Inluence of solidiication cell structure on the martensitic transformation in additively manufactured steels

A key feature when using martensitic steels is the proportion of retained austenite present in the inal compo- nent. Martensitic steels manufactured by laser powder-bed fusion (LPBF) have been shown to have more retained austenite than when conventionally manufactured. The LPBF microstructure is characterised by small grains containing ultraine solidiication cells (<1μm). This study shows that the solidiication cells can fully suppress thermal martensite. The retained austenite is highly metastable, and will readily transform to deformation martensite either in-build from thermal strain or post-build from deformation. This raises concerns around sample preparation methods causing incorrect phase quantiication in LPBF-built martensitic steels.


Introduction
17-4 PH is a precipitation-hardened stainless steel, widely used in both additive and traditional manufacturing. Under conventional processing it is martensitic, but the laser powder-bed fusion (LPBF) literature reports a dual-phase structure of α'-martensite with a signiicant proportion of retained γ-austenite. Diferent studies have reported a wide variation in the relative phase fractions, even when using similar build parameters and build environments (Table 1). Fig. 1 shows an example microstructure from LPBF-built 17-4 PH [5], which is representative of the hierarchical microstructure observed in many LPBF-built steels. The microstructure is characterised by small grains, generally between 10 μm-100 μm, each containing a forest of solidiication cells. The solidiication cells are elongated in the direction of growth, and typically range from 0.2 μm to 2 μm in diameter, depending on local solidiication rates. The cells are separated by dense walls of geometrically-necessary dislocations, caused by adjacent cells having slightly diferent crystallographic orientations [6].
This deinition of grains and solidiication cells follows the approach outlined in [6], where the solidiication cells are of the micron length scale, while grains (up to 200 microns) contain multiple solidiication cells of the same growth direction. This hierarchical structure, with grains and cells of these characteristic length scales, is observed across a wide range of additively manufactured steels including 316 L [6,7], 17-4 PH [1,8], H13 [9] and M300 [10,11].
A number of the studies listed in Table 1 attribute the elevated level of retained austenite to the efect of a reduction in austenite grain size suppressing the thermally-induced martensite start temperature [12]. However, they do not explore whether the martensite present is thermally-induced or deformation-driven, and do not consider whether the controlling length scale in the suppression of thermal martensite is the grain size or the solidiication cell size. The solidiication cell size has been shown to be the controlling length scale in the mechanical performance of LPBF-built steels, speciically for yield strength [6].
Further, much of the literature on LPBF-built 17-4 PH reports quantitative phase analysis from samples which have undergone some degree of surface preparation. It has been shown that the retained austenite in additively manufactured 17-4 PH is highly metastable, and can transform to deformation martensite either as a result of in-situ thermal strain during the build [5,13] or as a result of post-build deformation (e.g. tensile testing, polishing) [4,14]. This places a level of uncertainty on the quantitative phase analysis summarised in Table 1, and suggests that the as-built austenite content may be higher than reported.
This study now considers whether the solidiication cell size is the controlling length scale for the thermally-induced martensite start temperature, M s , in additively manufactured steels and how this afects the likelihood of forming thermally-induced or deformation-driven martensite in the as-built state.

