Researchers at Northwestern University and Case Western Reserve University have unveiled a digital twin framework designed to optimize laser-directed energy deposition (DED) using machine learning and Bayesian optimization. The system integrates a Bayesian Long Short-Term Memory (LSTM) neural network for predictive thermal modeling with a new algorithm for process optimization, establishing one of the most complete digital twin implementations for metal additive manufacturing to date.
The study, titled “Towards a Digital Twin Framework in Additive Manufacturing: Machine Learning and Bayesian Optimization for Time Series Process Optimization,” was published in the Journal of Manufacturing Systems. It builds toward real-time adaptive control in DED, an additive process known for its potential in functionally graded materials and repair applications, yet often limited by thermal instability and part-to-part variability.
A step toward true digital twins in additive manufacturing
In metal AM, DED systems direct a laser beam onto a substrate while feeding metallic powder or wire, building parts layer by layer. While the process allows near-net-shape fabrication and material customization, it also produces cyclic thermal fluctuations that lead to microstructural heterogeneity and mechanical anisotropy.
Existing feedback control approaches, such as PID or feedforward controllers, have mitigated some thermal variation but remain limited by their one-way information flow and inability to anticipate unmodeled disturbances. The Northwestern team sought to overcome these challenges by creating a fully bidirectional digital twin, a system that continuously exchanges data between simulation and the physical DED process to enable dynamic, model-predictive control.
Unlike so-called “digital shadows,” which only mirror process behavior, the proposed framework updates its virtual model on the fly and sends optimized process parameters back to the machine. The authors argue this capability is central to achieving the Industry 4.0 vision of cyber-physical manufacturing systems capable of intelligent self-correction.
Real-time prediction with uncertainty quantification
At the core of the framework is a Bayesian LSTM surrogate model trained to predict temperature histories across the part during laser deposition. The model replaces computationally expensive finite element method (FEM) simulations, which are too slow for real-time operation.
The neural network was trained on high-fidelity thermal data generated using Northwestern’s Additive Rapid Prototyping Instrument (ARPI), simulating thin-wall builds in Inconel 718. Each dataset varied laser power, frequency, and scan path to capture a wide range of thermal behaviors.
By combining Monte Carlo dropout with variational Bayesian inference, the model estimates not only the expected temperature but also its confidence interval, quantifying uncertainty caused by material variability, measurement noise, or model simplifications. This capability allows the digital twin to make informed adjustments under uncertainty, an essential feature for dynamic process control.
The model achieved an R² score of 0.75 when validated against independent datasets, demonstrating robust accuracy for real-time use. Predictions were most precise near the active melt pool, where feedback data is available, and became less certain in deeper or thermally lagged regions of the part.
According to the authors, this surrogate modeling approach provides a foundation for real-time Model Predictive Control (MPC) in future physical DED-DT systems.
Bayesian optimization for time-series control
To complement the predictive model, the researchers developed Bayesian Optimization for Time Series Process Optimization (BOTSPO), a data-efficient algorithm for determining the optimal laser power profile over time.
Traditional optimization of DED parameters is computationally expensive due to the high dimensionality of time-series data. BOTSPO addresses this challenge by representing the laser power curve as a modified Fourier series defined by only ten tunable parameters, including amplitude, frequency, phase, linear trend, seasonality, and their rates of change.
A Gaussian Process (GP) model evaluates how each candidate profile affects a target objective, in this case the heat treatment time, the period the material spends within a temperature range (654–857 °C) critical for Inconel 718’s γ” phase precipitation hardening.
The optimizer then uses an Upper Confidence Bound (UCB) acquisition function to balance exploration and exploitation, dynamically adjusting its search based on model uncertainty.
Applied to the simulated thin-wall geometry, BOTSPO increased the average heat-treatment time from 30.3 s to 38.2 s, a 26% improvement. Early iterations explored widely, while later ones refined the power trajectory toward an optimal sequence: high initial laser power to promote uniform heating, followed by gradual reduction as thermal dissipation increases with build height.
The result, while improving mechanical performance potential, revealed remaining challenges in achieving thermal homogeneity throughout the part, a focus for the team’s planned multi-objective extensions.
Bridging simulation and control through digital twins
The proposed digital twin architecture connects the offline optimization achieved by BOTSPO with online predictive control driven by the Bayesian LSTM.
In practice, the optimized laser power profile serves as a reference trajectory for the digital twin’s real-time controller, which can modify inputs dynamically in response to in-situ sensor feedback.
This workflow represents a major step toward autonomous AM process control, combining physics-based models, data-driven learning, and dynamic decision-making within a unified system. According to the authors, the next phase of research will focus on integrating this architecture with a physical DED system and developing continuous model updates that correct for simulation drift during fabrication.
Toward data-driven DED manufacturing
By merging uncertainty-aware neural networks with adaptive Bayesian optimization, the Northwestern-led team demonstrates a practical path toward intelligent digital twins in metal additive manufacturing. The study underscores a broader shift in AM research, from static process monitoring to predictive, feedback-based systems capable of self-optimization.
The work was supported by the NSF Engineering Research Center for Hybrid Autonomous Manufacturing: Moving from Evolution to Revolution (ERC-HAMMER), the National Science Foundation Graduate Research Fellowship, and Northwestern’s Predictive Science & Engineering Design Cluster.
Digital twins for additive manufacturing
Digital twin technology is rapidly expanding across industrial and AI research domains. Earlier this year, Second Life secured $10 million in funding to develop AI-powered digital twin capabilities aimed at improving manufacturing process intelligence. In the maritime sector, a UK consortium launched the world’s first digitally enabled, additively manufactured propeller project, demonstrating how integrated simulation and performance monitoring can accelerate certification and design validation.
Meanwhile, Meta unveiled the world’s largest 3D model digital twin catalog to support AI-driven research, underscoring the growing convergence between additive manufacturing, simulation, and artificial intelligence. The work led by Northwestern University extends these advances by translating digital twin principles into real-time, closed-loop optimization for metal additive manufacturing.
To stay up to date with the latest 3D printing news, don’t forget to subscribe to the 3D Printing Industry newsletter or follow us on Twitter, or like our page on Facebook.
While you’re here, why not subscribe to our YouTube channel? Featuring discussion, debriefs, video shorts, and webinar replays.Want to help shape the future of AM? Join the 3DPI Expert Committee.
Featured image shows schematic of the proposed Digital Twin Framework for DED. Image via Journal of Manufacturing Systems
