A systematic review published in IEEE Access by researchers from the University of Porto, Fraunhofer IWS, Luleå University of Technology, Oxford University, INESC TEC, and the Technical University of Dresden has mapped the emerging use of artificial intelligence (AI) in laser-based additive manufacturing (LAM) process control. Analyzing 16 studies published between 2021 and 2024, the review found that 62.5% had deployed AI-driven controllers in real-world environments, while 56% applied AI specifically for control strategies such as reinforcement learning (RL). AI was used for process modeling or monitoring in 62.5% of the reviewed work, and 68% of studies targeted instability caused by underheating or overheating — a primary source of defects in LAM.
LAM, a subset of additive manufacturing, builds metal parts layer by layer using either Powder Bed Fusion (PBF) or Directed Energy Deposition (DED). PBF operates at lower laser power, typically under 1 kW, with scanning speeds up to 2,000 mm/s, producing high-precision parts on a small scale. DED uses powder or wire feedstock with laser power reaching 40 kW, enabling higher deposition rates and larger structures. These process differences produce distinct melt pool characteristics, thermal gradients, and defect profiles, requiring separate control strategies.
The review identified five key stages where AI can be applied: intelligent sampling, process monitoring, modeling, controller design, and performance evaluation. Despite the high cost of LAM trials, classical design of experiments (DoE) methods remain the most widely used for sampling, while adaptive techniques such as Bayesian optimization — which adjusts data collection toward regions of rapid change or high nonlinearity — are less common but show efficiency gains in other manufacturing contexts.
Monitoring approaches often combine thermal and visual sensing to capture melt pool geometry and temperature, parameters that correlate strongly with part quality. Sensors are mounted on-axis, aligned with the laser, or off-axis to monitor externally. In reviewed work, convolutional neural networks processed image-based data, while artificial neural networks proved effective in multimodal monitoring, integrating visual, thermal, and acoustic signals. One study found that this multimodal ANN approach achieved the highest defect-prediction accuracy among tested models. Data labeling remains a limitation: many studies relied on simplified melt pool signatures or final inspection results rather than continuous real-time labels, limiting responsiveness in high-speed processes such as PBF.
Modeling methods span finite element modeling (FEM), recurrent neural networks (RNNs), long short-term memory networks (LSTMs), and system identification techniques such as Dynamic Mode Decomposition (DMD) and Sparse Identification of Nonlinear Dynamics (SINDy). FEM offers detailed thermal and mechanical simulations but requires significant computational resources and precise calibration. Data-driven models can capture nonlinear dynamics from observed data, with RNNs and LSTMs effective for time-series predictions. Hybrid models combining FEM with AI have been explored to balance predictive accuracy with reduced computational load.
Controller strategies in the reviewed studies included proportional–integral–derivative (PID) systems, model predictive control (MPC), and reinforcement learning. PID controllers, often tuned manually, remain common but are less effective in managing process uncertainty. MPC, which uses a surrogate process model to run optimization routines inside the control loop, was shown in one study to outperform PID when uncertainty was explicitly modeled. RL methods, including Q-learning and model-based RL (MBRL), were used to adjust parameters such as laser power and scanning speed based on real-time feedback. In one example, MBRL achieved optimal process parameters more quickly than a model-free RL baseline, producing lower surface roughness and higher cumulative rewards.
Most real-time control implementations in the review were limited to layer-wise adjustments rather than continuous closed-loop control. This limitation reflects both the high data rates required for melt pool monitoring and the computational demands of advanced AI methods. PBF presents particular challenges, with sub-millisecond melt pool dynamics and scanning speeds up to 2,000 mm/s.
Machine learning techniques outside of control included support vector machines for high-dimensional defect detection and Gaussian regression models for predicting disturbances affecting part quality. Deep learning methods such as convolutional neural networks assessed surface roughness and process stability, while recurrent networks modeled parameter changes over time. These methods, though effective, require large datasets and significant computational resources, constraining their industrial use.
The review noted several systemic gaps in AI adoption for LAM. Data acquisition is constrained by cost, limited availability of representative samples, and the need for edge-case scenarios. Computational demands from FEM, RL, and deep learning require specialized hardware and expertise. Many approaches have been validated only in simulations or limited-scale experiments, with few demonstrated in production settings. Model interpretability is also a barrier; none of the reviewed studies incorporated all three critical evaluation metrics — explainability, uncertainty, and robustness — in assessing AI performance.
Evaluation metrics, where applied, used techniques such as SHapley values or LIME for explainability, Bayesian networks and Gaussian processes for uncertainty, and adversarial training or robust optimization for stability. However, these were reported inconsistently, making cross-study comparisons difficult. The absence of standardized metrics was identified as a major obstacle to technology transfer into industrial environments.
The authors identified several priorities for future work. Expanding adaptive sampling methods such as Bayesian optimization could improve data efficiency. Increasing multimodal sensor integration, for example, combining visual, thermal, and acoustic monitoring, could enhance defect detection and process stability. Developing hybrid modeling frameworks that merge FEM and AI may yield more computationally tractable yet accurate process simulations. Scaling MPC and RL to production settings will require optimizing algorithms for latency and reducing computational overhead, along with building robust real-time data pipelines. Standardized evaluation frameworks that integrate explainability, uncertainty, and robustness metrics would facilitate meaningful comparison across studies and accelerate industrial adoption.
While AI in LAM remains in an early phase, the review shows a gradual shift from monitoring toward active process control. Traditional methods such as PID controllers and classical sampling still dominate, but advanced strategies like model-based reinforcement learning and hybrid FEM–AI modeling are gaining traction. Demonstrated benefits include reduced surface roughness, improved process stability, and faster convergence toward optimal manufacturing parameters. If combined with standardized metrics, adaptive sampling, and real-time continuous control, these methods could support fully autonomous metal additive manufacturing systems capable of self-monitoring and self-correction.
Want to help choose the 2025 3DPI Awards winners? Join the Expert Committee today.
Ready to discover who won the 2024 3D Printing Industry Awards?
Subscribe to the 3D Printing Industry newsletter to stay updated with the latest news and insights.
Featured image shows overview of two primary laser-based additive manufacturing processes-Directed Energy Deposition (DED) and Powder Bed Fusion (PBF). IEEE Access.
