How a drug crystal dissolves in a polymer melt: diffusion, channels and fragmentation of indomethacin in copovidone

Background and motivation

Hot melt extrusion (HME) is one of the main ways to make amorphous solid dispersions (ASDs) of poorly soluble drugs. A powder blend of crystalline drug and amorphous polymer is heated and mixed until it forms a single homogeneous melt, which then cools to a glass. When the process runs above the drug's melting point, the drug melts and mixes; when it runs below, the polymer acts as a very viscous solvent and the drug crystals have to dissolve in it (Introduction, citing Moseson and Taylor 2018, ref 32) C2. Crystals that do not finish dissolving remain as residual crystallinity, which can seed crystallization of the amorphous drug later.

How crystals dissolve in such a melt has been described with the same physics used for dissolution in water: the Noyes–Whitney equation, in which the rate depends on the crystal's surface area, the diffusivity of the drug and its solubility (refs 31, 32, 40, 41) C2. This picture treats each crystal as a particle that shrinks smoothly as drug diffuses away. The authors point out that it leaves out two things: the properties of the crystal itself, such as defects, and mechanical effects such as breakage during mixing C2.

The wider literature on organic crystals gives reasons to think defects matter. Vacancies, impurities, dislocations and grain boundaries are common, carry higher energy and mobility than the perfect lattice, and "may" initiate melting, solubilization or mechanical failure; crystal defects have been linked to faster dissolution, and defects impose local stress (refs 42–53, including Burt and Mitchell 1981, ref 48) C3. Different crystal faces dissolve at different rates in water (Maddar 2017, ref 2; Adobes-Vidal 2016, ref 3), even though most dissolution models assume a sphere C3. Lattice Monte Carlo simulations of minerals and amorphous solids have produced etch pits and irregular dissolution at defects (refs 63–66) C4. The authors also argue more generally that dissolution is usually followed from the solution side, by the appearance of dissolved molecules, with little attention to how the crystal itself changes; this "potentially" leads to wrong conclusions about mechanisms (refs 1–3) C1.

The direct trigger for this paper was the group's own earlier work. In a nominally crystal-free hot melt extruded ASD, they had found nanometre-scale residual crystals that contained a range of defects, with preliminary signs that the crystals had been dissolving by fragmentation (Moseson 2018, ref 67) R22. But an extruder applies strong shear and kneading, so fragmentation there could have mechanical causes; the authors note that "other explanations are also possible" C5. The question for this paper was therefore whether crystals fragment while they dissolve in a polymer melt when there is no mechanical input at all C6.

To answer it, the authors held indomethacin (IDM) crystals suspended in copovidone films at 130 °C without any stirring and followed them by X-ray diffraction (XRD), polarized light microscopy (PLM), scanning electron microscopy (SEM) with energy-dispersive X-ray analysis (EDX), and transmission electron microscopy (TEM), from the scale of whole crystals down to crystallites a few nanometres across P1·P3–P8. They also built a two-dimensional lattice Monte Carlo (MC) model of dissolution with three different surface-energy maps P9. The authors "believe" this is the first study of the mechanism of crystal dissolution into a polymer melt and the first report of a fragmentation-based microstructural progression C7.

Background concepts

Dissolving below the melting point. A crystalline drug can dissolve in a polymer at temperatures well below its own melting point if the two are miscible; this is the familiar melting point depression of drug–polymer mixtures (refs 32, 85–87) C38. Indomethacin and copovidone (Kollidon VA64, a vinylpyrrolidone–vinyl acetate copolymer) have favourable specific interactions and are miscible at high temperature (refs 32, 87–91) R26. From the group's temperature–composition phase diagram, the solubility temperature of the composition used here, 1:5 drug:polymer by weight, is about 78 °C (ref 32) R21. The experiments ran at 130 °C, which is 31 °C below the drug's melting point of 161 °C and above the polymer's glass transition temperature (Tg) of 104 °C, so all of the drug could in principle dissolve if given enough time M1·M2·C38. The authors describe 130 °C as "approximately 20 °C above" the Tg; with the printed Tg of 104 °C the margin is 26 °C C51·X4.

