Enabled Autonomous Manufacturing
Wei Chen¹, Vispi Nevile Karkaria1, Yi-Ping Chen¹ and Ying-Kuan Tsai,1
1 Department of Mechanical Engineering, Northwestern University, Evanston, IL, USA
E-mail: weichen@northwestern.edu
Status
Digital twins are beginning to transform manufacturing beyond early “digital shadows,” which only mirror system states for monitoring, by enabling bidirectional feedback for real-time predictive control and decision-making [1,2]. According to the National Academies report [3], a digital twin is “a set of virtual information constructs that mimics the structure, context, and behavior of a physical system or system-of- systems, is dynamically updated with data from its physical counterpart, has predictive capability, and informs decisions that realize value.” Figure 1 illustrates our digital twin framework, where physics-based
modeling and simulation build a machine learning surrogate for model predictive control (MPC), enabling bidirectional feedback with the physical system; this approach has been demonstrated in directed energy deposition (DED) using a Time Series Dense Encoder (TiDE) to predict melt pool features and optimize laser power for improved quality and reduced defects [4].
Figure 1. Digital twin framework for manufacturing systems
Despite these recent advances, most deployed digital twins in manufacturing still fall short of being fully autonomous, and need human experts in the loop to validate model updates and resolve edge-case decisions, which are rare or unusual scenarios outside normal operating conditions [5]. Looking ahead, the status quo in manufacturing digital twins must evolve from digital shadows to self-optimizing agents that not only forecast failures and optimize process but also learn evolving process dynamics, adapt to new products, and collaborate across multiple manufacturing processes, laying the foundation for the next generation of digital twins that drive lean, resilient, and sustainable manufacturing ecosystems [6].
Current and future challenges
Modern manufacturing processes such as additive manufacturing and precision welding involve complex spatio-temporal physics that are difficult to model in real time. While high-fidelity methods like finite element or computational fluid dynamics can capture these dynamics, they are too computationally expensive for closed-loop control, making fast machine learning surrogates essential [7]. By extracting patterns from sensor and simulation data, these surrogates can deliver millisecond-scale predictions on edge devices, enabling ultra-low latency inference. However, most digital twins still model only temporal or spatial dynamics in isolation, and many treat processes as black boxes rather than learning the underlying physical laws, creating challenges in both physics discovery from sparse data and generalization to new conditions [8]. These difficulties are amplified in high-cost processes where data are scarce, especially for rare defects, limiting calibration and adaptation. While offline retraining can restore accuracy, continuous online updating without disrupting production remains challenging, with methods such as Koopman operators and reduced-order models offering potential but requiring further research [9].
Transitioning from surrogate modeling to real-time decision making in the digital twin paradigm requires uncertainty-aware machine learning and decisions throughout the system’s lifecycle, which often involves a
tradeoff between model fidelity and computational efficiency. For example, when implementing MPC with discrete event simulation [10], the choice of model and optimization method, such as quadratic programming with simplified linear approximation or numerical optimization with a detailed physics-based model, strongly affects accuracy, decision frequency, and the overall quality of control. Moreover, as system behavior may change under different conditions or at various stages of life cycles, effective and data-efficient methods for continuous model adaptation remain an unsolved challenge.
The challenges pointed out so far highlight the importance of continuous verification, validation, and uncertainty quantification (VVUQ) which is still a hurdle in the deployment of digital twins for manufacturing [3,6]. While standards such as ISO 23247 (digital twin framework for manufacturing) offer structural guidance, they do not address credibility assessment or VVUQ. Digital twins are inherently complex and face epistemic, aleatoric, and even unknown uncertainties, requiring robust VVUQ to ensure reliability. The dynamic nature of digital twins requires ongoing validation to remain accurate as conditions evolve, raising questions about update frequency and computational feasibility.
Microstructure plays a critical role in manufacturing, as its evolution depends on geometry and process conditions and ultimately governs the mechanical, thermal, and chemical properties that determine component performance. Understanding the process–microstructure–property–performance relationships is essential for optimization [11], yet predicting part-scale evolution remains challenging due to the high cost of simulations, stochastic physical phenomena, and limited in-situ monitoring in high-speed or high-temperature processes. In additive manufacturing, for example, rapid thermal cycling produces nonlinear, path-dependent grain structures that are difficult to model [12]. These complexities also impact the co-design of geometry, materials, and processes, where microstructural heterogeneity complicates optimization. Addressing this challenge requires frameworks that incorporate microstructure-related objectives into early design decisions and refine them through adaptive process optimization.
