal histories for microstructure solidification evolution. From the results in Section 3.3, it can be inferred that UNet yields lower errors across the domain compared to the L-MOGPAR model. UNet’s errors are relatively lower and mostly localized near the dynamic melt-pool boundary, while the L-MOGPAR model exhibits larger error magnitudes and a broader error region. This is expected, as the L-MOGPAR regression assumes a smooth kernel in its covariance function, and this can smooth out localized extremes. UNet’s ability to preserve fine spatial details is due to its encoder-decoder architecture with skip connections, which allows 30 Solid: Simulation, Dashed: MOGPAR, Dotted: UNet Solid: Simulation, Dashed: MOGPAR, Dotted: UNet 25 Input [1.38e-4,1.02,68.27] 25 Input [1.38e-4,1.02,68.27] Input [1.54e-4,1.17,72.86] Input [1.54e-4,1.17,72.86] Solid growth y-direction Solid growth x-direction 20 20 15 15 10 10 5 5 4 6 8 10 4 6 8 10 Simulation time Simulation time (a) (b) Figure 24. Solid growth in the x- and y-direction for process inputs 2 (138.85 µm, 1.02 m/s, 68.27 W) and 3 (154.16 µm, 1.17 m/s, 72.86 W). These results are for a process modeling simulation time of 444.4 µs and spatial point 3 shown in Figure 22. Solid growth is compared with time predictions made using the process modeling inputs (i.e., temperature gradient and cooling rate) obtained using the multiphysics simulation, UNet, and L-MOGPAR models. The simulation time in this figure refers to the phase-field simulation time. Table 4. Maximum absolute percentage errors (lower is better) for solid growth in the x- and y-direction across the phase-field simulation times. The average and maximum error values across the three process modeling times and three spatial locations are presented for the three process modeling inputs: 1 (138.97 µm, 1.41 m/s, 74.11 W), 2 (138.85 µm, 1.02 m/s, 68.27 W), and 3 (154.16 µm, 1.17 m/s, 72.86 W). (a) Solid growth x-direction UNet L-MOGPAR Input Avg Max Avg Max 1 1.812,857,48 5.119,880,94 1.785,512,63 3.528,869,07 2 1.332,360,3 2.476,041,26 1.295,243,11 2.131,036,64 3 1.290,943,39 2.256,371,18 1.829,348,05 3.874,456,54 (b) Solid growth y-direction UNet L-MOGPAR Input Avg Max Avg Max 1 2.494,422,35 5.195,978,6 2.550,245,38 5.813,821,41 2 1.047,677,99 1.751,822,62 2.731,353,7 8.738,024,94 3 3.027,190,52 8.869,450,23 5.889,294,22 18.437,887,04 it to better reconstruct the steep temperature gradients at the melt-pool boundary. The checkerboard artifacts [52] observed in Figure 11 are caused by transpose convolution due to uneven overlap, stacked over the multiple upsampling blocks. A bilinear interpolation upsampling layer at the end smoothens and helps reduce 31 this issue [53]. To a certain extent, UNet can better preserve details and spatial consistency, which is crucial for preserving accurate spatial information in flow field prediction tasks. Qualitatively, UNet reproduces the melt-pool shape and size more accurately, whereas L-MOGPAR’s predictions are more conservative at the melt-pool edges. These trends are consistent with the findings in the literature that CNN-based surrogates excel at capturing high-resolution features in thermal fields [54]. When discussing the trade-offs and considering which machine learning approach to deploy in production, it is important to not only evaluate the model’s performance, as measured by standard metrics, but also other properties, such as model interpretability [55] and robust UQ [56]. While deep learning models offer exceptional performance, they are difficult to interpret [57]. On the other hand, L-MOGPAR leverages the flexibility and interpretability of GPs while capturing structure across outputs [58, 59]. Additionaly, the robust UQ of deep learning models such as UNet’s is often computationally expensive and can lead to misrepresentations of predictive uncertainties [60]. On the other hand, the standard deviations inherently obtained from the L-MOGPAR model provide a measure of uncertainty in the model’s prediction. When the standard deviation is high, it indicates that the model is less confident about its prediction, which often