In-situ process monitoring and adaptive quality enhancement in laser additive manufacturing: a critical review

Nomenclature

• It also defines machine-learning and analytical methods including ANN, CNN, SVM, PCA, GPR, and RMSE.Source: Nomenclature · pp. 3–4
• LPBF, LAM, and LDED are defined as core laser additive manufacturing terms.Source: Nomenclature · p. 3
• The nomenclature includes optical and imaging techniques such as OCT, MWIR, STFT, and X-CT.Source: Nomenclature · pp. 3–4

Motivation

• LAM’s stochastic process behavior can generate porosity, cracks, and distortions through complex thermal and material interactions, even with pre-optimized parameters, threatening critical-part reliability.The resulting need is for early defect detection and correction to improve as-built quality, process reliability, and reproducibility.Source: §1 · p. 6

Research gap

• Prior reviews often narrow their scope to specific sensors, ML, AM methods, or non-destructive testing, leaving insufficient coverage of recent monitoring advances and real-time process control.These limitations impede a holistic view across LAM techniques and underrepresent in-situ monitoring and control.Source: §1 · pp. 7–8

Review contribution

• The review comprehensively examines optical, acoustic, infrared thermography, and other in-situ sensing methods alongside ML-assisted defect detection, emphasizing standardization, reliability, sensitivity, data interpretation, and decision-making.It also reviews adaptive quality enhancement methods as a pathway toward zero-defect autonomous manufacturing.Source: §1 · p. 8

Review structure

• The review progresses from LPBF and LDED fundamentals and common defects, through in-situ and ML-assisted monitoring, to closed-loop adaptive quality enhancement and future research directions.Its roadmap targets fully autonomous, zero-defect LAM production and industrial-grade in-situ monitoring and control applications.Source: §1 · p. 9

2.1. Laser Powder Bed Fusion

• LPBF builds three-dimensional parts by selectively scanning and melting a thin, recoated layer of fine metal powder according to cross-sectional geometry, followed by rapid solidification; achieving fully dense parts can be challenging.Source: §2.1 · p. 11

2.2. Laser Directed Energy Deposition

• LDED deposits powder or wire with a focused laser, enabling repair, coating, larger parts, multimaterial components, and functionally graded materials, but with coarser microstructures and poorer precision.Higher laser power and larger spot sizes increase build rate, while localized heat accumulation and residual stress contribute to distortion, defects, and microstructure inhomogeneity.Source: §2.2 · p. 12

2. Defects in LAM

• LAM defects arise from parameter sensitivity, stochastic melt-pool dynamics, heat accumulation, and residual stress, making consistent quality and repeatability difficult.Defects can emerge even after parameter optimization, while transitions between conduction, keyhole porosity, and lack-of-fusion states remain difficult to determine.Source: §2.3 · p. 13

2.3. Defects in LAM

• LAM defects span micrometre-to-millimetre scales, including porosity, cracking, lack of fusion, balling, discontinuity, delamination, and microstructure inhomogeneity.Insufficient energy causes lack-of-fusion pores and balling, whereas excessive energy produces keyhole pores; porosity and cracking reduce strength, ductility, fatigue life, and corrosion resistance.Source: §2.3 · pp. 13–14
• Optimized parameters cannot eliminate defects in LAM, making in-situ detection and real-time remediation necessary to reduce post-production inspection and improve repeatability.Defects include microstructure inhomogeneity, roughness, geometric distortion, and process-specific bonding or thermal-expansion failures.Source: §2.3 · pp. 14–15

2.4. Comparative Analysis of LPBF and LDED Dynamics: Implications for Defect Formation and In-Situ Monitoring

• LPBF and LDED exhibit different process dynamics: LPBF cools at 10^5–10^7 K/s and is prone to microcracks, whereas LDED may use up to 15 kW, cools at 10^2–10^5 K/s, and can develop gas porosity.These fundamental differences necessitate specialized in-situ monitoring.Source: §2.4 · p. 17
• LPBF monitoring may require coaxial cameras up to 10 kHz, whereas LDED can use 30 Hz–1000 Hz sensing with larger fields of view.LDED’s noisier environment, larger laser spot, robotic-arm placement, and calibration requirements complicate acoustic monitoring and sensor deployment.Source: §2.4 · pp. 17–18
• LPBF generates high-volume monitoring data that challenges processing, while scarcer LDED data can reduce model generalizability; sensor placement, data handling, and ML models must therefore be tailored.The review emphasizes aligning monitoring solutions with the distinct needs of LPBF and LDED.Source: §2.4 · p. 18

