Gas-atomized austenitic 304L stainless steel powder was supplied by Jiangsu Vilory Advanced Materials Technology Co., Ltd. The powder exhibited predominantly spherical morphology, and its chemical composition is provided by the company’s product quality certificate, as detailed in Table 1. The particle size distribution ranged from 15 to 53 μm, with a median particle size (D50) of 32.1 μm. Powder characteristics included a hall flow rate of 16.48 s/50 g, apparent density of 4.06 g/cm³, and tap density of 4.90 g/cm³.

LPBF 304L stainless-steel specimens were fabricated on an EOS M290 system (laser spot diameter ≈ 100 μm) under an argon protective atmosphere, with the oxygen content controlled below 500 ppm. To mitigate thermal stress and reduce the risk of warping, the substrate plate was preheated and maintained at 80°C before and during printing. To reduce residual-stress accumulation and anisotropy, the scanning direction of each layer was rotated by 67° relative to the previous layer.

Process parameters were designed using a Doehlert matrix scheme and re-optimized for three discrete layer thicknesses (60, 90, and 120 μm) as determined in the previous study [27]. The distribution of test points and the parameter-space coverage are shown in Fig. 1.

Relative density was quantified from optical cross-sectional images using ImageJ. Pores were segmented, and the pore-area fraction was statistically analyzed to calculate the relative density (%). To ensure the statistical reliability of the density measurements, especially for near-fully dense samples, a comprehensive sampling strategy was employed. For each specimen, 25 non-overlapping optical micrographs were captured at 50× magnification, covering a cumulative inspection area of approximately 120 mm². This extensive surface characterization ensures that the reported relative density is representative of the bulk material. Image processing was conducted using Image J with a consistent global thresholding protocol to distinguish intrinsic porosity from surface artifacts, thereby maintaining the objectivity of the quantitative results.

This work contains two machine-learning tasks: (i) baseline model screening to select robust data-driven learners for density prediction; and (ii) physics-guided residual learning to correct a mechanistic porosity model and improve predictive fidelity.

Quasi-static tensile tests were performed at room temperature using a 50 kN Zwick Proline universal testing machine (Germany) in accordance with the GB/T 228.1-2021 standard. Tensile loading was applied perpendicular to the building direction (BD). Duplicate specimens for each condition were tested to minimize experimental variability. The printing layout for the small square and tensile specimens is shown in Fig. 3.

Phase identification was conducted using a Bruker D8 ADVANCE X-ray diffractometer operated with a scanning range of 30° to 90° (2θ) at a rate of 0.02°/s.

Microstructural analysis and tensile fracture characterization of LPBF 304L stainless steel specimens were performed using a JEOL JSM-IT500HR scanning electron microscope. Specimens were sequentially ground using 240 to 3000 grit SiC abrasive papers and mechanically polished using 0.05 μm SiO₂ suspension for 30 min to obtain a mirror-finish surface. Electrochemical etching was conducted in 10% oxalic acid solution at 5 V and 1 A for 45 s.

EBSD analysis was performed using a ZEISS Gemini500 SEM, employing a step size of 1.02 μm to characterize grain size, crystallographic texture, and geometrically necessary dislocation (GND) distributions. Specimen preparation for EBSD followed the same grinding and polishing protocol as described above, omitting the electrochemical etching step. EBSD data were processed using Oxford Instruments Aztec Crystal software. Grain boundaries were classified as low-angle (LAGBs: 5°–15°) or high-angle (HAGBs: >15°) based on misorientation angles.

Nanoscale microstructural characterization was conducted using a JEOL JEM-2100F field emission transmission electron microscope (TEM). TEM specimens were prepared as 3 mm diameter discs and mechanically thinned to 50 μm thickness. Final thinning to electron transparency was achieved by electrolytic polishing using a Struers TenuPol-5 system with a solution of 5% perchloric acid in ethanol at −20°C. Quantitative analysis of oxide particles and cellular substructures was performed using ImageJ software. Chemical composition analysis was conducted using an energy-dispersive X-ray spectroscopy (EDS) system integrated with the TEM.