leads to regions where the absolute error of the mean prediction is also high. The spatial distribution of the model’s standard deviation stems from the truncated PC components used in the multi-output GP and/or test data points that are outliers compared to the training data. The improvement in the standard deviations decreases and saturates with the number of PCs used to train the L-MOGPAR model, hence the standard deviation contribution is likely only from the distance of a test data point from the training cloud. Upon training and validating the L-MOGPAR against an increasing number of PCs, we concluded that 25 PCs are sufficient to capture about 99% of the explained variance without compromising the training time of the surrogate. Hence, by leveraging the uncertainty estimates, further experiments or simulations can be set up using active learning to improve the confidence in the test predictions by reducing the epistemic uncertainty. Linking the machine learning models’ accuracy in predicting melt-pool dynamics, or the lack thereof, to microstructural evolution is a unique aspect of this work. This enabled us to evaluate the influence of process conditions on the formation of the solidification rate and subgrain morphologies in the additively manufactured material. The melt-pool predictions of both the L-MOGPAR and UNet models resulted in highly satisfactory solidification rates. Although the L-MOGPAR model has higher percentage errors for predictions under certain process inputs and for certain spatiotemporal locations, on average either the L-MOGPAR or UNet model is viable for the solidification simulations. 4.3.2. Chemical segregation prediction Here, we present the segregation prediction from the alloy solidification model in conjunction with the process model. Similar to what has been shown in Section 4.3.1, we simulate multiple simulations with varying conditions. Here, we present the segregation behavior corresponding to the thermal conditions of Point 1 at 266.7 µs. This is selected for the demonstration as this point has the maximum cooling rate, among the conditions simulated. Figure 25a demonstrates the solidification via growth of a planar solid-liquid interface. It is observed that the solidification rate is faster during the initial stage of solidification and it reaches a steady-state growth rate after some time. The solidification rate for the three cases vary due to variation in the cooling rate resulting from the variations in the process conditions. Rate of solidification increases with increasing cooling rate. Figure 25b shows the position of the solid-liquid interface at simulation time t = 50. Figure 25c and Figure 25d shows the corresponding variation in Cr and Ni concentration around the interface. It is observed that both Ni and Cr segregation occurs ahead of the solid-liquid interface and a concentration depletion is observed in the solid side of the interface. This primarily occurs due to the combined effect of solute drag and solute rejection. Higher level of segregation values observed in these simulations compared to the experimental observations. This could be attributed to the initial supersaturation assumption in the liquid. Further analysis and model validation will be pursued in the future. 32 Case I 1.0 Case I Case II Case II 50 Case III Case III 0.8 Chnage in Solid Fraction (%) 40 Solid Order Parameter 0.6 30 0.4 20 10 0.2 0 0.0 0 200 400 600 800 1000 0 2 4 6 8 10 Simulation Time Position (a) (b) 0.50 0.25 Case I Case I Case II Case II 0.45 Case III Case III 0.20 0.40 Cr Concentration Ni Concentration 0.15 0.35 0.30 0.10 0.25 0.05 0.20 0.00 0.15 0.10 −4 −2 0 2 4 6 8 10 −4 −2 0 2 4 6 8 10 Position Position (c) (d) Figure 25. (a) Solid growth over time, (b) Interface location at simulation time t = 50, (c) Cr concentration, and (d) Ni concentration across the solid-liquid interface. Case I, Case II and Case III corresponds to process inputs 1 (138.97 μm, 1.41 m/s, 74.11 W), 2 (138.85 µm, 1.02 m/s, 68.27 W) and 3 (154.16 µm, 1.17 m/s, 72.86 W), respectively. These results are for the conditions obtained from a process modeling simulation time of 266.7 µs and spatial point 1 shown in Figure 22. 