Section scope

• In-situ monitoring is reviewed as a key approach for early defect detection, preventing quality deterioration and potential build failure in LAM.The section presents a critical review of recent monitoring advances and connects them to prior discussions of LAM defects.Source: §3 · p. 18

Monitoring methodologies

• The review covers optical, acoustic, laser line scanning, operando X-ray, and multisensor monitoring with data fusion.These methods are organized across Sections 3.1–3.5 and provide complementary approaches to in-situ sensing.Source: §3 · p. 18

Process signatures

• These sensing methods characterize critical process signatures, including melt pool dynamics, thermal histories, and acoustic features from laser–material interactions.Such signatures support understanding of process behavior relevant to defect detection.Source: §3 · p. 18

ML-assisted detection

• The section evaluates ML integration and models tailored to in-situ defect detection while synthesizing current capabilities and future monitoring directions.The evaluation provides a comprehensive view of state-of-the-art approaches and their potential development.Source: §3 · p. 18

3.1. Optical-Based Monitoring

• Optical monitoring captures melt-pool visual and thermal features used for process-quality assessment, defect detection, anomaly detection, and fault diagnosis.Infrared thermal imaging provides localized quality assessment and temperature-related information for LAM process control and optimization.Source: §3.1 · pp. 18–19

3.1.1. Melt Pool Dynamics And Visual Feature Extractions

• Coaxial cameras provide direct overhead melt-pool views, avoiding the image transformation and calibration required by oblique off-axis configurations.Off-axis systems are simpler to install but complicate dimensional measurement, whereas coaxial visual characteristics correspond to melting and cooling phenomena.Source: §3.1.1 · p. 19
• Melt-pool geometry and temperature features link process conditions to stability, solidification, grain size, and material-property-relevant process structure.Higher input energy density produces broader, deeper, and more asymmetrical melt pools, while melt-pool area can indicate stability and solidification state on sloped substrates.Source: §3.1.1 · p. 20
• Melt-pool imaging connects cooling conditions with microstructural outcomes: higher cooling rates increase columnar grain size, while heat accumulation enlarges melt pools at constant laser power.These relationships support controlling melt-pool size to influence grain size and subsequent material properties.Source: §3.1.1 · p. 21
• Infrared imaging reveals temperature distributions and solidus–liquidus phase boundaries, offering real-time insight into melt-pool heat transfer and material phase change.Laser power and gas-flow rate significantly influence temperature distributions, and IR features can be more informative than visible-spectrum monochrome images.Source: §3.1.1 · pp. 23–24, Figure 7 · p. 24

3.1.2. In-Situ Defect Detection With Melt Pool Features

• Melt-pool images support ML- and DL-based quality prediction, including porosity detection from morphology, temperature, spatter, and other process-related features.Vision-based approaches include supervised ML, unsupervised learning, and deep learning; reported methods distinguish pores from normal states using features such as circumference and coordinates.Source: §3.1.2 · pp. 24–25
• CNN-based melt-pool monitoring achieved 91.2% accuracy for porosity classification and predicted micropores below 100 µm in sponge-titanium samples.Traditional supervised methods require manual feature engineering and struggle with complex pore mechanisms and large datasets, whereas CNNs learn features automatically.Source: §3.1.2 · p. 25
• A real-time ensemble deep-learning system achieved 1.1 ms latency on low-power embedded boards and identified sagging, lack of fusion, irregular width, and lack of penetration on-the-fly.However, most optical defect-detection studies still rely on offline batch training, and real-time online inference remains underdeveloped for layered LAM processes.Source: §3.1.2 · p. 27

3.1.3. Anomaly Detection And Fault Diagnosis

• Supervised learning identified LoF, overheated, and unfused-powder anomalies, while a recoater-based line sensor captured powder-bed images with 99.15% accuracy; anomaly locations could be classified and stacked into 3D visualizations.Monitoring could predict component strength and density in addition to identifying defects.Source: §3.1.3 · pp. 27–28