To reduce the impact of uneven layer thickness on the evaluation, a layer thickness-based stratification strategy was adopted for data partitioning; the sample sizes of the training/test sets at 60/90/120 μm were 30/24/30 and 7/6/9, respectively, as shown in Fig. 4a,b. This ensured that all three layer thicknesses were covered during the modeling and validation phases. The density distribution showed that the 60 and 90 μm samples were highly concentrated in the 99.8%–100% range, while the 120 μm distribution broadened significantly and showed a low density tail, as shown in Fig. 4c. The density-energy curves fitted by the mechanism showed clear differences in layer thickness: 60 μm rapidly entered a high density plateau, 90 μm slightly declined at the high energy end, and 120 μm showed a significant decrease in the high energy region. This is consistent with the pattern of transformation from under-fusion (LoF) to high-energy unstable deep fusion/keyhole-related defects, as shown in Fig. 4d–f. Based on the mechanism-residual fusion framework, the model achieved R² = 0.833 and MAE = 0.104 on the test set. Under the background of high density ceiling effect, MAE better reflects the actual prediction deviation and is consistent with the physical understanding of the thermal input scale-layer thickness constraint control of melt pool stability and defect window.

Although the proposed mixed mechanistic-residual framework achieves a test R² of 0.833 and MAE of 0.104, Fig. 4 indicates that the density-energy relationship is influenced by additional mechanisms not fully captured by the current two-regime mechanistic from (LoF decay plus keyhole triggering). First, collapsing the process space into a single volumetric energy density E is inherently lossy, since different (P, V, S, H) combinations can yield the same E while producing distinct melt-pool dynamics and defect modes. Second, stochastic process variations (e.g., powder-bed fluctuations, spatter/denudation, local thermal accumulation and remelting variability) introduce heteroscedastic dispersion in density at a given E, which sets an irreducible error floor. Finally, the mechanistic component assumes a single transition scale and threshold per thickness; multi-regime transitions (e.g., unstable melt-pool/balling behavior) may require richer structure. In addition, because the measure densities are concentrated near the high-density ceiling (≈99%–100%), R² becomes more sensitive to small absolute deviations; therefore MAE is reported as a complementary metric to reflect the practical prediction error under the ceiling effect.

Future endeavors will focus on advancing predictive fidelity beyond current R² benchmarks. This will be achieved by enriching the input representation with physically-informed descriptors and interaction terms to better capture defect mechanisms, while evolving the existing dual-regime mechanistic model into a multi-regime or mixture framework to characterize complex transition behaviors. To enhance cross-thickness generalization, we will implement continuous parameterization of P₀(H) and E_k(H) via spline or Gaussian Process (GP) methods. Furthermore, by integrating heteroscedastic residual learning for uncertainty quantification and employing targeted sampling near regime-transition boundaries, we aim to establish a robust, high-fidelity framework for predicting the mechanical properties of LPBF 304L stainless steel.

To elucidate the effect of layer thickness on density and defect formation in LPBF-processed 304L stainless steel, pore distribution characteristics were characterized by optical microscopy across three layer thickness conditions (with three surfaces examined per specimen). Although a full quantitative statistical analysis of morphological parameters (e.g., aspect ratio and roundness) was not performed due to the distinct geometric contrast, the distinction between gas pores and LoF defects remained unambiguous. The former presented as discrete, microsized circular voids, while the latter exhibited large, irregular morphologies with sharp edges. The 60 μm condition achieved high density (porosity ∼0.017%) with predominantly uniformly distributed spherical pores (Fig. 5a,b), indicating stable molten pool dynamics. This thickness facilitated effective laser penetration and heat transfer, enabling controlled molten pool dimensions and enhanced interlayer bonding [23]. At 90 μm layer thickness, porosity increased to ∼0.041% with coexisting gas pores and LoF defects (Fig. 5c,d), suggesting that molten pool expansion increased susceptibility to fusion defects. Thicker powder layers absorbed greater laser energy, enhancing molten pool temperature gradients that promoted pore formation and LoF defects [28]. The 120 μm condition exhibited significantly higher porosity (∼0.417%), characterized by densely distributed pores and extensive LoF regions (Fig. 5e,f). The elevated porosity at 120 μm is primarily attributed to the increased thermal resistance of the thicker powder bed and the attenuation of energy penetration. While a thicker layer can locally prolong melt pool lifetime by reducing the cooling rate toward the substrate, it simultaneously demands significantly higher energy to ensure effective re-melting of the previous layer. In this study, the effective energy density at 120 μm was insufficient to overcome the increase powder volume, leading to a shortened effective metallurgical bonding time and restricted liquid phase capillary flow, which ultimately resulted in extensive LoF defects.

