Minerals Engineering 58 (2014) 8089

Contents lists available at ScienceDirect

Minerals Engineering
journal homepage: www.elsevier.com/locate/mineng

Particle shape effects in otation. Part 1: Microscale experimental observations q

David I. Verrelli a,,1, Warren J. Bruckard a, Peter T.L. Koh b, M. Philip Schwarz b,2, Bart Follink c
a

CSIRO Process Science and Engineering, Bayview Avenue, Clayton, VIC 3168, Australia CSIRO Computational Informatics, Bayview Avenue, Clayton, VIC 3168, Australia c Ian Wark Research Institute, University of South Australia, Mawson Lakes, SA 5095, Australia
b

a r t i c l e

i n f o

a b s t r a c t
There has long been speculation as to whether some particle shapes are more oatable than others, which might be due to differences in the induction period required to achieve attachment between the particles and the air bubbles in the pulp. To resolve this, we used the Milli-Timer apparatus to directly observe the process of particlebubble interaction and attachment by means of a magnied, high-speed video recording, thus providing a direct measure of the induction period for attachment. To assess the inuence of particle shape on induction time we used two varieties of methylated borosilicate glass particles spheres and angular frit in a range of tightly-sized fractions. Other factors that could affect the induction time, such as the polar angle of sliding commencement, and approach velocity, are accounted for using multiple nonlinear regression. Our results illustrate the importance of particle shape on induction period, with angular particles exhibiting induction periods that were an order of magnitude lower than those of spheres. Furthermore, the induction period was seen to decrease with increasing particle velocity, or kinetic energy on approach, but increased as the trajectory approached the limit of just grazing the bubble. These results indicate that attention should be paid to the shape of particles obtained from the grinding operation, besides particle size. Crown Copyright 2014 Published by Elsevier Ltd. All rights reserved.

Article history: Received 19 August 2013 Accepted 5 January 2014 Available online 13 February 2014 Keywords: Flotation bubbles Flotation kinetics Froth otation Grinding Induction time Mineral processing

1. Introduction Flotation is a key unit operation employed in mineral processing and a range of other industries. Successful otation hinges on the attachment of (certain) particles to bubbles. In real industrial systems the particles fed into a otation cell will exhibit a variety of shapes, varying from approximately spherical (e.g. zircon, slags) (see Wotruba et al., 1991), to cuboidal (e.g. galena) (Dippenaar, 1982), to platy (e.g. talc, chlorite) or acicular (e.g. tremolite) (Kursun and Ulusoy, 2006). How important is this shape in determining attachment? Anecdotally it is expected that particle shape can have a significant effect on oatability: the common view is that angular par-

q Extension of paper presented at IMPC conference, New Delhi, India, 2428 September 2012. Corresponding author. Address: ASAM, 2 Technology Place, Macquarie University, NSW 2109, Australia. Tel.: +61 2 9850 2755; fax: +61 2 9850 2701. E-mail addresses: David.Verrelli@mq.edu.au (D.I. Verrelli), Warren.Bruckard@ csiro.au (W.J. Bruckard), Peter.Koh@csiro.au (P.T.L. Koh), Phil.Schwarz@csiro.au (M.P. Schwarz), Bart.Follink@unisa.edu.au (B. Follink). 1 Currently at Macquarie University. 2 Currently at CSIRO Process Science and Engineering.

ticles are easier to oat than rounder particles. However, such overall tendencies leave it unclear as to which otation subprocesses are most affected by particle shape. For example, in the pulp we could consider collision rate, attachment efciency, stability against detachment, and entrainment, besides a number of froth characteristics. Traditionally the efforts to model the interaction of particles and bubbles have typically assumed that both objects are perfect spheres (see Nguyen and Schulze, 2004). Some exceptions are studies on the nal moments prior to breaching of the liquid gap, in which bubble deformation is modelled (Chan et al., 2011), and studies that examine the equilibrium conguration of a non-spherical particle after attachment has occurred in the pulp (Huh and Mason, 1974; Schulze, 1984: 199ff) (cf. Coghill and Anderson, 1923: 44ff) or the froth (Dippenaar, 1982; Morris et al., 2011). Hence, modelling is not yet at a stage to fully explain the relative importance of particle shape or roughness. 1.1. Measurement of particle shape and roughness In otation the particles of interest are typically sized at around 10150 lm (Shergold, 1984: 231). If the deviations occur on a scale comparable to the particle size, then we shall refer to this as a

http://dx.doi.org/10.1016/j.mineng.2014.01.004 0892-6875/Crown Copyright 2014 Published by Elsevier Ltd. All rights reserved.

D.I. Verrelli et al. / Minerals Engineering 58 (2014) 8089

81

difference in shape. If the deviations occur on a scale much smaller than the particle size, then we shall refer to this as a difference in surface roughness. An equivalent distinction is between structure and texture (Ahmed, 2010). An important element of working with real particles is to characterise their shape and surface roughness. Unfortunately, there is no single unambiguous parameter to fully characterise either of these properties, which has hindered research into this important area (Holt, 1981; Yekeler et al., 2004; Kursun and Ulusoy, 2006; Ahmed, 2010). The topic of roughness is scarcely more tractable: the topography of a solid surface can exhibit different characteristics on different lengthscales, such as rough or smooth asperities. Attempts to deal with these problems through use of fractal dimensions (e.g. Filippov et al., 1999; Ahmed, 2010) are subject to question, in general, as the features of a given particle are unlikely to conform to a true fractal rule over a broad enough range of lengthscales to be physically useful unlike the features of aggregates (Verrelli, 2008). 1.2. Inuences on particle shape, and roughness Different views exist on the dominant reason why particles have the shapes that they do, which can be summarised as the effect of the inherent material character, and the effect of the processing (Holt, 1981). For example, malleable materials such as some metals can deform plastically (e.g. Ofori-Sarpong and Amankwah, 2011), while brittle materials such as quartz will fracture (e.g. Holt, 1981). Laminated materials preferentially fail along certain orientations, while amorphous materials do not; on the other hand, the easy slip plane in the layered material may yield a smoother failure surface. The tendency to fail along grain boundaries or through grains also varies between materials, depending inter alia upon the structural homogeneity (Gaudin, 1926). In the industrial context, working with a given material, the type of grinding takes on increased importance. A number of studies have been reported in the literature in which the effects of different grinding processes on a given sample are reported. A typical result is that more angular particles are produced from rod mills (attributed to impact processes), compared to rounder particles obtained from ball mills (attributed to abrasion and chipping) (Yekeler et al., 2004), although this depends on the details of operation (Gaudin, 1926). Yet at the same time the particles ground in a rod mill have been reported to be smoother than particles from a ball mill (Yekeler et al., 2004).3 Autogenous milling produced intermediate results in each parameter in the work of Yekeler et al. (2004), while Forssberg and Zhai (1985) reported that autogenously ground ore particles were less elongated than those from either ball or rod milling. Hiylmaz et al. (2006) characterised pyrite particles of a given size fraction obtained from dry autogenous grinding as being rougher (by BET surface area) and more acute (by aerodynamic resistance) than those obtained from dry ball milling. This differed from their earlier results for barite (Hiylmaz et al., 2005), in which dry ball milling produced rougher and more acute particles than dry autogenous milling, for a given size fraction. This emphasises the effect of the feed material on the outcome (Hiylmaz et al., 2006). For the barite samples the smaller size fractions were reported to be rougher and more acute in shape (Hiylmaz et al., 2005), while the opposite trend was reported for the pyrite samples (Hiylmaz et al., 2006).