33 5. CONCLUSION In conclusion, we investigated the microstructural evolution of 316 SS during the LPBF process in connection to the process conditions using physics-based and machine learning models. we developed and evaluated a probabilistic, multi-output Gaussian-process autoregressive (L-MOGPAR) surrogate for modeling spatiotemporal temperature and melt-pool dynamics in laser powder bed fusion (LPBF), and contrasted it with a fast UNet baseline. Both models were trained on the same high-fidelity Arbitrary Lagrangian Eulerian–finite element method Navier–Stokes plus enthalpy data. We further propagated surrogate predictions to a grand- potential phase-field model, enabling an end-to-end assessment of how process-level prediction accuracies affect microstructure evolution. We demonstrated the solidification behavior and quantified associated micro-segregation for different process conditions. Such segregation behavior is crucial for evaluating creep and irradiation performance of the AM materials. While comparing the point accuracy of the surrogate models, UNet achieved higher coefficients of determination on temperature and spatial gradients (RT2 = 0.986, R∂2x T = 0.918, R∂2y T = 0.938) than L-MOGPAR (RT2 = 0.931, R∂2x T = 0.746, R∂2y T = 0.85). However, L-MOGPAR provides calibrated predictive uncertainty over space and time, revealing where epistemic uncertainty is large (e.g., in high- gradient regions near the melt-pool boundary) and where predictions are reliable. These spatially resolved uncertainty estimates correlated well with the spatial errors. When the machine learning predictions of the temperature and melt-pool dynamics were passed to the phase-field model on solidification, both models satisfactorily captured the dendritic structure formation. The predictions propagated well to the chemical microsregregation predicted at the solid-liquid interface of 316 SS. Quantitatively, the segregtaion values obtained from the simulations were higher that what has been observed in experiments. More one-to-one validation is needed to improve this further. The model predictions can be improved further by generalizing the physics choices (laser/path space, boundary conditions, and selected surface forces) to broader process windows (e.g., multitrack/multilayer), to add additional surface physics (e.g., thermocapillary shear) and incorporate experimental measurements. Additional physics, process windows, and experimental measurements would further strengthen the viability of uncertainty-informed machine learning models for advanced manufacturing (AM) process predictions. Overall, our results indicate that risk-aware process decisions benefit from probabilistic surrogates: even when a deterministic model is slightly more accurate on average, quantified uncertainty is pivotal for quantifying robustness. The phase-field model can be easily extended to capture segregation on phase boundaries and grain boundaries due to its multi-phase multi-order parameter nature. In summary, we present the computational framework necessary for capturing different aspects of the AM process encompassing different length-scales and associated microstructural evolution at the subgrain level. This effort provides a strong foundation for reaching the overall goal of establishing a reduced-order process-structure-property-performance (PSPP) correlation for AM materials in a cost-effective way and thus supporting AMMT’s rapid qualification goals. This work will also provide crucial physics-informed information for the larger length scale process models developed by the AMMT team at Oak Ridge National Laboratory. The generated models can also be used within an active learning framework to identify data need for accelerated material development and qualification for AMMT program. 34 [1] C. Sun, Y. Wang, M. D. McMurtrey, N. D. Jerred, F. Liou, J. Li, Additive manufacturing for energy: A review, Applied Energy 282A (116041). [2] M. McMurtrey, R. O’Brien, C. Sun, C.