The diffusion picture of dissolution. The paper writes the classical model as the Noyes–Whitney equation combined with the Nernst–Brunner diffusion layer (Equation 2) E2:

Here $C$ is the concentration of dissolved drug in the liquid (here, the molten polymer) at time $t$, $D$ is the diffusion coefficient of the drug, $A$ is the crystal surface area available for dissolution, $C_s$ is the equilibrium solubility of the crystalline drug in the molten polymer, $h$ is the thickness of the boundary layer through which drug diffuses away from the crystal surface, and $V$ is the volume of the liquid E2. In plain terms: dissolution is faster when the drug diffuses quickly, when there is more crystal surface, and when the surrounding melt is far from saturated; it is slower when the diffusion layer is thick. For a very viscous medium such as a polymer melt, the authors note that the diffusivity can be estimated with the Stokes–Einstein equation (Equation 3) E3:

Here $k_B$ is the Boltzmann constant, $T$ the absolute temperature, $\eta$ the viscosity of the medium and $r$ the radius of the diffusing molecule E3. In plain terms: the more viscous the melt and the larger the molecule, the slower it diffuses. Together, the two equations say that dissolution in a melt is controlled by diffusivity, temperature, viscosity, solubility, surface area and boundary-layer thickness C39. In this picture (Figure 16a) a boundary layer forms on all crystal surfaces and the crystal shrinks evenly from every face until it is gone, provided the solubility limit is not reached C40. The paper uses these equations only as a framework; it reports no values of $D$, $h$ or $\eta$ E2·E3.

Crystal defects and surface energy. A real crystal is not a perfect lattice. Growth irregularities, grain boundaries (interfaces between regions of different lattice orientation), dislocations and point defects all raise the local energy, and imperfect surfaces show both a higher and a more uneven surface energy than smooth ones (refs 60–62) C3. The authors' central idea is that such high-energy sites dissolve faster than the rest of the crystal and so cut channels into it C44.

Reading crystallinity in TEM. In bright-field TEM of a drug–polymer film, amorphous material looks like a homogeneous continuous phase, whereas crystals show up as darker areas, as diffraction contrast (contrast that changes as the crystal or the focus is tilted), and, at high magnification only, as lattice fringes (refs 67, 81–84) C18. Lattice fringes are only visible where the crystal is thin enough and suitably oriented; samples must be roughly 100 nm thick to be electron-transparent C18. Where fringes are visible, a fast Fourier transform (FFT) of the image gives spots, each corresponding to one set of lattice planes at one orientation. The authors mask one spot pair, transform back, and colour the result; this shows where in the image that particular lattice orientation lies, so fragments that have rotated relative to each other appear in different colours P8. Bend contours are dark bands that appear where a thin crystal is bent or strained R3.

Lattice Monte Carlo dissolution. In the MC model, the crystal is a square grid of sites, each either crystal or dissolved. At each trial step one crystal site is picked at random and the move "remove this site" is accepted or rejected with a probability that depends on energy. The paper gives this probability as Equation 1 E1:

Here $P(x)$ is the probability of accepting the removal, $E_{\mathrm{s}\mathrm{u}\mathrm{r}\mathrm{f}\mathrm{a}\mathrm{c}\mathrm{e}}$ is the surface energy assigned to the trial site by the chosen map, and $\Delta E_{\mathrm{i}\mathrm{n}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{f}\mathrm{a}\mathrm{c}\mathrm{e}\mathrm{s}}$ is the change in energy from creating or removing crystal–melt interfaces, with a dimensionless penalty of 4.0 for each new interface E1·P9. The move is accepted when $P(x)$ is at least as large as a random number drawn uniformly between 0 and 1 E1. In plain terms: removing a site that would leave many new exposed faces is unlikely, removing a site at a corner or edge is likely, and the site's own surface energy also enters. The number of trial steps has no direct time scale; the authors use the number of steps needed to reach a given fraction dissolved as a stand-in for relative dissolution rate C35. As printed, Equation 1 would make a high-energy site less likely to dissolve; the published code does the opposite, which is what the simulations show. This is discussed under Methods X2.

Methods in detail

The paper has three contributions: a multi-scale imaging study of crystals dissolving in a still melt, a lattice Monte Carlo model with defect-like surface-energy maps, and a conceptual model that sets defect-driven fragmentation beside the diffusion picture. The workflow is shown below.