3.1.4. Infrared Thermal Imaging For Localized Quality Predictions

• Infrared thermal histories produced voxel-level porosity predictions, micropore predictions with F1 scores above 0.96, and localized mechanical-property predictions using wavelet transforms and CNNs.The maximum IR radiance T_max was the most significant feature for voxel-state prediction, while thermal imaging remains limited for complex geometries when the camera view is blocked.Source: §3.1.4 · p. 29
• Existing infrared monitoring lacks sufficient industrial readiness because thermal imaging is largely restricted to thin-walled parts and cannot reliably capture profiles in complex geometries.Multiple thermal cameras from different viewpoints are suggested to address field-of-view blockage by the built structure.Source: §3.1.4 · p. 29

3.1.5. Challenges, Limitations, And Future Development

• Real-time vision monitoring must process immense data volumes at high speed, requiring robust computational resources and efficient algorithms.Neural-network pruning is identified as one possible way to accelerate predictive processing.Source: §3.1.5 · pp. 30–31
• Visual sensors differ in temperature range, acquisition frequency, field of view, and focal distance, making accuracy, cost, and integration difficult to balance.In Altenburg et al., a visible-spectrum high-speed camera with a NIR bandpass filter achieved the highest temporal and spatial resolution, but its limited dynamic temperature range could restrict process monitoring.Source: §3.1.5 · p. 31
• Accurate melt-pool temperature profiles remain elusive because infrared cameras require emissivity calibration affected by temperature, wavelength, phase, roughness, and other factors.Incorrect calibration introduces uncertainty and error into monitored data.Source: §3.1.5 · p. 31
• Integrating vision sensors can require custom coaxial laser-head designs or off-axis image transformation, while occlusion can hide process regions and reduce reliability.These integration issues are especially problematic for off-axis melt-pool monitoring.Source: §3.1.5 · p. 31

3.2. Acoustic-Based Monitoring

• Acoustic monitoring provides flexible sensor setups, fast dynamic responses, reduced hardware costs, and sensitivity to melting, solidification, crack propagation, and pore formation.The approach is readily integrated with existing LDED and LPBF equipment.Source: §3.2 · p. 32

3.2.1. Origins Of Acoustic Emissions And Acoustic Feature Extraction

• Acoustic signatures vary across penetration states, while stable processes show consistent mean intensity and anomalous processes show substantial intensity variation.These observations demonstrate the feasibility of acoustic monitoring for penetration-state recognition and link process anomalies with acoustic emissions.Source: §3.2.1 · p. 32, Figure 12 · p. 33
• Acoustic emissions originate from laser-material interactions, and time-, frequency-, and time-frequency features support process-state and quality recognition.STFT, wavelet transforms, and MFCCs have been used; a BPNN using cepstrum coefficients achieved 80%-90% accuracy in predicting welding penetration.Source: §3.2.1 · pp. 32–33

3.2.2. In-Situ Acoustic Monitoring In LPBF

• In LPBF, airborne microphones offer flexible, low-cost configurations, whereas substrate-mounted AE sensors detect structure-borne crack signals that vision and thermal sensing cannot capture.Microphone data quality depends on installation angle, distance, and sensor characteristics, while AE signals are susceptible to machine-movement and gas-flow noise and provide less setup flexibility.Source: §3.2.2 · p. 35
• Microphone installation angle strongly changes static-pressure signal energy, while laser-material-processing acoustic signals remain unaffected by microphone angle.Static-pressure fluctuations typically occur below 22.4 Hz and can have substantially higher amplitude than acoustic pressure signals.Source: §3.2.2 · p. 36
• Time-domain, frequency-domain, and time-frequency acoustic representations form distinct clusters for LoF, conduction, and keyhole-pore regimes across stainless steel, bronze, and Inconel.These clusters demonstrate acoustic-feature efficacy for distinguishing LPBF defect regimes regardless of alloy type.Source: §3.2.2 · p. 37
• LPBF acoustic features combined with ML achieved 89% accuracy for porosity classification, 93% for balling versus overheating, and 97% for keyhole-pore prediction.The 97% keyhole-pore result used an SVM classifier, while the 93% result used a deep belief network on band-pass-filtered raw acoustic signals.Source: §3.2.2 · p. 39
• A semi-supervised learning approach differentiated defect-free regimes from defects and achieved 97% accuracy; CNN models could be adapted by training only the final two fully connected layers.Reducing WT-feature extraction windows from 160 ms to 5 ms enabled faster and more precise defect identification; material-specific CNN development required balanced datasets and was expensive and time-consuming.Source: §3.2.2 · pp. 39–40