-H. Shiau, F. Teng, Irradiation damage and iascc of printed 316l for use as fuel cladding, in: 19th International Conference on Environmental Degradation of Materials in Nuclear Power Systems - Water Reactors, Boston, MA, USA, 2019. [3] M. McMurtrey, C. Sun, R. Rupp, C.-H. Shiau, R. Hanbury, N. Jerred, R. O’Brien, Investigation of the irradiation effects in additively manufactured 316l steel resulting in decreased irradiation assisted stress corrosion cracking susceptibility, Journal of Nuclear Materials 545 (2021) 152739. [4] M. Song, M. Wang, X. Lou, R. Rebak, G. Was, Radiation damage and irradiation-assisted stress corrosion cracking of additively manufactured 316l stainless steels, Journal of Nuclear Materials 513 (2019) 33–44. [5] M. McMurtrey, M. Messner, Qualification challenges for additive manufacturing in high temperature nuclear applications, in: the ASME 2021 Pressure Vessels & Piping Conference, PVP 2021, 2021. [6] M. Bayant, 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 (102278). [7] M. M. Francois, A. Sun, et al., Modeling of additive manufacturing processes for metals: Challenges and opportunities, Current Opinion in Solid State & Materials Science 21 (2017) 198–206. [8] A. Lindsay, R. Stogner, D. Gaston, D. Schwen, C. Matthews, W. Jiang, L. K. Aagesen, R. Carlsen, F. Kong, A. Slaughter, et al., Automatic differentiation in metaphysicl and its applications in moose, Nuclear Technology 207 (7) (2021) 905–922. [9] A. Lindsay, G. Giudicelli, P. German, J. Peterson, Y. Wang, R. Freile, D. Andrs, P. Balestra, M. Tano, R. Hu, et al., Moose navier–stokes module, SoftwareX 23 (2023) 101503. [10] G. Giudicelli, A. Lindsay, L. Harbour, C. Icenhour, M. Li, J. E. Hansel, P. German, P. Behne, O. Marin, R. H. Stogner, J. M. Miller, D. Schwen, Y. Wang, L. Munday, S. Schunert, B. W. Spencer, D. Yushu, A. Recuero, Z. M. Prince, M. Nezdyur, T. Hu, Y. Miao, Y. S. Jung, C. Matthews, A. Novak, B. Langley, T. Truster, N. Nobre, B. Alger, D. Andrš, F. Kong, R. Carlsen, A. E. Slaughter, J. W. Peterson, D. Gaston, C. Permann, 3.0 - MOOSE: Enabling massively parallel multiphysics simulations, SoftwareX 26 (2024) 101690. doi:https://doi.org/10.1016/j.softx.2024.101690. URL https://www.sciencedirect.com/science/article/pii/S235271102400061X [11] D. R. Noble, P. K. Notz, M. J. Martinez, A. M. Kraynik, Use of aria to simulate laser weld pool dynamics for neutron generator production., Tech. rep., Sandia National Laboratories (SNL), Albuquerque, NM, and Livermore, CA … (2007). [12] X. Chen, W. Mu, X. Xu, W. Liu, L. Huang, H. Li, Numerical analysis of double track formation for selective laser melting of 316l stainless steel, Applied Physics A 127 (2021) 1–13. [13] L. Caprio, A. G. Demir, B. Previtali, Influence of pulsed and continuous wave emission on melting efficiency in selective laser melting, Journal of Materials Processing Technology 266 (2019) 429–441. 35 [14] X. Zhang, S. A. Mantri, G. I. Vukovic, J. Listwan, D. Rink, E. Listwan, Development of process parameters and post-build conditions for qualification of lpbf 316 ss, Tech. rep., Argonne National Laboratory (ANL), Argonne, IL (United States) (2023). [15] S. Biswas, S. L. Dhulipala, P. German, A. M. Jokisaari, D. Yushu, M. D. McMurtrey, Multiscale and machine learning modeling for process-informed microstructure prediction in additively manufactured materials using malamute, Tech. rep., Idaho National Laboratory (INL), Idaho Falls, ID (United States) (2024). [16] J. Lamb, R. Ochoa, A. Eres-Castellanos, J. Klemm-Toole, M. P. Echlin, T. Sun, K. Fezzaa, A. Clarke, T. M. Pollock, Quantification of melt pool dynamics and microstructure during simulated additive manufacturing, Scripta Materialia 245 (2024) 116036. doi:https://doi.org/10.1016/j. scriptamat.2024.116036. URL https://www.sciencedirect.com/science/article/pii/S135964622400071X [17] R. Sharma, Y. B. Guo, M. Raissi, W. G. Guo, Physics-informed machine learning of argon gas-driven melt pool dynamics, Journal of Manufacturing Science and Engineering 146 (8) (2024) 081008. arXiv: https://asmedigitalcollection.asme.org/manufacturingscience/article-pdf/146/ 8/081008/7342881/manu\_146\_8\_081008.pdf, doi:10.1115/1.4065457. URL