Experimental method
The 17-4 PH powder used was manufactured by EOS, in the 15 μm-45 μm size range, to the chemistry shown in Table 2  The M300 (18-Ni300) powder was manufactured by LPW, now Carpenter Additive, again to a 15 μm-45 μm size range, to the chemistry shown in Table 3.
All samples were built on a Renishaw SLM125, using argon as the build environment. The build parameters are shown in Table 4, and the builds used a meander scan strategy with 67°rotation between layers. The samples were cylinders, 10 mm high and 8 mm diameter, and 10 mm cubes. After building the samples were removed from the baseplate by electro-discharge machining (EDM).
The 17-4 PH build did include samples with smaller point spacing and hatch spacing (higher energy density) than shown in Table 4, but these are not part of this analysis and results from them are discussed elsewhere [5,13]. Data from those samples is included in Fig. 3b and c, but the low energy density sample under consideration here, corresponding to the build parameters above, is circled.
Optical microscopy and scanning electron microscopy (SEM) were carried out on a mechanically sectioned surface from a cube of 17-4 PH. The surface was ground and polished, down to 1 μm diamond suspension, and etched with either Kallings #2 or 2% Nital. This surface preparation is expected to have caused deformation-driven phase transformation of retained austenite to martensite, but as this is a displacive transformation the physical arrangement and length scales of grains and solidiication cells should be unafected. The SEM images for 17-4 PH were taken using a FEI Nova 450 at the University of Sheield, operating at 20 kV. The SEM images for M300 were taken using a FEI InspectF at the University of Sheield, operating at 10 kV.
Vibrating sample magnetometer (VSM) data for 17-4 PH was generated from a 1.5 mm thick slice across the top of the cylinder, cut by EDM. The VSM was a MicroSense Model 10 at the University of Manchester. All measurements were carried out with the ield perpendicular to the build direction. The measurement program and the method for converting the saturation magnetisation observed in VSM into a martensite phase fraction (in wt%) are covered in the Appendix. While the VSM output is information-rich (including saturation magnetisation, coercivity etc) it is restricted to small, thin samples (8 mm diameter, 2 mm thickness), is comparatively slow (1-2 hours per sample), and cannot easily be used to track martensitic transformation during mechanical or thermal processing.
For 17-4 PH, Feritscope measurements were taken from slices EDMcut from further down the cylinder samples. The Feritscope was a Fischer Feritscope MP30, and at least 4 measurements were taken on each sample to generate a mean and standard error. The Feritscope was calibrated in wt% using a set of standards purchased from Fischer, covering a range from 0.69 wt% to 84.4 wt%, so there was no requirement to convert between Ferrite number and wt%. Feritscope is only a surface measurement technique, but has the beneit of instant quantitative readings that can be taken during a trial and there is no restriction on sample size.
XRD was carried out on the EDM-cut surfaces previously used for Feritscope measurement, with no further surface preparation, using a PANalytical X'Pert3 Powder with Cu Kα radiation (not mono-chromated) at a step size of 0.0394°and a time per step of 1120s. Additional scans were carried out with a time per step of 5000 s.
EDM is judged to be a low deformation cutting technique when compared with mechanical sectioning. It is acknowledged that it causes some thermal processing, and results in the formation of a thin (5 μm-10 μm) recast layer [15]. This was observed in the XRD data as a face-centred cubic (FCC) peak at slightly lower 2θ than the bulk austenite peak, corresponding to a lattice parameter of a =3.64 Å (Appendix Figure A8). This is consistent with the structure and lattice parameter expected from a compositional blend between the bulk 17-4 PH austenite (FCC, a =3.59 Å [12,16,17]) and the Cu65:Zn35 brass wire (FCC, a =3.69 Å [18]) used for cutting.
XRD is a surface measurement technique; 95% of the data is generated from the top 2 μm of material [19]. The XRD traces are therefore   A d d it iv e M a n u f a c t u r in g 3 0 ( 2 0 1 9 ) 1 0 0 9 1 7 far more strongly afected by the presence of the recast layer than the VSM measurements, which are taken from a bulk sample. Rietveld reinement was attempted on the XRD data, but it was not possible to get a repeatable convergence. This was attributed to the overlap between the peaks for the recast layer and the bulk austenite, and the inherent texture of additively manufactured materials. Instead the XRD data has been used for phase identiication, and for comparative analysis between the same sample in diferent preparation conditions. While Feritscope was used for quantitative phase analysis in the cooling trial, this was judged acceptable based on comparisons between Feritscope and VSM data, and the way the data was being used. There is more detail on this in the results section, where data from the cooling trial is presented.
The irst cooling trial used methanol in a table-top laboratory chiller. The 17-4 PH sample was placed in the methanol bath, and progressively cooled to −37°C. Measurements of wt% martensite were taken by Feritscope before and during the cooling process, sampling four diferent measurement locations at each temperature.
The second trial submerged the sample in liquid nitrogen for at least 1 h. For 17-4 PH, twelve measurements of wt% martensite were taken by Feritscope at room temperature before cooling, and again when the sample returned to room temperature after cooling. The sample then measured by XRD after it had returned to room temperature. For M300 the sample was measured by XRD before and after cooling in liquid nitrogen.
For 17-4 PH the cooling trials were carried out on 1 mm thick slices, EDM-cut parallel to the baseplate from a 10 mm cube. For M300 they were carried out on half-cylinders, made by EDM-cutting along the cylinder axis. Both of these were judged to have suiciently small thermal mass to experience high cooling rates and reach equilibrium quickly when submerged in liquid nitrogen from room temperature.