Dry and wet milling were carefully studied by Feng and Aldrich (2000), who found that the dry-ground particles had relatively rough surfaces containing microstructural defects, whereas the wet-ground particles had smoother, cleaner surfaces. It should be clear that comparison between studies on different materials is fraught with difculty; yet even when the material is identical, the situation may not be straightforward. Besides the type of material and type of machine, the resulting particle shapes depend inter alia on: size of feed material; speed of operation of equipment; geometry of equipment; product particle size, and particle size distribution (Gaudin, 1926; Holt, 1981). 1.3. Effect on otation 1.3.1. Shape One of the key resistances to attachment is the hydrodynamic resistance arising as liquid drains out of the gap between a bubble and an approaching particle. Aspherical particles could experience a lower resistance (depending upon their shape and orientation), and hence require less time for the intervening gap to thin sufciently to be breached, and attachment occur. There have been few systematic studies on the effect of particle shape or roughness upon otation performance. One of the earliest notable studies on this topic was carried out by Anfruns and Kitchener (1977). They concluded that angular particle shapes have much higher attachment efciencies than spheres, which they attributed to easier rupture of the wetting lm. In that work the angular particles were composed of quartz (Brazilian rock crystal), while the ballotini spheres were of lead glass. Despite care taken to ensure the same size (31 lm equivalent Stokes diameter), similar density, and similar surface properties, an effect of the difference in materials cannot be entirely ruled out (cf. Verrelli et al., 2012). In distilled water, Anfruns & Kitchener found collection efciencies about four times greater for angular quartz than for spherical glass particles. According to their experimental description, the collision rates would be the same in each case, entrained particles were eliminated, and detachment was implied to be negligible; hence, the differences should be due to variation in the ease of attachment. Studies on talc have suggested that rounder (but rougher) particles produced by ball milling are less easily recovered in a microotation cell (Yekeler et al., 2004) and in column otation (Kursun and Ulusoy, 2006), compared to more elongated (but smoother) particles produced by rod milling. These experiments used a relatively broad class of particle sizes (40250 lm). Despite the similar particle size distributions from each mill (Yekeler et al., 2004), narrower fractions would avoid the potential issue associated with dependence of shape upon particle size. Even within a given narrow size fraction, the predominant shape may depend upon whether it comes from the lower end of a coarse grind, or the upper end of a ne grind. Wotruba et al. (1991) reported that prismatic zircon particles of a given size oated better than those with more spherical or ellipsoidal shapes in a 1.5 L cell. (Conrmed independently by Gl (2006).) The particles all had essentially the same composition by XRF. The prismatic particles tended to have smoother faces and sharper edges than the other types of particle in the otation feed. However, even after roughening the faces and rounding down the sharp edges, the prismatic particles remained more oatable. Wotruba et al. proceeded to characterise the ease of detachment of the original particles from a bubble, and found that the energy necessary for detachment was signicantly greater for prismatic particles attached parallel to their long axes (at-attached) than for either rounded particles or end-attached prismatic particles. This can be related to the length of the three-phase contact line (TPCL) formed. Wotruba et al. accepted that rounded particles may also be

3 The rod-milled particles were also claimed to be more hydrophobic than the ballmilled product (Yekeler et al., 2004). However, that conclusion rests on comparing extrapolations of surface tension that are subject to uncertainties larger than their differences, besides being founded on Zismans theory, which is itself subject to criticism (Siboni et al., 2004).