Fig. 6 presents the X-ray diffraction analysis patterns of LPBF 304L stainless steel fabricated with layer thicknesses of 60, 90, and 120 μm. XRD analysis reveals the presence of both austenite and ferrite phases in all samples. Austenite is identified as the primary phase, with ferrite as a minor constituent.

To elucidate the influence of layer thickness on melt pool dynamics and microstructural evolution during LPBF, the horizontal (X–Y) and vertical (X–Z) planes of specimens fabricated with 60, 90, and 120 μm layer thicknesses were systematically examined using optical microscopy and etching techniques. As shown in Fig. 7a, etched horizontal surfaces reveal laser scan tracks with an interlayer rotation angle of approximately 67°, consistent with the preset processing parameter. On etched vertical sections, the melt pool dimensions vary significantly with layer thickness. The smallest melt pool, with a width of 113.3 μm and a depth of 93.5 μm, was observed at a 60 μm layer thickness as shown Fig. 7b. When the layer thickness increased to 90 μm, the melt pool widened to 177.2 μm and deepened to 141.1 μm as shown Fig. 7c. A further increase to 120 μm resulted in a width of 197.5 μm and a depth of 243.7 μm as shown Fig. 7d. Under both the 90 and 120 μm conditions, the melt pool exhibited a gradual widening and deepening. This strong correlation between increased layer thickness and larger melt pool dimensions indicates enhanced heat accumulation, which promotes grain coarsening and expands the heat-affected zone.

Electron backscatter diffraction (EBSD) analysis in Fig. 8 indicates that the sample with a 60 μm layer thickness has an average grain size of 13.5 μm, with LAGBs and HAGBs constituting 27.6% and 72.4% of the total, respectively. When the layer thickness increases to 90 μm, the average grain size increases significantly to 24 μm, accompanied by a sharp rise in the LAGB fraction to 38.6% (and a corresponding decrease in HAGBs to 61.4%). This suggests an increase in dislocation density and the occurrence of incomplete recrystallization. This trend aligns with the findings of Alexander et al., who observed that while the Kernel Average Misorientation (KAM) values—typically associated with local lattice distortion—slightly decreased at higher layer thicknesses, the absolute dislocation density measured by X-ray diffraction actually increased. Specifically, their results demonstrated that increasing the layer thickness from 20 to 80 μm led to a rise in dislocation density from 1.9 × 10¹⁴ m⁻² to 7.2 × 10¹⁴ m⁻² [10]. The underlying mechanism is that a thicker powder layer necessitates higher laser energy input for complete melting, resulting in a larger melt pool volume and an extended solidification time.

This prolonged thermal cycle induces great cumulative thermal stress, which facilitates the generation and multiplication of dislocations [29]. The elevated LAGB fraction (38.6%) in the 90 μm sample is a direct consequence of this increased dislocation density. Instead of being annihilated through post-solidification recovery, these dislocations rearrange themselves into low-energy subgrain configurations within the coarsened grains. This process, driven by the complex thermal history of the LPBF process, manifests as the high density of LAGBs observed in the EBSD orientation maps [30]. For the 120 μm sample, the grain size decreases to 21.48 μm and the LAGB fraction drops to 21.6% (HAGBs: 78.4%), exhibiting a grain boundary character distribution similar to that of the 60 μm sample, as evidenced in Fig. 9.