3.2.3. In-Situ Acoustic Monitoring In LDED

• LDED acoustic monitoring is particularly challenging because protective-gas flow and powder impact on the substrate create complex noise that obscures laser–material interaction sounds.In LPBF, protective-gas noise is concentrated in low frequencies and can be filtered, whereas LDED has a more complex noise composition.Source: §3.2.3 · p. 40
• LDED studies linked acoustic signals to build conditions and part quality, while flexible microphones, deep-learning denoising, and wavelet-integrated graph methods enabled broader monitoring and flaw-onset detection.Wavelet-based processing detected porosity, spatters, and line-width variations in wire-based DED, and deep-learning denoising generalized across machines and alloy materials.Source: §3.2.3 · pp. 40–41
• LDED acoustic emissions correlated with powder mass flow and laser power deviations, with higher laser power and mass flow producing increased AE intensity.The increased intensity indicated transfer of laser energy to powder particles and stimulation of acoustic waves.Source: §3.2.3 · p. 42
• LDED acoustic signatures distinguish defects: keyhole pores have the strongest 0–5000 Hz magnitude, while cracks and pores concentrate more low-frequency energy than defect-free deposition.Pores produce higher-energy, shorter-decay, lower-amplitude AE events than cracks, and acoustic energy best differentiates defects.Source: §3.2.3 · pp. 43–44
• MFCC-CNN classified LDED cracks and keyhole pores with 89% accuracy, 93% keyhole-pore prediction accuracy, and 98% AUC-ROC, outperforming other models.Denoising improved classification accuracy, while acoustic energy was the most important feature for distinguishing pores and cracks.Source: §3.2.3 · pp. 44–45

3.2.4. Challenges, limitations, and future development

• Acoustic monitoring requires improved noise removal, temporal and frequency resolution, multilayer interpretation, sensor standardization, fault diagnosis, and integration with optical or thermal sensing.Noise can arise from gas flow, machine motion, and environmental factors; microsecond-scale signal changes, inconsistent sensor setups, overlapping fault signatures, and difficult data fusion constrain robustness.Source: §3.2.4 · p. 46

Operating principle

• Laser triangulation converts shifts in reflected laser lines caused by surface irregularities into detailed topographical maps for defect identification.A laser line scanner projects a line onto the surface, while an angled optical sensor measures changes in the reflected line position.Source: §3.3 · p. 47

Data processing and applications

• Laser line scanning supports surface-defect detection in LDED by processing point clouds or 2D depth images to identify bulges, dents, over-extrusion, under-extrusion, and over- or under-built regions.Statistical outlier removal, clustering, RANSAC, CNN classification, and window-based depth-to-point-cloud processing are used to clean and interpret the measurements.Source: §3.3 · p. 48

Capabilities and limitations

• Laser line scanning has shown substantial promise for in-process surface defect detection in LDED, but raw point clouds are noisy and shadow effects reduce accuracy at edges and corners.Outliers, unwanted substrate surfaces, abrupt height changes, and measurement noise can impair subsequent defect detection.Source: §3.3 · p. 48

System integration and deployment

• Detection may miss minuscule defects, while limited scanning speed, dust, smoke, lighting variation, alignment, and spatial constraints can cause bottlenecks or false readings.Integration is easier in robotic LDED systems but more difficult in CNC-based LDED and LPBF because coordinate, TCP-motion, and available-space constraints complicate implementation; powder beds can disrupt the laser line.Source: §3.3 · p. 48

Future directions

• Future work should develop adaptable, miniaturized scanners and robust algorithms, while enabling real-time coordinate and TCP-motion estimation and reducing background-noise disruption.Additional sensors and machine-learning methods may help acquire or estimate machine coordinates and motion data across CNC, LPBF, and other LAM systems.Source: §3.3 · p. 49