82

D.I. Verrelli et al. / Minerals Engineering 58 (2014) 8089

disadvantaged in forming an attachment to a bubble, although they did not isolate this step in their own experiments. Schmidt & Berg have described an opposite tendency: they found that spherical printer toner particles oated better than disc or platy particles (Ahmed, 2010). Ahmeds work showed increases in otation recovery and greater resistance to detachment with increases in either particle asphericity or particle roughness (Ahmed, 2010). As the two parameters were highly correlated, it is difcult or impossible to tell from his data which was the more important factor. Furthermore, as with Anfruns & Kitchener, Ahmed used particles of somewhat different composition, although it was nominally accounted for by methylation of each sample to an equal degree of hydrophobicity. Finally, Hiylmaz et al. probed the effect of particle shape on the attachment process alone by testing recoveries in a modied Hallimond tube across a series of size fractions. The barite (Hiylmaz et al., 2005) and pyrite (Hiylmaz et al., 2006) particles that were characterised as rougher and more acute in shape showed lower recoveries. However, the results may have been inuenced by differing degrees of surface oxidation, and also skewed in favour of rounder particles by the use of low collector dosages (Hiylmaz et al., 2006). 1.3.2. Roughness Particles with rough surfaces composed of microscale projections might achieve attachment more easily, if those bumps or jags protrude through the gap, so that the separation between particle and bubble is smaller than it might otherwise seem. A three-phase contact would then be expected to occur on one of the projections. Nanoscale topological features seem less likely to have much effect on the ease of breaching the interjacent liquid; however, they may still affect the expansion of the three-phase contact line in dewetting (Oliver and Mason, 1977). Moreover, chemical composition and surface properties can vary on the atomic scale in layered crystalline materials (Yekeler et al., 2004; Franks and Gan, 2007), which would be accentuated on rough surfaces. A further feature of rough surfaces is that they may more readily harbour small air bubbles (microbubbles or nanobubbles), which promote attachment by some mechanism possibly including hydrophobic forces (Krasowska et al., 2007). Feng and Aldrich (2000) found a large number of differences in the particles prepared by wet and dry milling, with the dry-ground samples exhibiting faster dissolution, faster reagent adsorption, more stable and higher-loaded froths, and faster otation kinetics. These researchers were very careful to attain comparable degrees of comminution in the two milling operations. Nevertheless, it is notable that no consideration of the effect of particle shape was made. The otation kinetics were evaluated from recovery times in a Leeds otation cell under typical operating conditions; hence, effects such as entrainment and detachment will have inuenced the results. As alluded to above, Wotruba et al. (1991) roughened the surfaces of a mixture of near-spherical, ellipsoidal, and prismatic zircon particles through attrition in a planetary mill. The particles size and basic shapes were nearly unaffected by this processing. Their results indicated a much better oatability of the roughened particles in comparison to the original otation feed, for a given collector dose. 1.3.3. Complicating factors Different grinding techniques do not just affect the physical form of the minerals, but can also affect the chemistry of the surface, for example through more or less reducing conditions (Forssberg et al., 1988; Bruckard et al., 2011), and activation of the mineral surface through creation of material defects or

disorder (Feng and Aldrich, 2000; Yekeler et al., 2004). Under some circumstances microscopic air pockets may persist in crevices at the surface of the particles (Harvey et al., 1944), making subsequent attachment to an air bubble easier (Coghill and Anderson, 1923: 48); this would be more relevant in the case of dry grinding. 1.4. Induction period The attachment between a bubble and a particle (see Fig. 1) is commonly described as requiring a minimum, critical time to occur, once the two bodies are brought into proximity; this is the induction period, s (Sven-Nilsson, 1934). In otation, the conventional theory compares s against the time available for a particle to slide over the bubbles surface (Nguyen and Schulze, 2004: 265ff). Based on hydrodynamics, the sliding time is known to depend upon factors such as particle size, and the approach trajectory. The induction period is expected to depend on the surface chemistry, and perhaps other factors such as particle shape. For a given particle and bubble pair, the induction period is tacitly presumed to be constant; however, the latest research indicates that s actually depends upon the particles approach trajectory (Verrelli et al., 2012). We are interested in using our technique of direct observation of the induction period in order to assess the role of the attachment process in the previously reported changes in oatability. This will establish whether the altered oatability is due to changes in the ease of attachment, following from changes in the time required for thinning of the interposed liquid, or whether it is due to other factors, such as detachment or entrainment. In order to assess the importance of particle shape and roughness on the induction period, s, it is necessary to account for as many other factors as possible. Hence, we conduct our experiments on narrowly sized borosilicate glass particles that differ only in their shape and surface roughness. We are unable to precisely control the particle approach trajectories or speeds, so instead we include corrections for these parameters in the analysis. 2. Methodology 2.1. Milli-Timer CSIROs Milli-Timer (Fig. 2) has been described in detail previously (Verrelli et al., 2011). Briey, it consists of a stationary bubble

Fig. 1. Schematic of the progress of a particle interacting with and attaching to a bubble (see also Albijanic et al., 2010).

D.I. Verrelli et al. / Minerals Engineering 58 (2014) 8089

83

Fig. 2. CSIROs Milli-Timer. Adapted from Verrelli et al. (2011). The bulb of the pipette is not squeezed.

that is held underwater at the end of a capillary, and onto which particles are dropped. The interactions are recorded on a highspeed video camera. In the present set-up the camera used was a Phantom v210 (Vision Research, Wayne, U.S.A), tted with a Zoom 6000 lens assembly (Navitar, Rochester, U.S.A) comprising a 2 magnifying F-mount extension tube, a zoom lens with focussing capability, and a 0.75 lens attachment, for a nominal magnication of 6.75 through the lens assembly. The actual magnication of the cameralens system is approximately 7.1. The camera was operated at a capture rate of 1000 frames per second, and exposure time of 10 ls. A KL 1500 LCD 150 W halogen cold light source (Schott, Mainz, Germany) provides intense white light. To minimise image blur due to differential refraction of different wavelengths of light, a BP635 colour lter (Midwest Optical Systems, Palatine, U.S.A.) was interposed, which passes wavelengths in the range 590670 nm. Induction periods are dened as the time that elapses from the moment at which sliding commences, until the moment that attachment is initiated (see Fig. 1). The latter moment is clearly evidenced by a jump in event, as described previously (Verrelli and Koh, 2010; Verrelli et al., 2011). The former requires some judgement. In the ideal case, the particle will approach the bubble and then undergo sliding within the plane perpendicular to the cameras viewing axis and passing through the bubbles centre (i.e. approximately the middle of the focal plane). In reality the particles are commonly observed to travel along slightly deviating azimuthal angles (Verrelli et al., 2012), so that it is more practical to determine sliding commencement by identifying a substantial reduction in the radial velocity of the particle with respect to the bubbles centre. In each run, a dilute swarm of particles settles through the water column toward a freshly blown bubble. Depending upon the attachment efciency, any given run might result in multiple opportunities to estimate induction time, a single opportunity, or no opportunities. Only interactions that could be clearly seen, that resulted in attachment, and that were not subject to any apparent interferences, were included in the data analysis. The polar angle of particle position at the moment that sliding begins is given the symbol4 /, and is obtained from the video recordings as above (schematically depicted in Fig. 1). The initial