Advances in science and technology to meet challenges
To model dynamic manufacturing processes that evolve across time and space, machine learning methods such as spatio-temporal neural operators, graph neural networks, and transformer encoders have been used to extract patterns from sensor data, simulations, and historical logs. Heterogeneous data fusion has been addressed with approaches like latent variable Gaussian processes (LVGP) and Temporal Fusion Transformers [13], while adaptive sampling, multi-fidelity modeling, transfer learning, and synthetic data pre-training help mitigate data scarcity [14]. Uncertainty is handled through models such as Gaussian processes, Bayesian neural networks, or deep ensembles, which provide calibrated confidence intervals for risk-aware decision making. To meet performance demands, lightweight architectures and optimized edge deployment enable fast, cost-efficient inference. Building on these advances, pre-trained neural networks now act as surrogates for rapid evaluation, support efficient policy learning to map states to actions, and generate probabilistic forecasts that guide uncertainty-aware, real-time decision-making [15].
To ensure that surrogate models remain reliable throughout the system lifecycle, recent research has focused on continuous model validation [16] and adaptation. This involves not only detecting when a model no longer aligns with the physical systems but also updating the model efficiently with least compromised performance. Emerging approaches for concept drift detection have improved the robustness of real-time monitoring, especially for neural networks and multivariate predictive frameworks [17]. Concurrently, lightweight adaptation methods such as low-rank fine-tuning enable rapid updates with minimal data, preserving fidelity while adapting to evolving manufacturing conditions. These developments are critical for
enabling trustworthy, self-improving digital twin systems that support resilient and adaptive decision-making across dynamic operational environments.
Recent advances in AI and ML are transforming how microstructure is predicted and optimized in manufacturing to address the impracticality of extensive experimentation and the high computational cost of physics-based modeling [18]. Multi-scale modeling approaches now couple macro-level processing conditions with micro-and meso-scale material behaviors, allowing for more accurate prediction of microstructural evolution and its effect on material performance. These models are increasingly being enhanced by data-driven techniques, such as Generative Adversarial Networks (GANs) that generate realistic microstructures for virtual design [19], and convolutional neural networks (CNNs) that predict stress–strain curves and material properties from microstructural images; and recurrent neural networks (RNNs) capture history-dependent plasticity by learning deformation-path-sensitive responses [12]. These capabilities are embedded in integrated optimization frameworks that use Bayesian optimization, evolutionary algorithms, and transfer learning to discover process conditions yielding optimal microstructures and properties.
By extending these methods to co-design of geometry, materials, and processes, deep learning and inverse optimization can account for heterogeneous microstructural evolution [20]. Although co-design is high-dimensional, differentiable physics-based modeling such as JAX-FEM, enables gradient-based optimization that directly links design variables to microstructural outcomes, accelerating the search for optimal solutions while preserving physical fidelity.
Concluding remarks
In summary, advancing manufacturing digital twins requires addressing key needs such as spatio-temporal modeling, surrogate-based real-time control, uncertainty quantification, and microstructure-aware co-design. State-of-the-art physics-informed machine learning approaches, including spatio-temporal neural operators, graph neural networks, and transformer encoders, provide unified process modeling capabilities. Reliable deployment also depends on continuous model updating through fast surrogates, multimodal data fusion, adaptive sampling, transfer learning, concept drift detection, and multi-scale predictive frameworks. Looking ahead, the next generation of autonomous digital twins will incorporate cyber-secure federated learning, scalable data infrastructure, trustworthy AI, and edge-to-cloud integration, enabling self-learning and adaptive decision-making. These advances will support resilient, sustainable, and agile manufacturing with concurrent optimization of processes, materials, and designs.
Acknowledgements
We are grateful for the grant support from the National Science Foundation’s Engineering Research Center for Hybrid Autonomous Manufacturing: Moving from Evolution to Revolution (ERC-HAMMER), under the Award Number EEC-2133630. Grant support from the Remade institute research program (DE-EE0007897) is greatly acknowledged. Yi-Ping Chen also acknowledges the Taiwan-Northwestern Doctoral Scholarship.
References
[1] Karkaria, V., Goeckner, A., Zha, R., Chen, J., Zhang, J., Zhu, Q., Cao, J., Gao, R. X., and Chen, W., 2024, “Towards a Digital Twin Framework in Additive Manufacturing: Machine Learning and Bayesian Optimization for Time Series Process Optimization,” J. Manuf. Syst.
[2] van Beek, A., Nevile Karkaria, V., and Chen, W., 2023, “Digital Twins for the Designs of Systems: A Perspective,” Struct. Multidiscip. Optim., 66(3), p. 49.
[3] 2024, Foundational Research Gaps and Future Directions for Digital Twins, National Academies Press, Washington, D.C.