• Operando X-ray imaging rapidly and accurately detects keyhole pores and can provide ground-truth labels for annotating in-situ monitoring data.It reveals material structure and dynamics during operation, supporting supervised deep-learning detection with sub-millisecond temporal resolution.Source: §3.4 · p. 51
• ICI monitored LDED surface topology and cracks from an off-axis 24° alignment without modifying laser-delivery optics, detecting cracking events with subsurface signals.ICI surface topology correlated with corresponding X-ray radiographs at >0.93 and revealed humping-induced cracking in surface valleys of nickel-superalloy thin-wall builds.Source: §3.4 · pp. 51–52
• Operando X-ray deployment is limited by scarce equipment, radiation-safety requirements, costly data analysis, and difficulty extending studies beyond single tracks or thin walls.ICI integration is complex, especially off-axis, and its optical measurements can miss small or deep defects and degrade under dust, smoke, or temperature fluctuations.Source: §3.4 · p. 52
• In LPBF, combined thermal and synchrotron X-ray imaging exposed subsurface dynamics and distinct keyhole oscillation forms during Ti-6Al-4V processing.Operando X-ray observations also distinguished keyhole pores, conduction mode, and lack-of-fusion pores using heterogeneous sensors.Source: §3.4 · p. 51, Figure 21 · p. 53
• In LDED, high-speed X-ray imaging revealed how laser–matter interaction drives powder flow and porosity formation, including individual particles entering the scanning melt pool.These observations provide process insight for tracking powder flow and porosity during deposition.Source: §3.4 · p. 51, Figure 21 · p. 53

3.5. Multisensor Monitoring And Data Fusion

• Multisensor fusion counterbalances individual sensor limitations, improves signal-to-noise ratio, and supports more comprehensive and predictable LAM process and defect assessment.Single sensors often detect particular defects well but do not capture the full process dynamics.Source: §3.5 · p. 53

3.5.1. Multisensor Monitoring Setups In LAM

• Heterogeneous setups improve multiscale defect detection: LPBF systems can identify porosity below 100 µm, layer inconsistencies from 100 µm to 1 mm, and geometry flaws above 1 mm.Robotic LDED examples combine coaxial visible melt-pool imaging, off-axis infrared thermography, and microphones; LPBF systems combine thermal, powder-bed, spatter, acoustic, emission, and trajectory sensing.Source: §3.5.1 · p. 54, Figure 22 · p. 55

3.5.2. Multisensor feature visualization and correlation analysis

• Correlated multisensor features distinguish process conditions and link anomalies such as cracks and keyhole pores to abrupt feature shifts, while sensor importance depends on modality and time window.BR and AE carried the most informative content; AE became more important with shorter windows, whereas BR produced more reliable and accurate results over longer integration periods.Source: §3.5.2 · pp. 55–56

3.5.3. Spatiotemporal data registration

• Spatiotemporal synchronization and registration are required for location-specific defect prediction, but missing real-time machine positions, differing acquisition speeds, and offline spatial data limit reliability and real-time localization.Registering multisensor features with real-time robot TCP positions enables localized quality prediction and defective-region estimation using timestamp-based approximate synchronization.Source: §3.5.3 · pp. 57–58
• Spatiotemporal registration aligns melt-pool images, machine logs, tool traces, and multisensor features without requiring system-to-system synchronization, but access to real-time TCP position information remains limited.Deep learning can estimate melt-pool coordinates in the machine coordinate system from process settings and melt-pool images.Source: §3.5.3 · p. 58, Figure 24 · p. 59