approach velocity of a given particle toward the bubble, v0, is obtained by measuring displacement of the particle from frame to frame as it rst enters the eld of view, when the effect of the bubbles presence is minimal. 2.2. Materials 2.2.1. Bubbles Bubbles of 2.0 lL volume were blown at the end of a blunt 20 gauge capillary shortly before each run using ambient air, giving diameters of 1.56 0.04 mm. 2.2.2. Particles The particles used were composed of borosilicate glass, in two forms: spheres, and angular frit (Mo-Sci Specialty Products, Rolla, U.S.A). The spheres were supplied in a 7590 lm size fraction, designated as Precision Grade, which are specied as being at least 90% in the nominated size range and at least 90% classed as spherical. The scanning electron microscope (SEM) images we have taken in Fig. 3(A) conrm a narrow size distribution, and high sphericity, with rather smooth surfaces. The frit is manufactured by comminution of large chunks of recycled borosilicate glass, rst in a hammer mill, and then in a disk mill (e.g. Bico plate pulveriser). SEM images of the frit in Fig. 3(B) show the variety of shapes encountered. Using new screens, we sieved the as-received frit into narrow size fractions, of which the 7590 lm fraction was used in the present work. Due to the origin of the frit, its composition is practically the same as that of the borosilicate spheres (see Verrelli et al., 2012). This is conrmed in assays reported in the online Supplementary Material. All particles were washed using progressive stages of water rinsing, alkaline washing and acid washing (Verrelli et al., 2012). The clean particles were then methylated using chloro(trimethyl)silane (CTMS) to achieve a surface coverage equal to 50% of the maximum achievable coverage (Verrelli et al., 2012), designated as 50% methylated. This process increases particle hydrophobicity by adsorption of a stable coating on the particle. Full details of the methylation procedure are provided in the Supplementary Material. 2.2.3. Characterisation of particles The specic surface area of the frit was characterised using the approach suggested by Brunauer et al. (1938). Measurements were attempted rst using a TriStar 3000 surface area analyser

4 The symbol /1 was used in our previous publication (Verrelli et al., 2012). The subscript is omitted here for clarity.

84

D.I. Verrelli et al. / Minerals Engineering 58 (2014) 8089

Fig. 3. Micrographs of the particles at various levels of magnication. (A) Spherical particles. (B) Angular frit. Magnications are (i) 100 (ii) 1000 and (iii) 15000.

(Micromeritics, Atlanta, U.S.A) operating with nitrogen gas, with sample masses between 0.6 and 0.7 g. However, this proved inadequate for the small surface areas involved. Hence, nal characterisation was carried out on a TriStar II 3020 (Micromeritics), operating with argon gas for improved sensitivity, with sample masses between 1 and 2 g. Condence intervals on each measurement were evaluated based on uncertainty in the regression (including covariance (Montgomery et al., 2001: 32,74,82f.,103f)); the intervals do not include errors from other sources. Micrographs were taken with a eld emission, environmental SEM (FEI, Oregon, U.S.A) operating at 10 kV and a probe current of approximately 300 pA. Additionally, energy-dispersive spectroscopic (EDS) measurements were made on selected regions of the particles with this instrument, as a check of their composition. 2.3. Multiple nonlinear regression In order to systematically investigate the inuence of three key mechanisms on induction period, rigorous statistical analysis was performed in the form of multiple nonlinear regression. The three key issues explored are the effect of approach velocity, approach position, and particle shape or roughness. These are described by the approach velocity, v0, the polar angle of sliding commencement, /, and a binary indicator variable, f, that takes a value of either 1 for frit or 0 for spheres. We performed the regression using the following general equation:

in which the bi are constant coefcients. We evaluated Eq. (1) for the set of all possible regressions (Montgomery et al., 2001: 302ff). The best regression for a given number of parameters was that which minimised the mean squared error (MSE). Further details are provided in the Appendix. 3. Results and discussion 3.1. Particle shape, and roughness The nominally spherical particles are seen in Fig. 3(A) to mostly approximate ideal spheres. They do, however, exhibit some roughness due to debris melded on their surfaces. This debris is of irregular shape, but of size approximately 0.15 lm, and is composed of the parent material, as conrmed by EDS measurements. A small number of angular particles are seen intermingled in Fig. 3(A)(i); those angular particles are readily discerned on the video recording, and are excluded from further analysis. The frit used in the present experiments was produced by comminution in a hammer mill and disk mill in sequence. It would be expected that the nal particle shape and roughness would be inuenced more by the later stage. Particle size reduction in disk mills is due principally to shear, with smaller contributions from particle compression, impact and other mechanisms, which is not so different from the hammer mill (Austin and Rogers, 1985). The few previous reports on disk mills in the literature (Husemann et al., 1976; Ofori-Sarpong and Amankwah, 2011) are not directly relevant, due to the different materials used. Arguably the most straightforward means of shape classication is still semi-qualitative descriptions obtained from visual inspection (cf. Ahmed, 2010). Micrographs of the frit are presented

s b1

10 0

b2

2 0

b3

v0

2 b4 v 0 b5 v 2 0 b6 / b7 / b8 cos/ b9

b10 f 1 f ;

D.I. Verrelli et al. / Minerals Engineering 58 (2014) 8089

85

0.160
Spheres : Predicted

0.140

Spheres (Argon) Frit (Nitrogen) Frit (Argon)

Measured surface area [m/g]

0.120 0.100 0.080 0.060 0.040 0.020 0.000 45 to 53 m 53 to 63 m 63 to 75 m 75 to 90 m

Frit : Predicted

90 to 106m

Size fraction
Fig. 4. Specic BET surface area of spherical and angular borosilicate glass. The gas in parentheses indicates the species used for characterisation. The prediction for spheres uses the average of surface areas computed for the two end-points of the respective size class; the prediction for frit is double the value for spheres. The error bars enclose 95% condence intervals based on the linear regression used in the BET technique. The source and preparation of all particles were the same.

in Fig. 3(B). It is apparent that the most common shapes are platy fragments, although other shapes are also observed, such as tetrahedra and cuboids. As with the spherical particles, the frit exhibits roughness due to debris of size approximately 0.15 lm stuck on the surface. In addition to this, step-like features, cavities, ridges, grooves and so forth can be seen in the substrate, of a similar dimension.

2012). The same level of consumption applied also to ground Ballotini (Verrelli et al., 2012), when its surface area was taken as being 1/0.41 = 2.44 times greater than that of an ideal sphere in the same size fraction, as suggested by BET surface area measurements (Koh et al., 2009). These revised estimates differ from earlier estimates made on the basis of pH measurement (Koh et al., 2009).)