Prediction of M s in 17-4 PH
The efect of austenite grain size on martensite start temperature, M s , is well known, and has been observed in a range of alloys [12,20,21]. A relationship between grain size and M s has been established for a range of steel compositions, validated for grain sizes in the range 1 μm-500 μm (Eq. 1) [12]. In this, f = 0.01 is the irst detectable fraction of martensite, m = 0.05 is the martensite plate aspect ratio and V is the volume of the austenite grain in μm 3 .
This is based on geometric partitioning analysis describing the relationship between the number of martensite plates per unit volume, the fraction of martensite, the aspect ratio of the martensite plates and the size of austenite compartments after being split by a martensite plate. This was combined with experimental data for the number of martensite plates per unit volume as a function of temperature, and the itting constants were determined from data on M s as a function of grain size for a range of alloys. The overall relationship is therefore theoretically determined, and only the itting constants are experimentally determined. The alloys used in the experimental data cover a wide range of both low-alloy and more richly alloyed steels. Their composition range covers that of 17-4 PH, except for chromium and copper.
Eq. 1 uses a theoretical large-grain martensite start temperature, M s 0 , which is deined as the point when martensite becomes thermodynamically possible; when the free energy change G from austenite to ferrite of the same composition is suicient to provide the driving force for nucleation. This critical value, G C , is itself composition dependent.
An expression has been published for calculating G from the atomic fraction X i of each element in the alloy (Eq. 2), but this was established at an experimentally determined M s , so will itself have been inluenced by the grain size of the samples [22,23].
Instead, the M s reported for a sample of 17-4 PH with a known grain size and chemistry was used to calculate M s 0 , with adjustments to account for the minor compositional diferences between that sample and the alloy used in this study [24]. This is covered in detail in the Appendix.
This composition-speciic M s 0 was used to predict the relationship between austenite grain size, L , and martensite start temperature for the exact composition of 17-4 PH used (Fig. 2 a). This relationship was    3 calculated for the range 1 μm-500 μm, corresponding to the range over which Eq. 1 was shown to be valid [12]. The prediction error is generated from a combination of error in grain size measurement from [24]. This shows a predicted M s for large grained material (1000 μm) much lower than that which might be expected, up to only 80°C compared with the quoted values for 17-4 PH of 105°C-132°C [24][25][26]. This shift is primarily due to the composition of the powder being considered. The alloy used here ( Table 2) had more silicon and nitrogen than was reported for the comparison alloy [24]. This resulted in an increased value of G as well as a shift in the free energy curves for austenite and ferrite.
The predicted M s was combined with the Koistinen-Marburger relationship (Eq. 3) to predict the volume fraction of retained austenite at room temperature, T RT , as a function of austenite grain size (Fig. 2b) [26]. The itting constant = 0.02955 is taken from a it to neutron difraction data from 17-4 PH on cooling [24].
While this calculates the volume fraction, given that the volume expansion on transformation from austenite to martensite is only ∼3%, this has been taken to be equivalent to weight fraction for this analysis.
From the prediction above, grain sizes in the range 10 μm-50 μm, at the lower end of the characteristic range, would be expected to suppress M s to 30°C-50°C, giving 40 vol%-70 vol% retained austenite at room temperature. This is a conservative approach; if the M s were higher than this prediction, closer to the values reported in literature, then the expected phase fraction martensite for a given grain size would be higher.
Comparing this with the data in Table 1, the values appear roughly consistent with suppression by the grain size, at a controlling length scale of 10 μm-50 μm. The main discrepancy is the fully austenitic structure reported in [5], which would suggest a controlling length scale closer to that of the solidiication cells.
All of the studies in Table 1, with the exception of [5], report quantitative phase analysis (QPA) measured only from XRD traces and the majority state that the sample had been mechanically prepared before measurement (cut, ground, polished). In contrast, the analysis in [5] used XRD for phase identiication, but then based the quantitative phase analysis on VSM, a bulk technique which would be less afected by any surface transformation. These results are summarised in Fig. 3, with the sample under consideration in this study circled in Fig. 3b-c. It has been shown that metastable retained austenite in LPBF-built 17-4 PH can transform to deformation martensite [4,5,14]. It is suggested that, for the other samples listed in Table 1, this could have occurred during the sample preparation, and the fraction austenite in the as-built state could have been higher than reported.
This would explain the wide range in reported phase fraction austenite, from samples with similar build conditions, and explain why a fully austenitic structure was reported in [5] when all other studies have reported some martensite. Therefore, the apparent correlation between the reported austenite content in Table 1 and the prediction in Fig. 2 is not suicient to conirm that the grain size is the controlling length scale.