3.5.4. Multimodal sensor fusion and quality prediction

• Multimodal fusion combines raw sensor data, extracted features, or model decisions, trading information retention against noise rejection, synchronization, computational cost, and processing complexity.Data-level fusion offers high precision with minimal information loss but struggles with asynchronous signals and large data volumes; feature-level fusion reduces dimensionality, whereas decision-level fusion can be less computationally intensive.Source: §3.5.4 · pp. 59–60
• Feature-level fusion outperformed data-level and decision-level fusion models in classification accuracy for layer-wise LPBF quality estimation using photodiode and microphone signals.The compared models used channel fusion, fused extracted signal features, or combined individual classification results.Source: §3.5.4 · p. 60
• Deep multimodal models enable localized quality maps and transient track-geometry prediction by combining visual, acoustic, process-state, and location information.A hybrid CNN fused visual and audio streams for localized prediction, while a VGG-based model used power, velocity, and encoded motion states to predict track geometry during transients and reduce dimension deviations.Source: §3.5.4 · pp. 60–61

4. Adaptive Quality Enhancement

• Adaptive quality enhancement targets defects from fixed process parameters, dynamic heat accumulation, feature-size variability, and position or speed deviations that produce pores, geometric errors, and microstructural inhomogeneity.The section focuses on closed-loop feedback control to pre-empt defect generation and in-process defect correction through additive or subtractive methods.Source: §4 · p. 61

4.1. Closed-Loop Feedback Control

• Closed-loop systems use sensor signals such as melt-pool temperature or width to adjust laser power, scanning speed, and powder feed rate, improving geometric accuracy, surface finish, and microstructure homogeneity.A representative controller maintains melt-pool peak temperature to reduce heat accumulation and suppress porosity defects.Source: §4.1 · p. 62
• Adaptive controllers eliminate manual parameter adjustment across varying LDED shapes, materials, toolpaths, and process parameters, remove unstable ripple dynamics through layer-by-layer height control, and enhance spiral-part geometric precision.Chen et al. introduced an adaptive PID controller with automatic parameter tuning for LDED processes. A feedforward PI controller enhances spiral-part geometric precision [passage does not provide a bibliography marker].Source: §4.1 · p. 62
• Closed-loop control remains limited by validation on single tracks or thin walls, linearized plant models, and insufficient analysis of complex geometries, microstructures, and mechanical properties.Simulation-guided control can improve surface finish and geometric accuracy but remains vulnerable to systematic simulation errors, while deep-reinforcement-learning thermal control neglects convection and radiation effects.Source: §4.1 · pp. 62–63
• Closed-loop laser-power control produces more homogeneous SS 316L microstructures by suppressing heat accumulation and reducing temperature-gradient and solidification-rate variation across layers.Open-loop samples transition from fine cellular structures to columnar grains and dendrites, whereas closed-loop samples retain a consistent cellular grain structure with varying cell sizes.Source: §4.1 · p. 66
• Reducing laser power accelerates melt-pool solidification and produces a microstructure dominated by uniform fine equiaxed grains, unlike uncontrolled samples with coarser and mixed structures.Figure 27 contrasts uncontrolled cellular-columnar and coarse-grained structures with the consistent cellular grain structure achieved under closed-loop control.Source: §4.1 · p. 66, Figure 27 · p. 67

4.2. In-Process Defect Correction

• Closed-loop control can pre-empt defects and improve microstructure homogeneity, but dimensional inaccuracies, subsurface defects, weak sensing signals, and suboptimal parameter adjustments limit single-input single-output controllers.These limitations arise from interacting error sources and process instabilities in complex geometries.Source: §4.2 · pp. 67–68
• In-process defect correction complements feedback control by using additive or subtractive methods to repair dimensional deviations and subsurface flaws during fabrication, providing a backup for as-fabricated quality.A laser-line-scanned point cloud can identify bulges and dents, after which corrective toolpaths select LDED deposition or subtractive machining; the approach can also address pores and cracks.Source: §4.2 · p. 68
• Height variances in wire-based LDED were effectively compensated within a few layers using laser-line height mapping and segment-specific wire-feed control.Integrating detection and correction can enhance LAM quality, but software unification, multiscale defect handling, and more precise adaptive toolpaths remain unresolved.Source: §4.2 · p. 68

Summary and Future Perspectives

• Section 5.1 reviews key findings and developments from preceding sections, while Section 5.2 discusses major challenges and future directions.The section provides a concise summary and future outlook for in-situ process monitoring and control in LAM.Source: §5 · p. 70