3.1.1. Surface area The specic surface area provides a quantitative indication of particle shape and roughness (Hiylmaz et al., 2006). Experimental results are given in Fig. 4. For the spheres used in the Milli-Timer, the BET surface area was found to be approximately 0.0365 m2/g, which is close to the expected value for a perfect sphere of 7590 lm diameter. For frit of the same nominal size fraction, the surface area was 0.0614 m2/g. This represents a ratio of 1/1.68 = 0.59, which can be interpreted as a sphericity-like measure (Koh et al., 2009), although conventionally the sphericity is dened as relating particles of equal volume (Wadell, 1935: 264), rather than equal size fraction. Recognising the uncertainty in estimating the small surface area of the spherical particles, and recalculating the ratio using the lower limit of the condence interval or the theoretically expected value for ideal spheres (which happen to be approximately equal here) the ratio becomes 1/1.85 = 0.54. Although in the present work only particles in a single size fraction were used, namely 7590 lm, the BET data indicates that particle asphericity increases as the particle size decreases; this is consistent with previous reports (Gaudin, 1926; Holt, 1981). Measured reagent consumption for the borosilicate spheres at 100% methylation indicates chemisorption of 10 mg/m2 of pure CTMS, assumed independent of particle size5, after correcting for reagent degradation and losses (Verrelli et al., 2012). The same consumption is obtained for borosilicate frit when its specic surface area is taken as 2.0 times the specic surface area of spherical particles in the same size fraction, implying a ratio of specic surface areas equal to 0.50. (For the sodalime glass Ballotini used previously, 18.5 mg/m2 of CTMS was consumed for full methylation (Verrelli et al.,
5

3.2. Induction period for attachment In Figs. 5 and 6 induction period results are presented for spheres and frit, respectively. We have previously described a theoretical rationale for the induction period to depend upon the polar angle at which the particle impinges on the bubble (Verrelli et al., 2012). Furthermore, models of the attachment process commonly posit that the attachment will be easier for particles approaching at higher velocities (i.e. greater kinetic energy) (Yoon and Luttrell, 1989; Yoon and Mao, 1996; Nguyen and Schulze, 2004: 265,268). Both of these inuences are depicted in the graphs. It should be noted that in the case of angular particles much greater variation in the results can be expected as a consequence of the anticipated dependence on the orientation of the impinging particle. For example, a cubic particle could impinge point-rst or face-on. The hydrodynamic resistances will assuredly be different, the relationship between gap and surface forces will be different, and even the surface chemistry may be different, so a different induction period could also be expected. However, as seen from the SEM images, the frit exists in a wide range of shapes, so that a simple quantication of the impinging particles orientation for each event is very difcult. Hence, the approach has been to gather more data for the angular particles, in order to provide a more representative selection of the possible combinations of particle shape and orientation. In short, the experimental data suggest that (on average) the induction period for attachment is reduced in the case of angular particles, and is also reduced at greater approach velocities. The direct dependence on polar angle seems weaker, but still suggests an increase in induction period for particles that impinge further from the apex (i.e. at larger polar angles), in accordance with predictions from numerical modelling (Verrelli et al., 2012).

In this calculation the surface area is estimated using the mid-range particle size.

86

D.I. Verrelli et al. / Minerals Engineering 58 (2014) 8089

(a) 200
180 160

(a) 60
Run A
Run C Run D Lowest v 0 Highest v 0

50

Induction time, [ms]

140

Induction time, [ms]

40

Run a Run b Run c Run d Lowest v 0 Highest v 0

120
100 80 60 40 20 0 0 10 20 30 40

30

20

10

0 0 10 20 30 40

Polar angle of sliding commencement, [ ]

Polar angle of sliding commencement, [ ]

(b) 200
180 160
Run A Run C Run D

(b) 60
50

Induction time, [ms]

140 120 100 80 60 40 20 0

Induction time, [ms]

Highest Lowest

40

Run a Run b Run c Run d Highest Lowest

30

20

10

10

20

30

10

20

30

40

50

Approach velocity, v 0 [mm/s]

Fig. 5. Induction period for attachment of 7590 lm borosilicate glass spheres to a 1.56 mm air bubble, as a function of (a) polar angle (Verrelli et al., 2012) and (b) approach velocity. Symbols are experimental observations; curves represent the range of the proposed correlation.

Approach velocity, v 0 [mm/s]

Fig. 6. Induction period for attachment of 7590 lm borosilicate glass frit to a 1.56 mm air bubble, as a function of (a) polar angle and (b) approach velocity. Symbols are experimental observations; curves represent the range of the proposed correlation.

It should be noted that the apparent dependence of induction period on approach velocity is not linked to the particle size, as might have been thought. In the experiments, the particles fall under the action of gravity, and it is known that larger particles (of the same shape and density), will fall faster. However, analysis of the data indicates that in the present set-up, variations in the approach velocity are predominantly due to the inuence of other particles i.e. swarm effects (Verrelli et al., 2012). Viewing individual scatterplots in isolation may not give a complete picture of the correlation if the parameter of interest depends on multiple variables (Montgomery et al., 2001: 84f). Hence, to assess the various proposed dependencies simultaneously we undertake multiple nonlinear regression, which permits a clearer evaluation of the importance of each explanatory variable. For this analysis data from all runs are combined. 3.2.1. Regression Regressions for each combination of terms produced varying levels of t, as indicated by the MSE in Fig. 7(a). The lowest values of MSE for each number of coefcients is shown in Fig. 7(b). Table 1

gives the best choice of coefcients in Eq. (1) for increasing numbers of included terms. The information in Table 1 corresponds to the points in Fig. 7(b). There is little benet in introducing more than 4 coefcients, and indeed 3 coefcients is sufcient to capture the inuence of each of the three main effects under consideration. The estimated coefcients are generally found to be statistically signicant; the only notable exception is for the case of 3 retained terms (when b1 is included). The most important inuence on induction time, statistically speaking, is the polar angle. If only one coefcient were allowed to be retained in Eq. (1), then the best t would be proportional to approximately 0.0222 /2; however, this is quite a crude approximation (seen by the large value of MSE in Table 1), and so it is desirable to retain more coefcients in the tting equation. When more coefcients are retained, the trend for larger induction times to be correlated with greater /2 is still consistently found; b7 is consistently estimated equal to approximately 0.1. This dependence accords with our previous numerical modelling (Verrelli et al., 2012). It does not appear to be described elsewhere in the literature. The difference between /2 and cos(/) is small, as