https://doi.org/10.1115/1.4065457 [18] A. N. Brooks, T. J. Hughes, Streamline upwind/petrov-galerkin formulations for convection dominated flows with particular emphasis on the incompressible navier-stokes equations, Computer methods in applied mechanics and engineering 32 (1-3) (1982) 199–259. [19] G. Giudicelli, A. Lindsay, L. Harbour, C. Icenhour, M. Li, J. E. Hansel, P. German, P. Behne, O. Marin, R. H. Stogner, et al., 3.0-moose: Enabling massively parallel multiphysics simulations, SoftwareX 26 (2024) 101690. [20] M. A. Álvarez, L. Rosasco, N. D. Lawrence, Kernels for vector-valued functions: A review, Foundations and Trends® in Machine Learning 4 (3) (2012) 195–266. doi:10.1561/2200000036. URL http://dx.doi.org/10.1561/2200000036 [21] V. Feinberg, L.-F. Cheng, K. Li, B. E. Engelhardt, Large linear multi-output gaussian process learning (2017). arXiv:1705.10813. [22] L.-F. Cheng, B. Dumitrascu, G. Darnell, C. Chivers, M. Draugelis, K. Li, B. E. Engelhardt, Sparse multi-output gaussian processes for online medical time series prediction, BMC medical informatics and decision making 20 (2020) 1–23. [23] S. Wold, K. Esbensen, P. Geladi, Principal component analysis, Chemometrics and intelligent laboratory systems 2 (1-3) (1987) 37–52. [24] P. Goovaerts, Geostatistics for natural resources evaluation, Oxford university press, 1997. [25] Y. W. Teh, M. Seeger, M. I. Jordan, Semiparametric latent factor models, in: International Workshop on Artificial Intelligence and Statistics, PMLR, 2005, pp. 333–340. [26] J. Lloyd, D. Duvenaud, R. Grosse, J. Tenenbaum, Z. Ghahramani, Automatic construction and natural- language description of nonparametric regression models, in: Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 28, 2014. 36 [27] Z. Dai, M. Álvarez, N. Lawrence, Efficient modeling of latent information in supervised learning using gaussian processes, Advances in Neural Information Processing Systems 30. [28] J. Requeima, W. Tebbutt, W. Bruinsma, R. E. Turner, The Gaussian process autoregressive regression model (GPAR), 22nd International Conference on Artificial Intelligence and Statistics (2019) 1860–1869. [29] D. P. Kingma, J. Ba, Adam: A method for stochastic optimization, arXiv preprint arXiv:1412.6980. [30] D. Ulyanov, A. Vedaldi, V. Lempitsky, Instance normalization: The missing ingredient for fast stylization (2017). arXiv:1607.08022. URL https://arxiv.org/abs/1607.08022 [31] V. Nair, G. E. Hinton, Rectified linear units improve restricted boltzmann machines, in: Proceedings of the 27th international conference on machine learning (ICML-10), 2010, pp. 807–814. [32] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, Z. Wojna, Rethinking the inception architecture for computer vision, in: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 2818–2826. doi:10.1109/CVPR.2016.308. [33] M. D. Zeiler, ADADELTA: an adaptive learning rate method, CoRR abs/1212.5701. arXiv:1212. 5701. URL http://arxiv.org/abs/1212.5701 [34] D. Bertsimas, D. B. Brown, C. Caramanis, Theory and applications of robust optimization (2010). arXiv:1010.5445. URL https://arxiv.org/abs/1010.5445 [35] S. Biswas, D. Liu, L. K. Aagesen, W. Jiang, Solidification and grain formation in alloys: a 2d application of the grand-potential-based phase-field approach, Modelling Simul. Mater. Sci. Eng. 30. [36] M. Plapp, Unified derivation of phase-field models for alloy solidification from a grand-potential functional, Physical Review E 84 (3) (2011) 031601. URL ://WOS:000294569000002 [37] L. K. Aagesen, Y. Gao, D. Schwen, K. Ahmed, Grand-potential-based phase-field model for multiple phases, grains, and chemical components, Physical Review E 98 (2018) 023309. [38] A. Dindale, Sgte data for pure elements, Calphad 15 (1991) 317–425. [39] J. Miettinen, Thermodynamic reassessment of fe-cr-ni system with emphasis on the iron-rich corner., Calphad 23 (1999) 231–248. [40] N. Moelans, A quantitative and thermodynamically consistent phase-field interpolation function for multi-phase systems, Acta Materialia 59 (3) (2011) 1077–1086. [41] K. N. Shampur, A grand potential based multi-phase field model for alloy solidification, Thesis, McGill University (2017). [42] A. Choudhury, M. Kellner, b. Nestler, A method for coupling the phase-field model based on a grand- potential formalism to thermodynamic databases, Current Opinion in Solid State and Materials Science 19. 