Monitoring methods

• Vision-based monitoring supports melt-pool, thermal-history, powder-bed, and powder-stream diagnostics, while deep learning improves real-time image-based quality prediction.Cameras and pyrometers can operate coaxially or off-axis across SWIR, MWIR, and visible spectra.Source: §5.1 · p. 70
• Acoustic monitoring offers flexible installation, rapid response, and lower hardware cost than vision-based monitoring, but its noise rejection requires improvement.Time-, frequency-, and time-frequency representations can associate acoustic patterns with defects and anomalies such as spatter events.Source: §5.1 · pp. 70, 73
• Multisensor fusion combines heterogeneous sensor data to predict defects more accurately than traditional single-sensor monitoring.The review covers vision, acoustic, laser line scanning, operando X-ray, and inline coherent imaging approaches.Source: §5.1 · p. 70

Defect-detection hierarchy

• Defect detection spans region-based, layer-wise, and real-time levels, trading higher accuracy or resolution against spatial-temporal limitations, inspection pauses, computational demands, and cost.A hierarchical multisensor system spanning all three levels may therefore provide desirable multiscale detection.Source: §5.1 · p. 71

Adaptive quality enhancement

• Closed-loop feedback control can reduce porosity, improve geometric accuracy, enhance microstructure homogeneity, and strengthen consistency and overall product quality.In-process defect correction uses subtractive or additive methods to eliminate detected defects and protect final-product quality.Source: §5.1 · p. 71

Overall benefits and limitations

• In-situ monitoring and adaptive enhancement enable nondestructive early correction, real-time process adaptation, microstructure control, improved mechanical properties, reduced waste, higher efficiency, and economic benefits.Operando X-ray and inline coherent imaging show high defect-identification accuracy but face industrial adoption barriers from cost and safety concerns.Source: §5.1 · pp. 71–72

5.2. Research Gaps And Future Perspectives

• The review identifies six research gaps and future R&D directions for in-situ monitoring and adaptive quality enhancement in LAM.These directions are visually represented in Figure 29.Source: §5.2 · p. 75

5.2.1. Standardization And Reproducibility

• Standardizing sensor setups and validation methods is needed to improve reproducibility and ensure ML defect-detection models are robust and adaptable across a range of conditions.Source: §5.2.1 · p. 75

5.2.2. Location-Specific Quality Prediction

• Location-specific quality prediction remains underdeveloped because models rarely connect in-situ monitoring with localized characteristics amid complex spatially dependent process physics.Future ML methods should synthesize spatial and temporal data, potentially using robot-based LDED Tool Centre Point data for real-time location information.Source: §5.2.2 · p. 76

5.2.3. Reliable Defect Detection Models For Real Production

• Defect-detection models trained on thin-wall or single-track parts may underperform in complex, multilayer, multitrack, or multi-material production scenarios.Future models should use real-production data capturing stochastic defect occurrence, varied defect mechanisms, and multi-material manufacturing; transfer learning is one proposed approach.Source: §5.2.3 · pp. 76–77

5.2.4. Multisensor Data Fusion

• Multisensor fusion requires optimal sensor combinations, noise-aware processing, uncertainty quantification, and cost-conscious selection of a minimum effective sensor set.These measures should improve sensitivity, reliability, robustness, and cost-effectiveness while accounting for heterogeneous data and sensor-specific relevance.Source: §5.2.4 · p. 77

5.2.5. Decision-Making Strategies Beyond Early Stopping

• LAM decision-making must move beyond anomaly-triggered process termination toward real-time parameter adjustment that mitigates and rectifies defects.This requires understanding defect formation and propagation while balancing productivity against material waste.Source: §5.2.5 · p. 78
• Physics-informed, data-driven, or rule-based decision-making should enable dynamic process adjustments, defect correction, and proactive parameter changes before defects occur.Proactive strategies could reduce defective-material waste, shorten correction time, and improve LAM efficiency, but enabling technology remains in its infancy.Source: §5.2.5 · p. 78

5.2.6. Hierarchical Multi-Scale Defect Identification And Prediction

• Hierarchical defect identification remains underexplored, requiring integrated monitoring across region-based, layer-wise, and real-time quality levels.Existing research commonly targets one level; a multisensor system could assign sensors to defect types and scales, improving holistic quality control, robustness, and reliability.Source: §5.2.6 · pp. 78–79
• Multimodal sensor fusion must minimize redundancy and cost while preserving reliable, robust defect detection, supporting future self-adaptive LAM systems.Addressing these gaps could enhance process predictability, quality, efficiency, and reliability across industrial applications.Source: §5.2.6 · p. 79

