Accelerating process development for 3D printing of new metal alloys

Additive manufacturing (AM), particularly laser powder bed fusion (L-PBF), is increasingly used across sectors such as healthcare, aerospace, and defense due to reduced lead times, production efficiency, part consolidation, and design freedom. However, variability in mechanical properties and dimensional accuracy across powders, machines, scan strategies, and process conditions remains a key barrier to reliable production. Process mapping of power–velocity (P–V) parameters to defects is essential for determining optimal conditions that ensure melt pool stability and adequate track overlap to prevent lack-of-fusion defects that degrade fatigue life. Conventional process development relies on ex situ characterization by trained technicians. In situ monitoring using high-speed imaging and temperature field measurements has been explored, but deployment is hindered by limited observable features, low algorithmic accuracy for static-image analyses, and complex, costly setups (e.g., pyrometers, calibration and alignment with the scan head). This work investigates whether temporal features of melt-pool dynamics captured via simple off-axial high-speed imaging and analyzed with video vision transformers can enable accurate, generalizable, and efficient in situ process mapping for new metal alloys.

Prior studies have examined sources of variability and process–property relationships in L-PBF, highlighting the influence of laser power, scan speed, and hatch spacing on melt pool stability and defect formation (e.g., keyholing, balling, lack-of-fusion). Ex situ process mapping is well established for qualifying parameter windows. In situ approaches using high-speed imaging and multi-phenomena sensing (including pyrometry and thermal fields) can improve monitoring but often require complex instrumentation and calibration. Earlier computer vision and CNN-based methods focused on static images, which may miss temporal signatures critical for detecting periodic defects (e.g., balling) and transient keyhole instabilities. Ultrafast X-ray and optical studies documented keyhole dynamics and fluctuation frequencies and linked them to porosity formation. This work builds on these insights by leveraging temporal-spatial representations via transformers to improve classification accuracy and generalizability across alloys without necessitating temperature field measurements.

An off-axial high-speed imaging setup was integrated with a TRUMPF TruPrint 3000 L-PBF machine at CMU’s Mill 19 facility. A Photron FASTCAM Mini AX200 monochrome camera recorded melt pool dynamics at 54,000 fps using magnification lenses and optical filters to attenuate plasma plume emissions. Exposure was tuned to avoid sensor saturation and blooming; argon shielding maintained O2 < 0.1%. Single-bead experiments across diverse P–V combinations covered four regimes: desirable, keyholing, balling, and lack-of-fusion. Three alloys were studied: SS316L, Ti-6Al-4V, and IN718; a total of 1340 tracks (6 mm length; ends not imaged) were printed. For variability mapping, Ti-6Al-4V single tracks (30 µm layer, ~100 µm spot) were printed at powers 200–400 W and speeds 800–1600 mm/s; each condition was repeated four times. Data processing: Videos were converted to grayscale, spatially registered to align the melt pool, and intensity-normalized (per-video linear scaling to [0,1]). Each input clip was shaped 80×160×15 (spatial resolution ~6.3 µm; temporal resolution ~18.5 µs). Active contouring and thresholding quantified melt pool attributes (width, area). Ground-truth labels were assigned via ex situ microscopy (ZEISS Axio Imager) using criteria: (1) keyholing when deep/narrow penetration likely to generate pores (width/depth < 1.2 or porosity observed); (2) balling when bead shows periodic humps/ball-like peaks (even if shallow); (3) lack-of-fusion when shallow track without balling; (4) desirable otherwise. If any defect morphology was observed for a P–V condition, that condition was labeled with the corresponding defect class. Model: A video vision transformer (ViViT) with tubelet (nonoverlapping 3D patch) embedding encoded spatiotemporal information. The backbone used a transformer encoder with multi-head self-attention, MLP, layer normalization, residual connections, and GELU activations. The architecture used 12 heads, 20 layers, and MLP hidden dimension 256. Regularization via elastic net (L1+L2) was applied to MLPs; minimal data augmentation was used to preserve intensity and geometric dynamics. Training used TensorFlow on an NVIDIA Tesla V100 32 GB GPU, 200–500 epochs to convergence, batch size 128. Cross-dataset evaluation tested generalizability: models trained on one alloy and tested on the others with unchanged hyperparameters. Benchmarking against CNN-based models (VGG16, ResNet152, MoViNet) and alternate transformer variants (TimeSformer, ViViT-B pretrained variants) was performed. Process maps were generated by averaging class predictions across samples for each P–V combination. Variability maps were derived from standard deviations and relative standard deviations of melt pool width and area over time and repetitions.

