Experimental and numerical investigations of oxide-related defects in Al alloy gravity die castings

A primary AlSi7Cu0.5Mg cast alloy (EN AC-45500 [28]) in the form of ingots and a permanent steel mould, designed according to CEN/TR 16748:2014 standards [29], were selected for the casting trials. The chemical composition of the alloy is shown in Table 1.

The die, illustrated in Fig. 1, was made of AISI H11 tool steel, and it enabled to cast a cylindrical-shaped bar with a total length and mean diameter of 190 mm and 15 mm, respectively (Fig. 2). Before the casting trials, a semi-permanent layer of Foseco Dycote® F34 coating [30] was sprayed over the die walls at a temperature of about 200°C, according to standard foundry practise. To ensure the good reproducibility of the tests and monitoring the variation of the mould temperature during the trials, a K-type thermocouple (Ø1 mm) was embedded into the die, and it was located in the middle of the cylindrical-shaped bar, 2 mm from the surface of the cavity.

The AlSi7Cu0.5Mg cast alloy was melted at 780 ± 5°C into a silicon carbide (SiC) crucible in an electric resistance furnace. After melting, the temperature of the bath was decreased to 750 ± 5°C for the degassing phase. To increase the quality of the molten metal and to remove impurities and old oxides, the bath was degassed with argon for 30 min and accurately skimmed with a coated paddle. This procedure ensured that the defects detected in the final castings were primarily correlated to the presence of young bifilms that were generated during the pouring and filling phases.

Afterwards, the temperature of the molten metal was further decreased to 725 ± 5°C and the die was homogeneously pre-heated at 400 ± 3°C in an electric resistance muffle furnace.

The melt was manually poured in 3 s at 725 ± 5°C, causing the die temperature to increase from 400 ± 3°C to 450 ± 2°C. The flow path of the molten metal was captured through the aperture of the riser using a high-speed camera with a sampling rate of 240 frames per second and a shutter speed of 1/8000 s.

A batch of 10 castings was produced, taking care to gently stir and manually skim the molten metal before any casting. Filtering operations, as well as chemical eutectic modification or grain refinement, were not performed in the present work. The selected casting parameters and the absence of the filter promoted high surface turbulence during the filling stage and induced the formation of entrainment defects inside the castings.

For the metallographic investigations, samples were drawn from the cross-section of the castings as shown in Fig. 2. The samples were mechanically prepared to a 3-μm finish with diamond paste and, finally, polished with a commercial 0.04-μm silica colloidal suspension. The polished specimens were then etched in a solution of 5 vol% HF and 95 vol% H₂O to better evidence the entrained oxide inclusions [31].

Microstructural observations were carried out using an optical microscope (Leica® DMLA) and a field emission gun scanning electron microscope (FEG-SEM, FEI® QUANTA 250) equipped with an energy-dispersive spectrometer (EDS, EDAX®). To quantify the microstructural features, the optical micrographs were processed using image analysis software, which was focused on the secondary dendrite arm spacing (SDAS) measured by applying the line intercept method [32].

The oxide-related defects were mapped throughout the entire cross-sections indicated in Fig. 2 by exploiting the automatic handling of the microscope stage. Contiguous micrographs, each with an area of 1.2 × 0.9 mm², were automatically collected throughout the whole section. The severity of the entrained oxide inclusions was evaluated for each field based on the reference micrographs, as shown in Fig. 3. Five severity grades (SGs) were defined according to the area of defects, which was quantitatively analysed using the image analyser. The total measured defect area in the micrograph was then divided by the initial area of the micrograph to determine the area fraction covered by the oxide-related defects (A_D). As shown in Fig. 3, each SG was associated with a different colour and a range of A_D values. The A_D ranges were selected to emphasise the low damage conditions. In general, a region containing few closed bifilms or porosity, referring to gas precipitation within a bifilm, with a maximum size of about 20 μm was categorised as grade 1. Regions showing coarser unfurled bifilms and porosity with a maximum size of 300 μm were categorised as grade 4.

The result of mapping the oxide-related defects throughout a cross-section is shown in Fig. 4a. Furthermore, the distribution of the oxide-related defects, representative of the experimental conditions, was obtained as a statistical average of the sections drawn from five different castings, as shown in the coloured map in Fig. 4b.