Transmission electron microscopy (TEM) images in Fig. 10 reveal the presence of cellular subgrains, delineated by dislocation networks, across all samples irrespective of layer thickness (Fig. 10a–c). The statistical analysis of cellular subgrain sizes is summarized in Table 4. To minimize statistical error, measurements were taken from ten distinct TEM images and averaged. The average cellular subgrain sizes measure 467.36, 403.21, and 606.4 nm for layer thicknesses of 60, 90, and 120 μm, respectively. These dense dislocation networks act as potent barriers to dislocation glide, constituting a key strengthening mechanism that contributes to the ultra-high YS observed in LPBF-processed stainless steel. This strengthening effect can be rationalized by the Hall-Petch relationship, where the cellular subgrain size effectively functions as the Hall-Petch grain size [31]. The formation of these cellular subgrains is attributed to a nonlinear self-organization phenomenon, driven by the complex thermo-physical conditions within the LPBF melt pool. The intense laser irradiation generates substantial temperature and surface tension gradients within the melt pool, thereby inducing vigorous Marangoni convection and triggering Bénard instability [32]. This fluid flow results in the formation of minute, ordered vortex cells within the solid-liquid two-phase zone at the solidification front. These convective cells template the spatial distribution of solute elements, thereby guiding the subsequent solidification process. Subsequently, under rapid cooling conditions, this transient cellular morphology is rapidly solidified, resulting in the formation of cellular subgrains comprising solute-enriched boundaries and dislocation networks [33]. When the layer thickness increases from 60 to 90 μm, the altered melt pool geometry and steeper thermal gradients enhance the Bénard instability. This, in turn, refines the scale of the convective vortex cells, leading to the templated formation of a finer solidification structure. This phenomenon represents a nonlinear dynamic process wherein microstructural evolution is not governed by a single thermodynamic parameter but arises from the complex coupling of multiple fields, notably the temperature and flow fields [34]. Conversely, a further increase in layer thickness to 120 μm leads to a significant coarsening of the cellular subgrains. This reversal in trend indicates that within this parameter regime, thermodynamic factors begin to dominate over fluid dynamic effects. A greater layer thickness necessitates deeper energy penetration, resulting in a larger melt pool volume and a significantly prolonged liquid phase lifetime [35]. Consequently, the overall cooling rate is markedly reduced, the solidification front velocity (R) decreases, and the thermal gradient (G) is lowered owing to increased heat accumulation. According to solidification theory, a decreased cooling rate and a lower G/R ratio (morphological parameter) promote coarsening of the as-solidified microstructure [32]. The extended duration in the liquid state provides enhanced kinetic conditions for cellular subgrain growth and coalescence, ultimately resulting in the observed microstructural coarsening.

Fig. 10 presents TEM images characterizing both the cellular subgrains and the oxide particles, with the size and number density of the latter quantified. The corresponding quantitative data are summarized in Table 4. Respectively, the average oxide particle sizes were measured to be 43.02, 43.21, and 64.23 nm for layer thicknesses of 60, 90, and 120 μm. The number density decreased markedly from 2.484 μm⁻² (60 μm layer) to 2.034 μm⁻² (90 μm layer), and further to 1.272 μm⁻² (120 μm layer). This particle coarsening is directly attributable to the altered thermal history of the melt pool, as described previously. An increase in layer thickness prolongs the melt pool lifetime and its residence time at elevated temperatures. This extended duration provides sufficient kinetics for diffusion and growth of nanoparticles nucleated in the liquid phase, thereby promoting Ostwald ripening—a process wherein smaller particles dissolve to feed the growth of larger ones. Further elemental analysis of the matrix and particles was conducted using energy-dispersive X-ray spectroscopy (EDS). Fig. 11 presents the elemental distribution maps obtained from the sample surface. Combined with the EDS point analysis results summarized in Table 5, the matrix is significantly depleted in Si (0.47 wt.%) and O (0.16 wt.%) compared to the Si- and O-rich particles, which exhibit two distinct contrast phases (bright and dark) under TEM. Both types of particles are predominantly composed of silicon and oxygen. The bright-contrast particles exhibit higher concentrations of Si (12.30 wt.%) and O (26.67 wt.%), whereas the dark-contrast particles are enriched in Cr (14.86 wt.%), suggesting they are primarily a Cr-silicate phase. The particles are distributed both intragranularly (within cellular subgrains) and intergranularly (along subgrain boundaries), indicating a uniform dispersion throughout the microstructure.