Results
Reviewing Fig. 2, if the grain size is the controlling length scale, then it would be expected that cooling to -30°C would result in 80 vol %-90 vol% thermal martensite. Cooling in liquid nitrogen (−196°C) would give > 99 vol% thermal martensite. To test this, samples of the LPBF-built 17-4 PH material were cooled to −37°C in methanol and to −196°C in liquid nitrogen.
In both cases measurements were taken by Feritscope on a surface cut by EDM. While Feritscope is a surface method, and therefore will be afected by the recast layer, the measurements are for comparative analysis, looking for a change in the martensite content as a result of the cooling process.
Data presented above (Fig. 3c) shows that Feritscope (surface) and VSM (bulk) measurements from an EDM-cut surface are in good agreement for low martensite contents, although they diverge at higher martensite contents, with Feritscope generally measuring lower than VSM, presumably due to the recast layer. However, if grain size is the controlling length scale, then sub-zero cooling would be expected to result in > 50 wt% martensite. The data in Fig. 3c gives conidence that this would register in the Feritscope reading, even if VSM were then required to measure a more accurate phase fraction.
In summary, the use of Feritscope in this trial is judged acceptable due to the comparative nature of the analysis, combined with the expected magnitude of the martensite phase fraction if grain size were the controlling length scale.
Before cooling both samples showed an initial level of (0.38 ± 0.01) wt% martensite; the martensite present was attributed Fig. 3. a) VSM data for 17-4 PH sample showing majority paramagnetic (austenite) contribution with only small ferromagnetic (martensite) contribution [5]; b) Martensite wt % in 17-4 PH determined from VSM saturation magnetisation against build energy density; c) Comparison of VSM and Feritscope data from 17-4 PH showing good agreement when martensite phase fraction < 10 wt%, error bars show standard error across repeat Feritscope measurements [13]; d) XRD data from 17-4 PH sample showing majority austenite and minority martensite, generated at 1120s per step.