D.I. Verrelli et al. / Minerals Engineering 58 (2014) 8089

87

(a) 900
800

Mean squared error [ms]

700 600 500 400 300 200 100 0 0 2 4 6 8 10

Number of parameters

(b) 10

Any combination
Without 1/ v 010

Mean squared error [ms]

values uctuate. This is in agreement with more general studies comparing oatability of angular and round particles, as discussed earlier. Finally the inuence of approach velocity is introduced. Although the 1=v 10 0 term is a good t for the frit data, there is no obvious physical reason for such an unusual dependence. If this term is disallowed, the penalty is not too great, and the regression is then optimised by including a 1=v 2 0 term, which is related to the initial kinetic energy of the falling particle. (The term does not fully account for the energy, as this would require the particle mass, which is especially difcult to determine for the angular frit particles). Although separate models exist relating probability of attachment to the induction period (Yoon and Luttrell, 1989; Nguyen and Schulze, 2004: 265,268) and kinetic energy (Luttrell and Yoon, 1992), it is not clear from these models how the induction period should depend upon kinetic energy. In all cases the regression predicts increase in s with decreasing v0: the slower the particles approach, the longer the induction period. This is conceptually consistent with models of particlebubble attachment in which the two bodies need to overcome an energy barrier in order for attachment to occur (Luttrell and Yoon, 1992). The hydrodynamically-based models of induction period (see Nguyen and Schulze, 2004: 507ff) tend to suggest dependence on 1/v0; this is still compatible with the overall trend observed presently the data do not allow a precise judgement of the best exponent. Similarly, the experimental data of Gu et al. (2003) for a particle-pickup device is best t by an equation of the form

10

t50% attach a1

1
:14 v1 a

a2 :

In this device the bubble is squashed against a bed of particles, approaching at a speed va, and the kinetics may be quite different to the more natural motions in the present system. The constant term, b9, is not needed to describe our data. The suggested t to the data is:
10
1

10

" 2 73:8mm=s

Number of parameters
Fig. 7. MSE for multiple nonlinear regression according to Eq. (1). (a) All combinations of coefcients. (b) Optimal combinations of coefcients, either allowing or disallowing inclusion of a 1=v 10 0 term.

v0

/  3:11

2 # 0:0792f 1 f ;

cos(/) 1 is directly proportional to /2 for small /. The /2 term does not have an obvious physical interpretation, while the cos(/ ) term does. The next-most important inuence is the particle shape. In all cases the frit is estimated to have induction periods that are only a few percent of the equivalent spheres, although the precise

with s in milliseconds. Eq. (3) indicates that the induction period is reduced for angular particles, high-velocity particles, and particles impinging nearest the apex of the bubble. The correlation has been plotted in Figs. 5 and 6. Given that each individual graph includes only one of either / or v0 on the horizontal axis, the span of the other parameter is accounted for by plotting two curves: one for the largest value and one for the smallest value of that missing parameter. Considering the scatter of the data, the ts are satisfactory. The correlation provides a reasonable prediction of the average induction period under specied

Table 1 Optimal regressions of s on combinations of the candidate variables, using Eq. (1) with 14 coefcients. (For 30 , the 1=v 10 0 term is disallowed: b1 is set to 0). The uncertainties indicate the 95% condence intervals. Entries in italic denote high uncertainty. Entries in bold are the adopted data set. Coefcient Corresponding parameter Number of coefcients 1 b1 b2 b3 b4 b5 b6 b7 b8 b9 b10 MSE 1 =v 1 =v 1/v0 Approach velocity,
10 0 2 0

3 7.63 1011 8.82 1011

30 5.45 103 3.04 103

4 84.3 103 6.3 103 23.9 3.5

[s10/mm10] [s2/mm2] [s/mm] [mm/s] [mm2/s2] [] [2] [] [ms] [] [ms2] 0.0222 0.0117 0.124 0.013

v0

v2 0
Polar angle, / /2 cos(/) Constant Frit ag, f

0.123 0.010

0.103 0.016 698 76

0.0569 0.0412 717.8 116.8

0.0303 0.0311 64.2

0.0792 0.0362 97.0

0.0299 0.0071 39.4

88

D.I. Verrelli et al. / Minerals Engineering 58 (2014) 8089

interaction conditions. The correlation does not describe the stochastic variability, although this may be signicant. 4. Conclusions In the literature a number of papers have suggested increasing oatability of more angular (or rougher) particles. A few studies have judged this by otation recovery, but have not been able to quantify the induction period directly. In the present work we have been able to observe particlebubble interactions, and thus directly obtain estimates of induction period. Statistical analysis indicated that the induction period, s, depended on three parameters. In order of decreasing statistical signicance, these are: the polar angle of sliding commencement, /, an indicator variable, f, (equal to 1 for frit and 0 for spheres), and the approach velocity, v0. As particles become more angular, approach at higher velocity, or impinge closer to the bubbles apex, the induction period is reduced. Particle shape depends in part on the milling process. The results presented herein indicate the level of attention that should be paid to the shape of particles obtained from the grinding operation, besides particle size. It may be anticipated that the more angular the particle, the lower the induction period. An extension of the present work would be to measure the effect of different degrees of angularity on the induction period. Acknowledgements We gratefully acknowledge the assistance of numerous colleagues at CSIRO who assisted with sample preparation (Cathy Edwards), chemical analysis (Cheryl McHugh & team) and microscopy (Matthew Glenn), or advised on other aspects. Funding for this project was provided from CSIROs Process Science and Engineering Capability Development Fund. Appendix A. Regression formul The simplest regression equation to apply would be

issue for /, because previous work (Verrelli et al., 2012) indicates that although s decreases with decreasing /, there is a natural lower limit of /, namely / = 0. A further concern is that it seems more likely for the induction period for the frit to be a multiple or fraction of the corresponding value for spheres, whereas the foregoing equations all propose that they are interrelated by an offset, namely b3f. A multiplier can be implemented like so:

s b1

v0

 b2 / b3 b4 f 1 f :