To analyse the distribution of the entrained oxide inclusions, a convolutional neural network (CNN) technique was applied. CNN is a deep learning algorithm designed for image analysis; it is based on the convolution and pooling operations at the pixel level. The convolution operation was used to extract a feature map from the image by using multiple filters (kernels), while the pooling operation, also called subsampling, permitted the reduction of the dimensionality of the feature map and, therefore, the local sensibility of the image [33].

A typical 2×2 kernel was defined with a stride length equal to 2, that is, the filter moved two pixels on the input image at one time. Groups of four (2 × 2) contiguous micrographs were generated, representing a surface area of about 4.3 mm² (2.4 mm × 1.8 mm). Therefore, an average pooling operation was carried out where the SG assigned to each kernel was equal to the average value of the micrographs’ group, as shown in Fig. 4c. This procedure resulted in a lower resolution version of the mapping, which still contained the important structural elements without the unnecessary fine details.

In the present work, the numerical analysis was performed using FLOW-3D, which is a CFD code based on the finite volume discretization method [34]. This code uses the Fractional Area-Volume Obstacle Representation (FAVOR) methodology to model geometries, and the volume of fluid (VOF) method to track the free surfaces. Detailed descriptions of these algorithms can be found in [35, 36].

A cumulative scalar technique existing in the commercial software was used to predict the formation of entrainment defects [27]. At each time step, the concentration of the oxides (O_x) accumulated in a cell of the mesh is provided according to the following equation:where F_s is the metal surface area in contact with the atmosphere, t is the exposure time and DFTSRF is the calibration coefficient that considers the chemical composition of the alloy and the specific weight of the oxide.

Fluid fluctuations or convective motions can change the value of the oxide concentration accumulated in a cell of the mesh. When these conditions are present, a transport equation is numerically solved using a second order scheme that preserves monotonicity; in this way, a scalar value can diffuse from a cell to adjacent ones. The diffusion rate is controlled by a molecular diffusion coefficient, which is a function of the dynamic viscosity and density of the fluid [34]. This surface defects model allows for predicting the distribution of casting defects due to surface turbulence.

FLOW-3D CAST v5.1 (2020) commercial software [34], with its workspace for gravity die casting, was used to simulate the filling and solidification phases of the casting process. The three-dimensional (3D) computer-aided design model of the die was drawn and imported into the simulation software. The physical properties of the die were defined as a linear function of the temperature in the working temperature range of 400–450°C. The thermal conductivity varied from 34.2 to 33.4 W/(m·K), while the specific heat was included in the range 620–650 J/(kg·K).

The properties of the alloy were set as a function of the temperature ranging from 25 to 750°C. The solidus and liquidus temperatures were imposed at 540°C and 616°C, respectively. The density ranged from 2690 to 2410 kg/m³, the thermal conductivity was included between 176 and 85 W/(m·K), while the specific heat varied from 886 to 1160 J/(kg·K).

The interfacial heat transfer coefficient (IHTC) was approximated by two constant values for completely liquid and solid states, while it followed a linear variation between these two values during the phase transition interval. In particular, the IHTC value was set equal to 2700 W/(m²·K) and 1000 W/(m²·K) above the liquidus and below the solidus temperatures, respectively. The IHTC controls the heat exchange at the metal–mould interface, and especially due to the progressive formation of an air gap between the die wall and the casting part, its value decreases during the solidification stage. Different models of IHTC are used during the casting simulation, e.g. the approximation with a constant value, the linear or exponential correlation with the casting temperature at the interface, the empirical Lewis–Ransing correlation [37, 38]. The model used in the present work is well known and diffused in numerical casting simulation because it provides a good compromise between reliable results and reduced calculation time.

To increase the reliability of the numerical simulation, the process parameters used during the experimental procedure were imported. Therefore, the pouring and die temperatures were 730°C and 400°C, respectively. The flow rate was set constant at 9.1·10⁻⁵ m³/s in the first 2.6 s, and then it progressively decreased in the remaining filling time (0.4 s).

The following models were used during the casting simulation:

Virtual vents were located along the parting line of the die to model the leakage of air.

