Computational, Data-Driven, and Physics-Informed Machine Learning Approaches for Microstructure Modeling in Metal Additive Manufacturing

Abstract: Metal additive manufacturing enables unprecedented design freedom and the production of customized, complex components. However, the rapid melting and solidification dynamics inherent to metal AM processes generate heterogeneous, non-equilibrium microstructures that significantly impact mechanical properties and subsequent functionality. Predicting microstructure and its evolution across spatial and temporal scales remains a central challenge for process optimization and defect mitigation. While conventional experimental techniques and physics-based simulations provide a physical foundation and valuable insights, they face critical limitations. In contrast, data-driven machine learning offers an alternative prediction approach and powerful pattern recognition but often operate as black-box, lacking generalizability and physical consistency. To overcome these limitations, physics-informed machine learning, including physics-informed neural networks, has emerged as a promising paradigm by embedding governing physical laws into neural network architectures, thereby enhancing accuracy, transparency, data efficiency, and extrapolation capabilities. This work presents a comprehensive evaluation of modeling strategies for microstructure prediction in metal AM. The strengths and limitations of experimental, computational, and data-driven methods are analyzed in depth, and highlight recent advances in hybrid PIML frameworks that integrate physical knowledge with ML. Key challenges, such as data scarcity, multi-scale coupling, and uncertainty quantification, are discussed alongside future directions. Ultimately, this assessment underscores the importance of PIML-based hybrid approaches in enabling predictive, scalable, and physically consistent microstructure modeling for site-specific, microstructure-aware process control and the reliable production of high-performance AM components.

• The paper highlights Physics-Informed Machine Learning (PIML), particularly PINNs, as a method to improve microstructure modeling in metal additive manufacturing by integrating physical laws into ML, offering enhanced accuracy and data efficiency compared to traditional data-driven or physics-based simulations.
• Data-driven methods excel with large datasets but often lack physical interpretability, while conventional physics-based simulations are computationally expensive and face scalability challenges in complex AM processes.
• PIML approaches, despite challenges in training stability and generalization, hold promise for real-time adaptive process control and improved predictive capabilities in AM through ongoing research in scalability, uncertainty quantification, and integration with advanced computing.

Physics-Informed Machine Learning Approaches in Metal Additive Manufacturing Microstructure Modeling

Metal additive manufacturing (AM) has emerged as a versatile technology capable of producing complex components with tailored geometries, which are not feasible through conventional manufacturing methods. However, the inherent rapid melting and solidification in metal AM processes often result in intricate non-equilibrium microstructures, posing a significant challenge to the prediction and control of their evolution across spatial and temporal scales. Conventional approaches like experimental techniques and physics-based simulations, while informative, are typically limited by high computational costs and scalability challenges. This paper presents physics-informed machine learning (PIML) as a potential solution to these shortcomings by embedding physical laws within neural networks, notably physics-informed neural networks (PINNs), offering enhanced accuracy, transparency, data efficiency, and generalization abilities.

Microstructure Modeling in Metal AM

The work underscores the importance of accurate microstructure prediction for process optimization and defect mitigation in metal AM, as heterogeneous microstructures directly affect mechanical properties. Traditional experimental characterization techniques, such as Scanning Electron Microscopy (SEM), Electron Backscatter Diffraction (EBSD), and X-ray Computed Tomography (XCT), provide insights into grain morphology, crystallographic texture, and phase distribution. Nevertheless, these methods often struggle to monitor microstructures in situ due to the rapid processing conditions of AM. Computational approaches like Finite Element Analysis (FEA), Phase-Field (PF) modeling, Cellular Automata (CA), and Kinetic Monte Carlo (KMC) simulations offer detailed predictions of microstructural evolution, but face scalability and computational efficiency issues.

Data-Driven Methods in Microstructure Prediction

Data-driven models, exploiting ML and deep learning (DL) architectures, present a promising alternative by leveraging large datasets. These models, including Convolutional Neural Networks (CNNs), Graph Neural Networks (GNNs), and Long Short-Term Memory networks (LSTMs), have shown success in predicting complex process-microstructure relationships. However, they often lack physical interpretability and exhibit "black-box" behavior, especially when extrapolating beyond the training domain. Hybrid approaches integrating physical knowledge improve robustness and cross-process applicability, as demonstrated through ML-based digital twin frameworks predicting melt pool behavior, grain size, and defect formation in metal AM parts.

PIML: Bridging Physics and Data Science

Physics-informed machine learning, represented mainly by PINNs, aims to bridge the gap between physics-based models and data-driven predictions. By incorporating governing physical laws into ML architectures, these models effectively enhance accuracy even in data-scare conditions, while reducing reliance on extensive datasets. PINNs offer mesh-free solutions to predict temperature fields, melt pool dynamics, and other microstructural features, achieving high precision without labeled data when provided boundary conditions are well-defined. Furthermore, PIML frameworks optimize process control by dynamically updating predictions with real-time sensor data, enabling adaptive manufacturing systems.

Challenges and Future Directions

While promising, PIML approaches face challenges in training stability, generalization between different materials and processes, and maintaining computational efficiency. Current development efforts focus on improving model scalability through techniques like domain decomposition and adaptive sampling, and refining UQ to ensure robust predictions. Enhancements like transformer-based PINNs, lightweight attention mechanisms, and integration with high-performance computing are anticipated to improve prediction capabilities and unlock real-time adaptive control possibilities in AM.

Ultimately, the paper highlights that the progression of PIML frameworks will benefit from interdisciplinary collaboration. The convergence of experimental insights, simulation advancements, and AI-driven techniques will catalyze the evolution of intelligent, predictive AM systems equipped for reliably producing high-performance components with application-specific microstructures.