Roadmap overview

• The roadmap targets fully autonomous, defect-free LAM that operates continuously without human supervision through coordinated sensing, prediction, and rectification.It combines AI, sensor technology, and robotics to support continuous operation despite potential anomalies.Source: §5.3 · pp. 80, 82

i. Multisensor Monitoring and Data Fusion Framework

• Multisensor data fusion should combine complementary sensors, reduce redundancy, and improve predictive accuracy, reliability, process understanding, and monitoring cost-effectiveness.Using a minimal sensor set can reduce noise, disturbances, and monitoring costs while preserving comprehensive process information.Source: §5.3 · p. 81

ii. Proactive Defect Prediction and Mitigation Models

• Physics-informed machine-learning models should predict defects from historical and real-time multisensor data and automatically adjust laser power or speed before defects occur.Using spatiotemporal knowledge such as thermal histories could improve quality consistency and reliability while reducing defective-material waste and correction time.Source: §5.3 · p. 81

iii. Multi-scale Hierarchical Defect Identification and Rectification Models

• Hierarchical defect management should identify micro- to macro-scale defects in real time, layer-wise, and by region, then rectify them automatically rather than merely halting.Robotic machining could remove detected defects, while specialized sensors support robust, reliable, and comprehensive quality control across operational scales.Source: §5.3 · p. 81

Table A1. Vision-based defect and fault detection

• Vision-based methods span coaxial and off-axis optical, thermal, and high-speed imaging for porosity, melting-state, anomaly, geometry, and process-quality detection, with reported accuracy, F1, and latency results.Examples include CNN melting-state classification accuracy of 0.997, layer-wise anomaly classification F1 of 86.6% and accuracy of 87.3%, and LDED CNN inference at 1.1 ms with 95.1% accuracy and 95.2% F1.Source: Table A1 · p. 83, Appendix A · p. 84
• Optical studies are frequently limited to single-track or thin-wall experiments, layer-wise inspection, restricted materials or geometries, limited spatial or temporal resolution, and incomplete spatial or location-specific information.Several studies additionally depend on specific thermal features, offer limited sensor flexibility, or lack generalizability beyond their tested settings.Source: Appendix A · pp. 83–84
• Vision studies also report localized porosity and defect-prediction performance, including classic-machine-learning F1 of 0.966, statistical TPR of 0.9 with FPR of 0.15, and ResNet F1 of 0.974 with accuracy of 0.967.Other reported results include semi-supervised CNN accuracy above 95%, variational-RNN ROC AUC of 0.944, and laser-power prediction failure rate below 1.1%.Source: Appendix A · pp. 83–84

Table A2. Acoustic-based defect detection

• Acoustic-based monitoring uses microphones, acoustic-emission transducers, and fibre Bragg grating sensors for regime, pore, anomaly, quality, and melting-state classification.Reported results include VGG-16 accuracy of 93%, SVM keyhole-pore accuracy of 97%, VAE/GAN accuracy of 96%–97%, CNN build-condition accuracy of 95%, and DBN melting-state accuracy of 93%.Source: Table A2 · p. 86, Appendix A · pp. 86–87
• Acoustic approaches remain constrained by limited spatial or temporal resolution, non-location-specific sensing, restricted defect types or process conditions, and limited material or process generalizability.Some methods emphasize general quality or process-condition classification rather than defect correlation and give limited attention to inference time.Source: Appendix A · pp. 86–87

Table A3. Multisensor fusion-based defect detection

• Multisensor fusion combines optical, thermal, powder-bed, acoustic, and reflection measurements, but surveyed systems remain limited to specific flaws, materials, geometries, settings, or general-quality objectives.Reported limitations include installation complexity, thin-wall study designs, and a lack of location- or defect-type-specific information.Source: Table A3 · p. 88, Appendix A · p. 88

• The authors declare no known competing financial interests or personal relationships that could have appeared to influence the reported work.Source: Declaration of competing interest · p. 90