37 [43] S. Biswas, S. L. N. Dhulipala, P. German, A. M. Jokisaari, D. Yushu, M. D. McMurtrey, Multiscale and machine learning modeling for process-informed microstructure prediction in additively manufactured materials using malamute, Tech. rep., Idaho National Laboratory (2024). [44] A. E. Slaughter, Z. M. Prince, P. German, I. Halvic, W. Jiang, B. W. Spencer, S. L. Dhulipala, D. R. Gaston, Moose stochastic tools: A module for performing parallel, memory-efficient in situ stochastic simulations, SoftwareX 22 (2023) 101345. [45] D. R. Gaston, C. J. Permann, J. W. Peterson, A. E. Slaughter, D. Andrš, Y. Wang, M. P. Short, D. M. Perez, M. R. Tonks, J. Ortensi, L. Zou, R. C. Martineau, Physics-based multiscale coupling for full core nuclear reactor simulation, Annals of Nuclear Energy 84 (2015) 45–54. [46] S. Biswas, D. Liu, L. K. Aagesen, W. Jiang, Solidification and grain formation in alloys: a 2d application of the grand-potential-based phase-field approach, Modelling and Simulation in Materials Science and Engineering 30 (2) (2022) 025013. [47] H. Safari, H. Wessels, Physics-informed surrogates for temperature prediction of multi-tracks in laser powder bed fusion, arXiv preprint arXiv:2502.01820. [48] X. Zhao, Z. Gong, Y. Zhang, W. Yao, X. Chen, Physics-informed convolutional neural networks for temperature field prediction of heat source layout without labeled data, Engineering Applications of Artificial Intelligence 117 (2023) 105516. doi:10.1016/j.engappai.2022.105516. [49] R. Raut, A. K. Ball, A. Basak, Scalable and transferable graph neural networks for predicting temperature evolution in laser powder bed fusion, Engineering Applications of Artificial Intelligence 153 (2025) 110898. doi:10.1016/j.engappai.2025.110898. [50] 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. [51] 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. [52] A. Odena, V. Dumoulin, C. Olah, Deconvolution and checkerboard artifacts, Distilldoi:10.23915/ distill.00003. URL http://distill.pub/2016/deconv-checkerboard [53] X. Wang, Interpolation and sharpening for image upsampling, in: 2022 2nd International Conference on Computer Graphics, Image and Virtualization (ICCGIV), IEEE, 2022, pp. 73–77. [54] O. Ronneberger, P. Fischer, T. Brox, U-net: Convolutional networks for biomedical image segmentation, in: International Conference on Medical image computing and computer-assisted intervention, Springer, 2015, pp. 234–241. [55] C. Huyen, Designing machine learning systems, ” O’Reilly Media, Inc.”, 2022. [56] V. Kuleshov, N. Fenner, S. Ermon, Accurate uncertainties for deep learning using calibrated regression, International conference on machine learning (2018) 2796–2804. [57] A. Radford, L. Metz, S. Chintala, Unsupervised representation learning with deep convolutional generative adversarial networks, arXiv preprint arXiv:1511.06434. 38 [58] W. Bruinsma, E. Perim, W. Tebbutt, S. Hosking, A. Solin, R. Turner, Scalable exact inference in multi-output Gaussian processes, in: H. D. III, A. Singh (Eds.), Proceedings of the 37th International Conference on Machine Learning, Vol. 119 of Proceedings of Machine Learning Research, PMLR, 2020, pp. 1190–1201. URL https://proceedings.mlr.press/v119/bruinsma20a.html [59] J. Chen, W. Mu, Y. Li, D. Li, On the identifiability and interpretability of gaussian process models, in: A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, S. Levine (Eds.), Advances in Neural Information Processing Systems, Vol. 36, Curran Associates, Inc., 2023, pp. 70267–70278. URL https://proceedings.neurips.cc/paper_files/paper/2023/file/ dea2b4f9012686bcc1f59a62bcd28158-Paper-Conference.pdf [60] G. D. Pasparakis, L. Graham-Brady, M. D. Shields, Bayesian neural networks for predicting uncertainty in full-field material response, Computer Methods in Applied Mechanics and Engineering 433 (2025) 117486. doi:10.1016/j.cma.2024.117486. 39