• The off-axial high-speed imaging captured clear melt pool shapes and temporal dynamics across regimes. Keyholing exhibited the largest intensity fluctuations; conduction/stable regimes showed higher temporal correlation between frames. Measured intensity fluctuation frequencies were approximately 8–17 kHz, exceeding reported keyhole depth fluctuation frequencies (~10 kHz), consistent with plume emissions affecting visibility.
• Cross-dataset generalization: Training on SS316L and testing on IN718 achieved top-1 accuracy 96.63% and top-2 100% (F1 up to 1.0); testing on Ti-6Al-4V achieved top-1 87.60% and top-2 95.87% (with lower F1 for balling, 0.69). Overall, top-1 accuracies up to ~98% were reported in some evaluations.
• Process maps generated in situ from transformer classifications agreed well with ex situ maps, with one noted misclassification at high power in Ti-6Al-4V (balling misidentified as desirable).
• Transformer models outperformed CNN-based baselines across cross-dataset tests. Pretrained ViViT-B and TimeSformer achieved >90% mean accuracy in many cases; ViViT-B showed balanced performance (e.g., up to ~98% when training on IN718 and testing on IN718/SS316L).
• Ablations indicated larger backbone capacity improves accuracy, but sufficient temporal context is essential to distinguish regimes with similar objects/actions but different dynamics. Regularization (and in some cases data augmentation) improved performance, especially for Ti-6Al-4V.
• Variability mapping in Ti-6Al-4V showed higher standard deviations of melt pool width/area at higher energy densities; balling zones exhibited relatively low width variability but large area variability due to undercuts around humps. Beads at 200 W and high velocities had smooth boundaries and low width variability. Notably, the machine-recommended P–V combination aligned with the lowest melt pool area variability, indicating increased stability.

The study demonstrates that incorporating temporal features of melt-pool dynamics via video vision transformers enables accurate, generalizable in situ classification of printing regimes and defects without requiring complex temperature-field measurements or machine modifications. The resulting process and variability maps can accelerate printability qualification and process development for new alloys. Temporal context is critical, as periodic phenomena like balling and transient keyhole instabilities may not be evident in single frames or isolated cross-sections. While overall accuracies are high, overlap between regimes (e.g., non-defective keyholing within desirable tracks) can cause confusion between classes, suggesting the utility of multilabel formulations. Performance on Ti-6Al-4V is relatively lower due to its distinct thermophysical properties (e.g., denser plume, higher absorptivity, lower thermal conductivity), but pretrained models and regularization mitigate this gap. The approach’s agreement with ex situ maps and identification of low-variability P–V regions support its practical value for in situ process development.

This work introduces a practical, generalizable in situ method for accelerating process development in metal L-PBF using off-axial high-speed imaging and video vision transformers to capture and exploit melt-pool temporal dynamics. The method achieves high cross-dataset accuracies, produces process maps consistent with ex situ characterization, and enables morphological variability mapping that highlights stable operating windows across alloys. Future research directions include developing multilabel models to handle overlapping regimes, optimizing backbones for improved temporal modeling with limited data, and advancing real-time deployment for closed-loop control by addressing data transfer and buffering constraints of high-speed imaging systems.

• Real-time deployment challenges: high data transfer rates are required to sustain the recording rates needed to capture high-frequency melt-pool dynamics; many commercial high-speed cameras are limited to short recording durations.
• Class overlap: keyholing signatures can appear within otherwise acceptable “desirable” regimes, leading to confusion between classes in some cases; a multilabel framework may better reflect ground truth.
• Material-specific variability: performance is lower on Ti-6Al-4V due to its distinct thermophysical and plume characteristics; although cross-dataset generalization is strong overall, some alloys may require adaptation or pretraining to achieve peak accuracy.
• Top-view limitation: depth information is not directly observed; variability analyses rely on width/area proxies and correlation with ex situ measurements for liquidus–solidus threshold estimation.