A mesh of 870,000 cubic cells with a uniform grid size of 1.25 mm was automatically generated by the software for the whole system (die and casting); a mesh of 330,400 cubic cells was made for the die cavity.

The free surface defects concentration was the software output used to analyse the numerical distribution of oxide inclusions. The model was previously described in Section 2.

To ensure the reliability of the experimental results, the repeatability of the casting process was first evaluated. The frames, captured during the casting trials, were analysed to evaluate the filling phase in the different tests.

Figure 5 shows the images at 0.24 s and 0.50 s from the beginning of the pouring phase. While Fig. 5a illustrates the initial metal flow inside the cavity of the test bar, the frames in Fig. 5b refer to the initial filling of the riser, which seems to begin in the zone that is furthest from the ingate.

In general, a good correspondence in the metal position and the die filling level appears evident, highlighting a similar fluid-dynamic behaviour in the different experimental tests. This consistency in the empirical results assures the validity of the collected data.

The accuracy of the numerical code to simulate the fluid-dynamic behaviour of the liquid metal was verified by comparing the flow path between the actual experiment and the numerical solution. Figure 6 shows the comparison between the numerical simulation result (right) and the flow visualisation experiment (left) through the open-type riser; the flow pattern inside the die cavity refers to different times from the beginning of the pouring phase.

Figure 6a shows the flowing metal inside the cavity of the test bar at 0.20 s from the beginning of the casting. The numerical prediction was in agreement with the experimental findings; the position of the fluid front, as well as the fluid path inside the die cavity, was comparable.

The liquid metal flowed along the cavity of the test bar and struck the far side of the die cavity, generating a returning wave (Fig. 6b). The flow front folded over itself and flowed back towards the ingate. This folding mechanism results in the formation of a bifilm, where the two sides of the solid oxide film can entrap a small volume of air/gas between them while surrounded by the liquid melt.

At 0.42 s from the beginning of the pouring, the liquid metal started to fill the region of the riser where two different fluid fronts flowing in opposite directions were generated (Fig. 7). The first front was given by the metal that rose from the middle zone of the passage linking the riser to the cavity of the test bar, then flowing towards the distant side of the die cavity (indicated as #1 in Fig. 7). The other flow front was the returning wave that was previously mentioned (indicated as #2 in Fig. 7). When the two fronts meet and join each other, the oxide skin covering the surface of the liquid can be partially entrained inside the bulk liquid, thereby promoting the formation of the oxide-related defects.

The die filling proceeded with the accumulation of liquid Al in the zone of the riser far from the ingate, visible in both the simulated and experimental results at 0.54 s and 0.90 s (Fig. 6c, d), respectively. From 1.24 s, the action of the unzipping waves completed the filling phase (Fig. 6e). The molten metal flowed under the surface skin, gradually increasing the level of liquid inside the die cavity. This filling mode mitigated the entrapment of bifilms because the surface film was not very perturbated.

As seen in Fig. 6e, f, some differences appear in the flow path between the actual experiment and the numerical results. While the fluid ripples were numerically evident, during the experimental tests they could only be slightly observed below the oxidised surface. This difference was related to the difficulty of the numerical code in modelling the surface oxide skin, which hid the fluctuations of the liquid metal. At 5.20 s (Fig. 6f), although the die cavity was completely filled, the fluid was not at rest: flow fluctuations could still be detected, especially inside the riser. These movements are clearly visible in the numerical results seen in Fig. 6f, and they gradually weakened over time until they stopped completely 25 s after the beginning of the pouring phase.

Even though oxide-related defects are primarily formed during the filling phase, their final distribution inside the casting is also influenced by the solidification stage. A premature or delayed solidification of the material affects the flow motions that are responsible for the redistribution of the defects. To increase the reliability of the numerical simulation, the simulated heat transfer has to be consistent with the empirical one. Thus, the solidification behaviour of the casting was studied to evaluate the reliability of the thermal parameters set in the numerical simulation and to ensure reliable modelling of the heat exchange. In particular, the correspondence between the experimental and numerical cooling rates was evaluated.

In the present work, the mean cooling rate (R) was estimated according to the empirical relation

SDAS = 39.4·R^−0.317 proposed by Wang et al. [42] for Al-Si foundry alloys, where SDAS represents the scale of the primary α-Al phase. In the central area of the test bar of both the investigated sections, the average value of SDAS was equal to 27 ± 3 μm, corresponding to a cooling rate of approximately 3.3°C/s.