Room-temperature uniaxial tensile tests were performed on LPBF 304L stainless steel specimens fabricated with layer thicknesses of 60, 90, and 120 μm to quantify the effect of this parameter on mechanical properties. To ensure full data transparency and represent the specific material response within the limted sample size (two duplicates per condition), individual raw values for each specimen are reported. As shown in Fig. 12, the specimen with a 60 μm layer thickness demonstrates the optimal combination of strength and ductility. Its tensile properties were recorded as 696 and 695 MPa for UTS, 532 and 530 MPa for YS, and 49.5% and 50.5% for EL. For the 90 μm specimen, the properties showed a noticeable decreased to 669 and 663 MPa for UTS and 505 and 499 MPa for YS, while the EL was 48.5% for both specimens. A further increase in layer thickness to 120 μm resulted in a more pronounced drop in strength to 656 and 645 MPa for UTS and 476 and 467 MPa for YS, with an EL of 49.5% and 44.5%, respectively. Notably, although increasing the layer thickness leads to a systematic reduction in strength (UTS by >6%, YS by >11%), the concomitant loss in ductility is relatively limited (EL by only 6%). This indicates that the process utilizing a larger layer thickness retains potential for engineering applications.

Fig. 13 shows scanning electron microscopy (SEM) fracture images of tensile specimens, revealing the failure modes of samples with varying layer thicknesses. Macroscopically, all specimens exhibit a visible 45° shear lip and necking band. Microscopically, a uniformly distributed dimple structure is observed, indicating a typical ductile fracture. Oxide particles are present at the bottom of the dimples, acting as nucleation sites for microvoid formation. The increase in dimple size with increasing layer thickness coincides with a decreasing elongation after fracture, but the ductile nature of the fracture remains unchanged.

In LPBF, process parameters synergistically dictate the volumetric energy density (VED), which in turn governs melt pool dynamics, cooling rates and defect formation. While increasing laser power enhances density [36], excessive scanning speeds or reduced hatch spacing can induce spattering and heat accumulation, leading to residual stresses or microcracks. It should be noted, however, that VED is a convenient first-order scalar indicator of global heat input but cannot fully represent the melt-pool geometry (e.g., width/depth and stability), the effective powder/solid absorptivity, or the extent of multi-track/multi-layer remelting and heat accumulation. Different combinations of (P, V, S, H) may yield similar VED values while producing distinct melt-pool shapes and defect modes; therefore, VED-based interpretation should be considered together with process-specific descriptors and the microstructural evidence discussed below. For the studied conditions, VED followed the order: 120 > 90 > 60 μm. Despite its high VED, the 120 μm sample exhibited the lowest density due to excessive layer thickness, which destabilizes the melt pool and increases susceptibility to LoF and keyhole defects [37]. Conversely, the 60 μm sample achieved the highest density as shown in Table 3, providing a robust foundation for superior mechanical strength by mitigating stress concentrations [38].

Microstructure analysis reveals a non-monotonic trend in grain evolution as shown in Fig. 9. The 60 μm sample exhibited the finest grain size (13.5 μm) due to high cooling rate [23]. Interestingly, while 60 and 120 μm samples possessed high HAGB fractions (72.4% and 78.4%) to hinder crack propagation, the 90 μm sample showed the highest proportion of LAGBs (38.6%). These LAGBs acted as dislocation storage sites, maintaining a higher yield strength (502 MPa) compared to the 120 μm sample [39,40]. This anomaly at 90 μm is attributed to a critical heat input threshold that triggers abnormal grain coarsening via strain-induced boundary migration and intense thermal cycling [41–43]. As thickness reaches 120 μm, the transition to a deep keyhole mode induces intense Marangoni convection and fluid agitation, promoting grain refinement despite the complex temperature fields [44–46].