F.S.H.B. Freeman, et al.
A d d it iv e M a n u f a c t u r in g 3 0 ( 2 0 1 9 ) 1 0 0 9 1 7 to deformation-driven transformation triggered by EDM when removing the sample from the baseplate and cutting the slice. Sample 1 showed no further transformation on cooling to 0°C. There was a very slight increase in martensite content on cooling from −30°C to −37°C, up to (0.66 ± 0.07) wt%, although this was accompanied by an increase in the standard error and may be within the measurement error of the Feritscope. Sample 2 showed similar behaviour with a very slight increase in martensite content, up to (1.37 ± 0.18) wt% after 3 h at −196°C. While both samples showed some martensitic transformation, the extent was much smaller than the > 80 vol% martensite expected from grain size limited transformation (Fig. 4a). The XRD trace (Fig. 4b insert) conirmed the Feritscope result, showing small peaks for martensite and strong peaks for austenite. As explained earlier, it was not possible to achieve a good quality Rietveld reinement on the XRD data for quantitative phase analysis.
To conirm this result, a similar trial was carried out on M300, also known as 18-Ni300. This has a higher martensite start temperature (280°C [27]), so would be expected to have a higher proportion of martensite at room temperature for the same grain size. The literature on LPBF processed M300 shows that it has the same characteristic hierarchical microstructure as illustrated above, and on the same length scale [10,28]. In the as-built condition, the phase fraction of martensite has been reported to vary between 88.6 % [10] and 94.2 % [28], although this was again from samples which had undergone some surface preparation before XRD.
Initially, the as-built surface was measured by Feritscope in three separate locations, with three repeats in each location. This gave a reading of (50.2 ± 1.5) wt% martensite, which is considerably lower than that reported in literature [10,28]. The measurement variation is comparatively high, due to measuring on the rough, as-built surface but this avoided any inluence from the recast layer on an EDM-cut face. This cannot distinguish between thermally-induced martensite which had transformed on cooling, or deformation-martensite resulting from in-situ thermal strain during the build.
If the martensite was thermally-induced, then further cooling would be expected to drive further transformation, in accordance with the Koistinen-Marburger relationship (Eq. 3). The sample was cooled in liquid nitrogen for 2 h, returned to room temperature and the measurement was repeated, giving a reading of (51.7 ± 1.8) wt% martensite (Fig. 5). This is within the error from the previous measurement, indicating that there had been no further transformation.
To test the susceptibility to deformation-driven transformation, the same sample was then subjected to very gentle polishing with P2500 grit paper, which would be expected to cause some local surface deformation. After polishing, the sample was re-measured and the Feritscope reading had now increased to (71.1 ± 0.2) wt% martensite. This was a signiicant increase, indicating considerable deformationdriven transformation had occurred. The reduction in error is due to the reduced roughness in the measurement surface.
Optical microscopy of the M300 sample showed clear keyhole melt pools, approximately 100 μm-150 μm in depth and 100 μm in width (Fig. 6a). At higher magniication it could be seen that these contained the expected hierarchical microstructure, with grains ranging from 10 μm to 100 μm in length, containing extremely ine solidiication cells (Fig. 6b-c & Appendix Figure A10). These were generally between 0.3 μm and 1.3 μm across ( Figure A11). This is consistent with the microstructure reported elsewhere for LPBF-build M300 [10,11].

Discussion
In this work, it is shown that 17-4 PH and M300 produced by LBPF have comparable microstructures, with micron-scale solidiication cells, and both show complete suppression of thermally-induced martensitic transformation, even on cooling in liquid nitrogen (−196°C). Both also show a deformation-driven transformation in response to either in-situ or post-build deformation.
If the grain size were the controlling length scale for the thermallyinduced transformation mechanism, then both materials would be expected to show transformation when cooled in liquid nitrogen. While the prediction of M s for 17-4 PH shown in Fig. 2 cannot necessarily be extrapolated to length scales below 1 μm, it does suggest that the controlling length scale must be of this order of magnitude to get the observed level of thermal martensite suppression. For M300 the   A d d it iv e M a n u f a c t u r in g 3 0 ( 2 0 1 9 ) 1 0 0 9 1 7 martensite start temperature should be higher, making it even less likely that full suppression could be achieved with grain size as the controlling length scale. Complete suppression of thermally-induced martensitic transformation has been reported elsewhere for material with an austenite grain size of 0.8 μm [20], which is comparable with the solidiication cell size. It was shown that, in these ultra-ine grains, the strain energy associated with multi-variant nucleation was extremely high and therefore very unlikely, but that single variant transformation required an unachievable level of undercooling. The result was that thermally driven martensitic transformation was fully suppressed, even after cooling in liquid nitrogen.
The importance of solidiication cell size on mechanical behaviour of LPBF-built materials has been demonstrated elsewhere [6]. This showed that, while the misorientation between the cells was comparatively low at only 1-2° [6,29], under tensile testing, the dense dislocation walls around the solidiication cells were suiciently robust to strengthen the material in a manner normally attributed to grain boundaries. Further, it was shown in [6] that there was a strong dislocation trapping and retention mechanism inside the cell walls.
In this study, it is suggested the solidiication cells are the controlling length scale for thermally-induced martensite; the dense dislocation walls being suiciently robust to prevent the displacive martensitic transformation crossing them. If the dense dislocation walls were not able to prevent martensite growth, then the controlling length scale would be the grain size, and there would be thermally-induced transformation on cooling to −196°C. As this did not happen, the conclusion is that the controlling length scale is the solidiication cell size, and that the dense dislocation walls must be suiciently robust to stop thermal martensite.
The microscopy of both 17-4 PH (Fig. 1) and M300 (Fig. 6) samples show that the solidiication cell size is not completely uniform across the material; it will vary depending on the local solidiication conditions. Eq. 4 is an empirical relationship between the cooling rate T and primary dendrite arm spacing 1 (solidiication cell diameter), where a and n are material dependent constants, determined to be a = 60-100 ms/K and n = 0.2-0.5 for steels [30][31][32]. The very low level of martensitic transformation seen in the 17-4 PH sample after cooling below 0°C (Fig. 4) may be due to limited thermallyinduced martensitic transformation taking place in the larger solidiication cells, or where the cell walls are slightly less heavily dislocated.