The neat thing about Eq. (8) is that b4 has a clear interpretation as the factor by which one needs to multiply s for spheres to obtain the corresponding value for frit. It is not possible to be sure in advance what relationship will best describe the inuence of v0, / and f upon s. To leave a variety of options open, we perform the regression using the following general equation:

s b1

2 b4 v 0 b5 v 2 0 b6 / b7 /  b8 cos/ b9 b10 f 1 f :

v 10 0

b2

v2 0

b3

v0

s b1 v 0 b2 / b3 f b4 :

The true induction period may take on arbitrarily small values fractions of a millisecond but should not become negative. This suggests perhaps an equation of the form

logs b1 v 0 b2 / b3 f b4 :

The variances in s were assumed to be of similar absolute value, irrespective of the magnitude of s. Hence Eq. (5) would produce an inappropriate minimisation of the residuals (Montgomery et al., 2001: 420). The scheme could be returned to homoskedasticity by rearranging the formula to

s exp b1 v 0 b2 / b3 f b4 ;

but this has the disadvantage of requiring nonlinear regression. The main concern associated with prediction of negative values of s stems from the (unbounded) approach velocity, for which large values may be correlated with small values of s. Therefore an alternative tactic is regress s on, say, 1/v0, according to

s b1

v0

b2 / b3 f b4 :

Note that this is still a linear correlation, because 1/v0 is treated as an explanatory variable, rather than needing to be decomposed into the reciprocal of a variable. We are less concerned about this

The rst term, involving 1=v 10 0 , was obtained by optimising n in a linear regression in the general form s b=v n 0 + constant, for the frit data only. The optimum n does not change much when a /2 term is added. There is less data for the spherical particles from which to obtain a reliable estimate of the optimal exponent. For s b=v n 0 + constant, n seemed to be optimal at about 3; however, inclusion of a /2 term brought the optimum exponent back up to about 10, for the sphere data alone. Nonlinear regression was performed with MATLAB (The MathWorks), making use of the Statistics Toolbox. Initially it was expected that a stepwise-type regression procedure would be most practicable (Montgomery et al., 2001: 310ff). However, given the discordant results seen from different kinds of stepwise regression in preliminary investigations, the complication introduced by nonlinearity, and the power of the computational environment selected, more helpful information was derived from considering the set of all possible regressions (Montgomery et al., 2001: 302ff). In the case of a linear regression equation with 10 coefcients, that would imply 210 = 1024 combinations. The nonlinearity of Eq. (9) means that when the effect of f was to be neglected, instead of setting b10 to zero it was set to unity; also the case of only the coefcient b10 being retained was considered physically meaningless, so that only 1023 combinations were computed. The best regression for a given number of parameters was chosen based on the objective of minimising the mean squared error (MSE), which is similar to but preferable to maximising the relevant coefcient of determination, R2 (Montgomery et al., 2001: 47f.,90,296ff). Furthermore, the principle of parsimony is employed, in which extra terms are only included in the model if they lead to a substantial reduction in MSE. One other helpful tool to ensure that an excessive number of terms has not been retained in the model is to ensure all retained coefcients are statistically signicant. Statistical signicance can be assessed by checking whether the condence interval on the estimated coefcient excludes zero. Here the 95% condence interval is computed from the diagonals of the asymptotic covariance matrix and the relevant Students t-statistic (Montgomery et al., 2001: 428ff.,434ff). Finally, for completeness it may be noted that an alternative, linear equation,

D.I. Verrelli et al. / Minerals Engineering 58 (2014) 8089

89

s b1

v 10 0

b2

v2 0

b3

v0

2 b4 v 0 b5 v 2 0 b6 / b7 / b8

cos/ b9 b10 f ;

10

which replaces the multiplier effect with a bias effect, was also investigated. This produced less favourable results, including coefcients with uctuating signs (see Montgomery et al., 2001: 120ff), and is not considered further herein. Appendix B. Supplementary material Supplementary data associated with this article can be found, in the online version, at http://dx.doi.org/10.1016/j.mineng.2 014.01.004. References
Ahmed, M.M., 2010. Effect of comminution on particle shape and surface roughness and their relation to otation process. Int. J. Miner. Process. 94 (34), 180191. Albijanic, B., Ozdemir, O., Nguyen, A.V., Bradshaw, D., 2010. A review of induction and attachment times of wetting thin lms between air bubbles and particles and its relevance in the separation of particles by otation. Adv. Colloid Interface Sci. 159 (1), 121. Anfruns, J.F., Kitchener, J.A., 1977. Rate of capture of small particles in otation. Trans. Inst. Min. Metal. 86, C9C15. Austin, L.G., Rogers, R.S.C., 1985. Powder technology in industrial size reduction. Powder Technol. 42 (1), 91109. Bruckard, W.J., Sparrow, G.J., Woodcock, J.T., 2011. A review of the effects of the grinding environment on the otation of copper sulphides. Int. J. Miner. Process. 100 (12), 113. Brunauer, S., Emmett, P.H., Teller, E., 1938. Adsorption of gases in multimolecular layers. J. Am. Chem. Soc. 60 (2), 309319. Chan, D.Y.C., Klaseboer, E., Manica, R., 2011. Film drainage and coalescence between deformable drops and bubbles. Soft Matter 7 (6), 22352264. Coghill, W.H., Anderson, C.O., 1923. Certain interfacial tension equilibria important in otation. U.S. Bureau of Mines/Department of the Interior/University of Washington, Washington, D.C., U.S.A., p. 55. Dippenaar, A., 1982. The destabilization of froth by solids. I. The mechanism of lm rupture. Int. J. Miner. Process. 9 (1), 114. Feng, D., Aldrich, C., 2000. A comparison of the otation of ore from the Merensky Reef after wet and dry grinding. Int. J. Miner. Process. 60 (2), 115129. Filippov, L.O., Joussemet, R., Irannajad, M., Houot, R., Thomas, A., 1999. An approach of the whiteness quantication of crushed and oated talc concentrate. Powder Technol. 105 (13), 106112. Forssberg, E., Sundberg, S., Hongxin, Z., 1988. Inuence of different grinding methods on oatability. Int. J. Miner. Process. 22 (14), 183192. Forssberg, E., Zhai, H., 1985. Shape and surface properties of the particles liberated by autogenous grinding. Scand. J. Metall. 14 (1), 2532. Franks, G.V., Gan, Y., 2007. Charging behavior at the aluminawater interface and implications for ceramic processing. J. Am. Ceram. Soc. 90 (11), 33733388. Gaudin, A.M., 1926. An investigation of crushing phenomena. Trans. Am. Inst. Min. Metall. Eng. LXXIII, 253316. Gu, G., Xu, Z., Nandakumar, K., Masliyah, J., 2003. Effects of physical environment on induction time of airbitumen attachment. Int. J. Miner. Process. 69 (14), 235 250. Gl, A., 2006. Separation of microlite from zircon by otation. Can. Metall. Quart. 45 (1), 111116. Harvey, E.N., Barnes, D.K., McElroy, W.D., Whiteley, A.H., Pease, D.C., Cooper, K.W., 1944. Bubble formation in animals. I. Physical factors. Journal of Cellular and Comparative Physiology 24 (1), 122. Hiylmaz, C., Ulusoy, U., Bilgen, S., Yekeler, M., 2005. Flotation responses to the morphological properties of particles measured with three-dimensional approach. Int. J. Miner. Process. 75 (34), 229236.