This non-monotonic trend suggests that the effect of layer thickness on grain size is governed by mole pool dynamics and thermal gradients. To provide a comprehensive synthesis of these coupled effects, Fig. 14 delineates the mechanism evolution process across three characteristic regimes (60–120 μm). Conduction mode (60 μm): at a thinner layer thickness, the process is dominated by high cooling rates and rapid heat dissipation. This stable conduction-mode melting effectively suppresses excessive grain growth and minimizes the propensity for pore formation, resulting in a refined microstructure with high relative density. Transition window (90 μm): as the layer thickness increases, the heat input approaches a critical threshold where the material undergoes intensified thermal cycles and “thermal tempering” effects. These thermodynamic conditions trigger abnormal grain coarsening or coalescence, while reduced cooling rates lead to the emergence of irregular LoF defect clustering. Keyhole mode (120 μm): under the highest energy density, the laser-material interaction transitions into a deep and narrow keyhole mode, driven by significant vapor recoil pressure and strong Marangoni convection. This dynamic instability generates a complex thermal field that interrupts columnar grain growth-leading to a mixed morphology-but simultaneously elevates the risk of keyhole-induced porosity.

In summary, this systematic evolution reveals that the stable conduction/transition window is the primary optimization target for tailoring the microstructure and enhancing the performance of LPBF 304L stainless steel.

As summarized in Table 4, the average cellular subgrain sizes for the 60 and 90 μm samples are comparable, measuring 467.36 and 403.07 nm, respectively. This similarity is attributed to the steep thermal gradients generated by the high laser power and optimized scanning speed, which promote the rearrangement of dislocations into well-defined, fine subgrains. In contrast, the 120 μm sample exhibits the largest cellular subgrain size, averaging 606.4 nm. The excessive layer thickness results in a reduced cooling rate, enabling substantial dislocation recovery and subsequent subgrain coarsening. This process diminishes the dislocation pinning effect [47]. The 60 μm sample contains the highest number density of oxide particles, which are also the smallest in size. These finely dispersed oxides effectively hinder dislocation motion via the Orowan strengthening mechanism. Although the oxide particles in the 90 μm sample are comparable in size to those in the 60 μm condition, their lower number density results in marginally inferior mechanical performance. The 120 μm sample exhibits the lowest oxide number density and the largest particle size, culminating in a significantly weakened strengthening effect [48].

In summary, an optimal volumetric energy density window exists for each layer thickness. Deviation from this optimal range, whether too high or too low, promotes defect formation [37]. Increasing the layer thickness necessitates the concurrent optimization of scanning speed, hatch spacing, and laser power to maintain melt pool stability. A high fraction of HAGBs, combined with Hall-Petch strengthening from fine grains, is crucial for enhancing both strength and ductility [49]. A high density of nano-oxides improves strength by effectively pinning dislocations, provided that particle coarsening is avoided. Consequently, the superior ductility of the 60 μm sample originates from its fine grain size and high HAGB fraction, whereas its high strength is derived from the combined effects of a high density of nano-oxides and fine cellular subgrains. Although coarse grains in the 90 μm sample weaken the strength, this is partially compensated for by a high fraction of LAGBs and fine cellular subgrains, albeit at the expense of reduced ductility. The 120 μm layer thickness is excessive, resulting in decreased density, coarsened cellular subgrains, a reduced number density of oxide particles, and increased particle size. Thus, further development and optimization of process parameters are required.

The contributions of various strengthening mechanisms to the YS of LPBF 304L stainless steel at different layer thicknesses are summarized in Fig. 15. The YS (σ_y) of polycrystalline metallic materials can typically be described as the linear sum of contributions from various strengthening mechanisms. For the LPBF 304L stainless steel investigated in this work, the overall YS can be expressed by the Eq. (8):(8)σy=σ0+σss+σgb+σd+σp,where σ₀ is the lattice friction stress, also referred to as the Peierls–Nabarro stress, representing the intrinsic resistance to dislocation glide in a perfect crystal lattice. For a given material such as 304L stainless steel, this term is constant. σ_ss denotes solid solution strengthening, which arises from lattice distortions induced by solute atoms that impede dislocation motion. Since the chemical composition is consistent across all three sample groups in this study, σ_ss is also considered constant. σ_gb, σ_d, and σ_p represent the contributions from grain boundary strengthening, dislocation strengthening, and second-phase particle strengthening, respectively.