Conclusions
A hierarchical microstructure is characteristic of LPBF-built steels, with grains of 10 μm-100 μm containing solidiication cells of 0.2 μm-2 μm, which are surrounded by dense dislocation walls. These walls have been shown to have a mechanical strengthening efect comparable with grain boundaries.
It has been previously shown that it is possible to achieve a fully austenitic structure in LPBF-built 17-4 PH. Here, it has been demonstrated that this is stable on cooling in liquid nitrogen, with no thermally-induced martensitic transformation. Similar results were observed in M300, with 50 wt% martensite in the as-built condition, but no further transformation on cooling in liquid nitrogen. This indicates that, in both materials, the thermally-induced transformation mechanism has been suppressed.
Suppression of thermally-induced martensite to this extent is not consistent with the grain size being the controlling length scale. It is therefore concluded that the solidiication cell size is the controlling length scale for thermally-induced martensite in LPBF of martensitic steels, and can actually suppress thermally-induced martensite completely. This results in an elevated level of retained austenite, which will readily transform to deformation martensite as a result of either insitu thermal strain or post-build processing (e.g. grit-blasting, machining, sample preparation).
To get accurate phase quantiication, it is essential to minimise, or at least understand the efect of, any post-build deformation or surface treatments. It may be more appropriate to use bulk measurement techniques (e.g. VSM), measure in the as-built condition without any surface preparation or use low-deformation techniques such as electro-polishing.
Much of the literature on LPBF-built martensitic steels reports phase fractions from ground and polished samples, where sample preparation may have caused the surface to transform, leading to a higher martensite content being reported. This should be taken into account when comparing the reported microstructure and mechanical properties. On a more positive note, this also suggests a method for achieving a spatially varied microstructure using LPBF to build a majority austenitic component, and then mechanically processing the surface to achieve a harder martensitic outer casing.

Funding
This work was supported by the Engineering and Physical Sciences Research Council in part by award reference 1686001 and in part by the Future Manufacturing Hub in Manufacture using Advanced Powder Processes (MAPP) (EP/P006566/1).

Contributions
FF carried out the builds in 17-4 PH and M300, and the XRD and A d d it iv e M a n u f a c t u r in g 3 0 ( 2 0 1 9 ) 1 0 0 9 1 7 6 Feritscope analysis. JS and FF carried out the methanol and liquid nitrogen trials. JX and FF carried out the microscopy on M300. FF, JX, JS and IT analysed the data and developed the concept. FF wrote the manuscript. All authors reviewed the manuscript.

Declaration of Competing Interest
None.

Appendix A. Supplementary data
Supplementary material related to this article can be found, in the online version, at doi:https://doi.org/10.1016/j.addma.2019.100917.