To assess the simulated cooling rate, a region of interest was defined in correspondence of the gauge length, and the variations in the cooling rate were plotted as a function of time. The mean cooling rate of the primary α-Al dendrites during solidification was equal to 3.3°C/s, which is in agreement with the experimental data. This could also ensure the reliability of the thermal parameters set in the numerical simulation and the accuracy in the heat exchange modelling. Figure 8 illustrates the numerical distribution of the cooling rate in the section next to the ingate (A-A) at the dendritic coherence point, when the solid fraction was approximately equal to 0.18, and the temperature was about 608°C [43].

The cooling rate decreased with the distance from the die walls and the thickness of the casting. It ranged from about 0.8°C/s in the upper central zone of the riser to a maximum value of about 10°C/s in some areas of the test bar that were in contact with the die surface. In the central region of the test bar, the cooling rate ranged between 2.5 and 5°C/s, with an average value of about 3.3°C/s. The same distribution of cooling rate was detected in section B-B, which was about 155 mm from the ingate.

The casting defects generated by entrainment of the bifilms were the typical casting defects detected throughout the microstructure, as shown in Fig. 9. Both large and small porosities were contoured by a thin oxide layer (Fig. 10a), as revealed by the EDS analyses (Fig. 10b, c). High levels of oxygen were identified in correspondence of the film, consistent with oxide entrapment [7]. Furthermore, the reduced thickness of the film, here observed typically in the range of 0.1 to 0.6 μm, demonstrates that its formation began in the pouring or filling stages. No thick oxides (i.e. old oxides) were observed, indicating the high efficiency of preliminary degassing used in the present work.

The presence of thin oxide layers contouring porosities suggested that they acted as formation points for the growth of solidification defects (see Fig. 9). The unbounded oxide surfaces of a bifilm can be easily separated with minimal gas pressure or minimal stress, forming pores or cracks. In contrast, the nucleation of volume defects, as an atomic-sized event, is extremely difficult due to the perfect atomic contact and the strength of the boundaries [7].

Figure 11 shows the distribution of the SGs related to the casting defects generated by the entrainment of bifilms in two different cross-sections of the castings after application of the CNN method. Section A-A was located close to the ingate (35 mm), while section B-B was about 155 mm from the ingate.

In both sections, the area fraction covered by the oxide-related defects increased from the region of the test bar towards the free surface of the riser. In particular, the average SG value in the tensile test sample was less than 1, and it gradually increased up to local values of about 3 in the upper part of the riser. The main difference detected between the two sections was seen in the lower zone of the riser. Slightly higher values of A_D, which means more oxide-related defects, were identified in the section next to the ingate.

To analyse the distribution of the oxide-related defects in detail, each cross-section was subdivided into three different regions (Fig. 12): the upper side of the riser (UR), the lower side of the riser (LR) and the zone of the cylindrical test bar (TB). To obtain a statistical average of the distribution, the number of measured fields corresponding to the different SG values was evaluated and listed in Table 2. The average SG was then calculated as weighted arithmetic mean in the different subdivided cross-sections.

The statistical analysis reported in Table 2 showed that the SG in the investigated sections was relatively low (lower than 1.7). These low values were not related to the presence of a few coarse defects; rather, they were linked to the formation of a large number of reduced-size defects. In detail, in the investigated cross-sections there were 1104 (~1192 mm²) grade 1 micrographs in comparison to 229 defect-free zones (~247 mm²). Moreover, only 177 of the 2193 fields showed A_D values between 0.8 and 1.5% (SG 3 or SG 4), and they were mostly located in the upper side of the riser. This is reasonable to expect from the casting geometry because the riser filled with material that has already flowed through the die cavity and the major part of entrained defects were accumulated in the “dead” riser.

However, the lack of old thick oxides confirmed the efficiency of the degassing treatment that was used during the experimental procedure.

To emphasise the entrainment of the oxide-related defects in the casting, filtering systems were not used in the present work. Although some studies [5, 26, 44] revealed how oxides that are micrometres thick are hardly trapped by millimetre-pore–sized filters, these remain highly effective to reduce the velocity of the melt flow and to control over surface turbulence during the filling process [5, 45].