Grain boundary strengthening can be described by the classical Hall–Petch relationship [50], as shown in Eq. (9):(9)σgb=kHPd,where k_HP is the Hall–Petch coefficient and d is the average grain size. This relationship describes the enhancement of YS resulting from grain refinement. For 304L stainless steel, the value of k_HP is taken as597.6 MPa·μm^0.5. Microstructural analysis reveals a predominance of fine equiaxed grains, with average grain sizes of 13.5, 24, and 21.48 μm for the 60, 90, and 120 μm layer thickness samples, respectively. Accordingly, the grain boundary strengthening contributions for these three samples are calculated to be 162.7, 122, and128.9 MPa, respectively.

The thermal stresses induced during the LPBF process result in the formation of a high-density dislocation network within the material, which significantly impedes dislocation glide during plastic deformation [51]. The contribution of dislocation strengthening is commonly described by Taylor’s hardening law, as shown in Eq. (10):(10)σd=αMGbρ,where α is the dislocation strengthening coefficient (taken as 0.3), M is the Taylor factor (M = 3.06), G is the shear modulus (78 GPa for austenitic stainless steel), b is the magnitude of the Burgers vector (0.25 nm), and ρ is the dislocation density [52,53]. Since the boundaries of the cellular subgrains comprise dislocation walls formed by densely entangled dislocations, the average cellular subgrain diameter, d_c, is intimately related to the overall dislocation density, ρ [54]. This relationship can be quantified using an empirical formula as shown in Eq. (11):(11)ρ≈Cdc2,where C is a constant, typically assigned a value of 3 in additive manufacturing literature to represent cellular structures. It should be noted that this calculation serves as an order-of-magnitude estimation rather than an absolute measurement of dislocation density, as the precise value of C may vary depending on the specific dislocation configuration and boundary characteristics. Nevertheless, this approach provides a consistent semi-quantitative basis for comparing the dislocation strengthening contributions across different processing conditions. Based on the measured cellular subgrain sizes provided in Table 4, the corresponding dislocation strengthening contributions are calculated to be 66.5, 77.1, and 51.3 MPa for the 60, 90, and 120 μm samples, respectively.

The LPBF process results in the formation of nanoscale oxide particles. When dislocations encounter these particles, they must bypass them rather than cut through them, requiring additional stress and thereby producing a strengthening effect. This contribution, known as Orowan strengthening, as shown in Eq. (12):(12)σp=0.4MGbπλ1−vln⁡dpb,where d_p is the average oxide particle diameter, ν is Poisson’s ratio (ν = 0.3), and λ is the effective interparticle spacing. In all calculations, d_p, b, and λ were expressed in the same length unit. Since experimental measurements typically provide the number of particles per unit area (N_A) in a two-dimensional plane, the following relationships are used to convert the areal density into a volume fraction f_v, which in turn is used to estimate λ. Assuming the oxide particles are spherical with an average diameter d_p, the number of particles per unit volume N_v is related to N_A through the stereological approximation, as shown in Eq. (13):(13)Nv=NAdp,where N_A has the unit of μm⁻² and N_v has the unit of μm⁻³.

The particle volume fraction f_v can then be written as shown in Eq. (14):(14)fv=Nv⋅πdp36=πNAdp26,which is dimensionless.

For a random particle distribution, the mean center-to-center spacing on the section is approximated as L=1/NA. Therefore, the effective edge-to-edge interparticle spacing used in the Orowan model is shown in Eq. (15):(15)λ=L−dp=1NA−dp=dp(π6fv−1).

Using the particle size and areal density data from Table 4, the Orowan strengthening contributions are calculated to be 77.6, 69.9, and 61.1 MPa for the 60, 90, and 120 μm samples, respectively.

In summary, the high strength of the LPBF-processed 304L stainless steel is primarily attributed to grain boundary strengthening and dislocation strengthening. In comparison, the strengthening contribution from the native oxide particles is relatively modest.

Table 6 presents the tensile properties of the LPBF 304L stainless steel produced in this study, alongside a comparison with previously reported mechanical properties of #300 series stainless steels fabricated via LPBF at varying layer thicknesses. Fig. 16 provides a graphical summary comparing the UTS vs. EL, as well as YS vs. EL, based on the data compiled in Table 6. The results clearly indicate that the material developed in this work achieves either superior mechanical performance at comparable production efficiency, or higher production efficiency at equivalent performance levels, when compared to existing studies, thereby striking an optimal balance between performance and manufacturing efficiency.