Hiylmaz, C., Ulusoy, U., Bilgen, S., Yekeler, M., Akdogan, G., 2006. Response of rough and acute surfaces of pyrite with 3-D approach to the otation. J. Min. Sci. 42 (4), 393402. Holt, C.B., 1981. The shape of particles produced by comminution. A review. Powder Technol. 28 (1), 5963. Huh, C., Mason, S.G., 1974. The otation of axisymmetric particles at horizontal liquid interfaces. J. Colloid Interface Sci. 47 (2), 271289. Husemann, K., Bernhardt, C., Heegn, H., 1976. Feinmahlung und mechanische Aktivierung von Eisenpulver. Powder Technol. 14 (1), 4149 (In German). Koh, P.T.L., Hao, F.P., Smith, L.K., Chau, T.T., Bruckard, W.J., 2009. The effect of particle shape and hydrophobicity in otation. Int. J. Miner. Process. 93 (2), 128134. Krasowska, M., Krastev, R., Rogalski, M., Malysa, K., 2007. Air-facilitated three-phase contact formation at hydrophobic solid surfaces under dynamic conditions. Langmuir 23 (2), 549557. Kursun, H., Ulusoy, U., 2006. Inuence of shape characteristics of talc mineral on the column otation behavior. Int. J. Miner. Process. 78 (4), 262268. Luttrell, G.H., Yoon, R.H., 1992. A hydrodynamic model for bubbleparticle attachment. J. Colloid Interface Sci. 154 (1), 129137. Montgomery, D.C., Peck, E.A., Vining, G.G., 2001. Introduction to Linear Regression Analysis, Third ed. John Wiley & Sons, New York. Morris, G., Neethling, S.J., Cilliers, J.J., 2011. An investigation of the stable orientations of orthorhombic particles in a thin lm and their effect on its critical failure pressure. J. Colloid Interface Sci. 361 (1), 370380. Nguyen, A.V., Schulze, H.J., 2004. Colloidal Science of Flotation. Marcel Dekker, New York, U.S.A. Ofori-Sarpong, G., Amankwah, R.K., 2011. Comminution environment and gold particle morphology: effects on gravity concentration. Miner. Eng. 24 (6), 590 592. Oliver, J.F., Mason, S.G., 1977. Microspreading studies on rough surfaces by scanning electron microscopy. J. Colloid Interface Sci. 60 (3), 480487. Schulze, H.J., 1984. Physico-Chemical Elementary Processes in Flotation: An Analysis from the Point of View of Colloid Science including Process Engineering Considerations. Elsevier, Amsterdam, The Netherlands. Shergold, H.L., 1984. Flotation in mineral processing. In: Ives, K.J. (Ed.), The Scientic Basis of Flotation. Martinus Nijhoff, The Hague, The Netherlands, pp. 229287. Siboni, S., Della Volpe, C., Maniglio, D., Brugnara, M., 2004. The solid surface free energy calculation: II. The limits of the Zisman and of the equation-of-state approaches. J. Colloid Interface Sci. 271 (2), 454472. Sven-Nilsson, I., 1934. Einu der Berhrungszeit zwischen Mineral und Luftblase bei der Flotation. Kolloid Zeitschrift 69 (2), 230232 (In German). Verrelli, D.I., 2008. Drinking Water Treatment Sludge Production and Dewaterability, Department of Chemical and Biomolecular Engineering (Ph.D. thesis). The University of Melbourne, Melbourne, Australia, 966 pp. <http:// repository.unimelb.edu.au/10187/3521>. Verrelli, D.I., Koh, P.T.L., 2010. Understanding particlebubble attachment: experiments to improve otation modelling, In: Chemeca 2010. IChemE in Australia/EA/RACI/SCENZ-IChemE N.Z., Adelaide, Australia. Verrelli, D.I., Koh, P.T.L., Bruckard, W.J., Schwarz, M.P., 2012. Variations in the induction period for particlebubble attachment. Miner. Eng. 3638, 219230. Verrelli, D.I., Koh, P.T.L., Nguyen, A.V., 2011. Particlebubble interaction and attachment in otation. Chem. Eng. Sci. 66 (23), 59105921. Wadell, H., 1935. Volume, shape, and roundness of quartz particles. J. Geol. 43 (3), 250280. Wotruba, H., Hoberg, H., Schneider, F.U., 1991. Investigations on the separation of microlithe and zircon the inuence of particle shape on oatability. In: XVII International Mineral Processing Congress. Federal Republic of Germany, Dresden, pp. 8395. Yekeler, M., Ulusoy, U., Hiylmaz, C., 2004. Effect of particle shape and roughness of talc mineral ground by different mills on the wettability and oatability. Powder Technol. 140 (12), 6878. Yoon, R.-H., Mao, L., 1996. Application of extended DLVO theory, IV: derivation of otation rate equation from rst principles. J. Colloid Interface Sci. 181 (2), 613 626. Yoon, R.H., Luttrell, G.H., 1989. The effect of bubble size on ne particle otation. Miner. Process. Extr. Metall. Rev.: An Int. J. 5 (14), 101122.