[4] Chen, Y.-P., Tsai, Y.-K., Karkaria, V., and Chen, W., 2025, “Uncertainty-Aware Digital Twins: Robust Model Predictive Control Using Time-Series Deep Quantile Learning,” J. Mech. Des., 148(021702).
[5] Zemskov, A. D., Fu, Y., Li, R., Wang, X., Karkaria, V., Tsai, Y.-K., Chen, W., Zhang, J., Gao, R., Cao, J., Loparo, K. A., and Li, P., 2024, “Security and Privacy of Digital Twins for Advanced Manufacturing: A Survey.”
[6] Shao, G., Hightower, J., and Schindel, W., 2023, “Credibility Consideration for Digital Twins in Manufacturing,” Manuf. Lett., 35, pp. 24–28.
[7] Karkaria, V., Tsai, Y.-K., Chen, Y.-P., and Chen, W., “An Optimization-Centric Review on Integrating Artificial Intelligence and Digital Twin Technologies in Manufacturing,” Eng. Optim., 0(0), pp. 1–47.
[8] Böttjer, T., Tola, D., Kakavandi, F., Wewer, C. R., Ramanujan, D., Gomes, C., Larsen, P. G., and Iosifidis, A., 2023, “A Review of Unit Level Digital Twin Applications in the Manufacturing Industry,” CIRP J. Manuf. Sci. Technol., 45, pp. 162–189.
[9] Kabashkin, I., 2024, “Digital Twin Framework for Aircraft Lifecycle Management Based on Data-Driven Models,” Mathematics, 12(19), p. 2979.
[10] Thelen, A., Zhang, X., Fink, O., Lu, Y., Ghosh, S., Youn, B. D., Todd, M. D., Mahadevan, S., Hu, C., and Hu, Z., 2022, “A Comprehensive Review of Digital Twin — Part 1: Modeling and Twinning Enabling Technologies,” Struct. Multidiscip. Optim., 65(12), p. 354.
[11] Kalia, K., and Ameli, A., 2023, “Understanding the Process-Microstructure-Property Relationships in Material Extrusion Additive Manufacturing of Polylactic Acid Microcellular Foams,” Addit. Manuf., 72, p. 103636.
[12] Mozaffar, M., Bostanabad, R., Chen, W., Ehmann, K., Cao, J., and Bessa, M. A., 2019, “Deep Learning Predicts Path-Dependent Plasticity,” Proc. Natl. Acad. Sci., 116(52), pp. 26414–26420.
[13] Comlek, Y., Ravi, S. K., Pandita, P., Ghosh, S., Wang, L., and Chen, W., “Heterogeneous Multi-Source Data Fusion Through Input Mapping and Latent Variable Gaussian Process.”
[14] Karkaria, V., Chen, J., Luey, C., Siuta, C., Lim, D., Radulescu, R., and Chen, W., 2024, “A Digital Twin Framework Utilizing Machine Learning for Robust Predictive Maintenance: Enhancing Tire Health Monitoring.”
[15] Chen, Y.-P., Karkaria, V., Tsai, Y.-K., Rolark, F., Quispe, D., Gao, R. X., Cao, J., and Chen, W., 2025, “Real-Time Decision-Making for Digital Twin in Additive Manufacturing with Model Predictive Control Using Time-Series Deep Neural Networks,” J. Manuf. Syst., 80, pp. 412–424.
[16] Bitencourt, J., Wooley, A., and Harris, G., 2025, “Verification and Validation of Digital Twins: A Systematic Literature Review for Manufacturing Applications,” Int. J. Prod. Res., 63(1), pp. 342–370.
[17] Zhang, K., Bui, A. T., and Apley, D. W., 2023, “Concept Drift Monitoring and Diagnostics of Supervised Learning Models via Score Vectors,” Technometrics, 65(2), pp. 137–149.
[18] Papadimitriou, I., Gialampoukidis, I., Vrochidis, S., and Kompatsiaris, I., 2024, “AI Methods in Materials Design, Discovery and Manufacturing: A Review,” Comput. Mater. Sci., 235, p. 112793.
[19] Hsu, T., Epting, W. K., Kim, H., Abernathy, H. W., Hackett, G. A., Rollett, A. D., Salvador, P. A., and Holm, E. A., 2020, “Microstructure Generation via Generative Adversarial Network for Heterogeneous, Topologically Complex 3D Materials,” JOM, 73(1), pp. 90–102.
[20] Shang, X., Liu, Z., Zhang, J., Lyu, T., and Zou, Y., 2023, “Tailoring the Mechanical Properties of 3D Microstructures: A Deep Learning and Genetic Algorithm Inverse Optimization Framework,” Mater. Today, 70, pp. 71–81.