To numerically investigate the entrainment of oxide inclusions, the filling velocity and the concentration of free surface defects, obtained from the simulation software, were considered. The first output was related to the action of the surface turbulence, which causes entrainment phenomena. The entrapment of bifilms and air/gas bubbles inside the bulk liquid is emphasised by a fluid velocity greater than the critical value for the generation of turbulence in Al alloys (0.5 m/s) or the formation of vortexes during the filling phase [5].

Figure 13 shows the variation in filling velocity over time from the beginning of the pouring phase at half width of the casting.

From the beginning of the filling phase, the fluid flow was turbulent. The metal velocity at the bottom of the pouring channel overcame the critical value for the generation of turbulence (Fig. 13a). The values remained high for the entire length of the bar cavity until the impact with the die surface at 0.40 s (see Fig. 13b). When the returning wave was generated and the metal began to flow inside the riser (Fig. 13c), the variation of cross-sectional area in the passage linking the bar cavity to the riser decreased the fluid velocity inside the riser. Figure 13d and e show how the metal velocity decreased from 0.4–0.7 m/s inside the bar cavity to 0.05–0.30 m/s inside the riser.

As seen in Fig. 13d and e, the effects related to the absence of a filter (not used in the present work) are visible. At 1.65 s, a vortex formed in the filter housing, which could become particularly damaging at the beginning of the filling of casting cavity. The combined action of air, which oxidises the molten metal, and the recirculation motions that entrap the oxide layers inside the bulk liquid, generally favours the formation and entrainment of bifilms. Furthermore, the fluid velocity downstream of the filter zone was still high because the absence of the filter did not facilitate the transition from a turbulent flow to a laminar regime flow.

At 5 s (Fig. 13f), although the die filling phase was completed, the fluid was not at rest; some metal fluctuations still existed, resulting in a non-stationary flow regime. These motions slowed as the time and solid fraction increased (Fig. 14) until 25 s from the beginning of the pouring phase when the flow velocity was equal to zero throughout the casting.

Figure 14 shows the distribution of the metal velocity and solid fraction in the investigated sections at 15 s. The flow velocity was at rest in the test bar’s cavity and in the lower zone of the riser, while slow recirculation motions were still detectable in the upper side of the casting. These vortexes promoted the backflow of metal from the free surface towards the inner region of the riser, as indicated by arrows in Fig. 14.

The second software output used to numerically investigate the entrainment phenomena was the concentration of free surface defects, whose distribution throughout the casting at the end of die filling is shown in Fig. 15. High values of defects could be observed in the zone of the riser that was in contact with the atmosphere, in particular in the regions of the upper side of the casting. The oxide concentration gradually decreased towards the inner zones of the casting; in particular, the cylindrical test bar appeared to be the soundest region. This distribution was depicted using the model described in Section 2. A scalar value, representative of the concentration of oxide defects, was accumulated in the cells where a surface of liquid metal was in contact with the atmosphere (free surface). The accumulation rate was constant over time, and the scalar could be transported from a cell to adjacent cells because of the flow movements or the convective motions.

In Fig. 15, the high accumulation of defects in the upper part of the open-type riser can be explained by the metal flow in the die cavity and the exposure time with air. Oxide defects were mainly accumulated at the top of the riser after that the liquid metal flowed through the whole die cavity. A longer flow path leads to a lower metal temperature and a greater oxidation time, which increases the probability of forming oxide skins, inclusions and cold laps.

Furthermore, the oxides that accumulated at the free surface of the riser could move towards the inner regions of the casting, especially at the farthest ends of the casting. This mechanism was due to the non-stationary flow regime at the end of the cavity filling phase. The presence of slow recirculation motions, indicated by arrows in Figs. 14 and 15, involved the redistribution of the oxide-related defects until 25 s from the beginning of the pouring phase when a steady-state condition was reached. In particular, along the investigated cross-sections, the decreasing gradient in the concentration of defects from the riser to the test bar was related to the vortexes illustrated in Fig. 14a, which promoted the oxide’s movement from the upper part of the riser to the lower part of the riser.