Contribution 1: Watching crystals dissolve in a still melt at four length scales

① Problem. Dissolution of crystals in an extruder happens in seconds to minutes, under shear, inside an opaque melt; residual crystals can be examined only after the fact. To see whether crystals fragment on their own, dissolution has to be slowed down and watched without any mechanical input C6·C51.

② How. The authors used crystalline indomethacin with a particle size distribution of D10 6.3 ± 0.1 µm, D50 21.9 ± 0.3 µm and D90 63.4 ± 0.9 µm by laser diffraction (n = 3) M1·P2. Because indomethacin is poorly soluble in water, the crystals could be suspended in an aqueous copovidone solution and cast as films in which the crystals stay crystalline; the ratio was 5:1 copovidone to drug, that is 1:5 drug to polymer P1. Films were cast on a zero-background silicon plate for XRD, on glass slides for PLM (and then SEM of the same slides), and on glow-discharged 300-mesh carbon-coated copper TEM grids P1·M3–M5. The paper does not report film thickness, the amount deposited or the drying conditions P1.

• XRD. The film was scanned at room temperature, heated to 130 °C over about 10 min in a high-temperature chamber, and scanned from 10 to 30° 2θ, each pattern taking 15 min, for up to 3 h (Figure 4 shows 18–23° 2θ) P3.
• PLM. Films on glass were held at 130 °C on a hot stage for up to 3 h and imaged between crossed polarisers with a 20× objective; Figure 5 shows frames every 15 min up to 120 min P4.
• SEM and EDX. The slides from the PLM work were platinum-coated and imaged at 10 kV with secondary and backscattered electrons; EDX maps and line scans gave the atomic percentages of C, O, N and Cl P5·P6. Copovidone contains no element that indomethacin lacks, but indomethacin contains chlorine, so the chlorine signal marks the drug, whether it is still crystalline or already dissolved in the film C15.
• TEM. Films on grids were held at 130 °C for 0, 30 min, 2 h or 3 h on the PLM hot stage, cooled, and imaged in bright field at 200 keV; at least three grid squares were examined per sample P7. Where lattice fringes were visible, single-spot FFT filtering coloured each lattice orientation P8.

③ Why it works. The melt is very viscous and the drug diffuses slowly, so dissolution at 130 °C takes hours rather than minutes, and fragments stay close to where they formed instead of drifting away C51·C30. This lets the authors see stages that would be invisible in a stirred liquid. Each technique covers a different scale: XRD the whole film, PLM single crystals tens of micrometres across, SEM micrometre-scale surfaces and the chlorine distribution, TEM crystallites down to a few nanometres A2.

④ Difference from earlier work. Earlier HME work looked at residual crystals after extrusion, where shear and dissolution act together R22·C5. Studies of dissolution kinetics in water typically measure dissolved drug, not crystal shape C1. Here the crystal is followed in a melt without shear, and the microstructure inside the crystal is examined.

Contribution 2: A lattice Monte Carlo model with defect-like energy maps

① Problem. Images show what happens but not why. The authors wanted to test whether high-energy sites in a crystal are enough, on their own, to produce channels and fragmentation C46.

② How. The SI gives the full Python code. A 500 × 500 grid has its border set vacant, leaving 498 × 498 crystal sites P9. At each trial step a random crystal site is chosen and the energy change of removing it is computed from its four nearest neighbours and from the surface-energy map (SI code; card equation E4) E4:

The removal is accepted if $\exp (-\Delta E)\ge 1$, or otherwise if $\exp (-\Delta E)$ exceeds a uniform random number E4. In plain terms: each neighbour that is still crystal means a new interface and costs 4, each neighbour already dissolved means an interface removed and gains 4, and the site's own surface energy makes removal easier. A site in the middle of a perfect crystal (four crystal neighbours) has $\Delta E$ = 16 and is almost never removed, while a site at a straight edge (three crystal neighbours, one vacant) has $\Delta E$ = 8 and a probability of about $3\times {10}^{-4}$ per trial (AI arithmetic from the code values) E4. Three surface-energy maps were used (Figure 14) P9·E5:

• No defects: surface energy 0 everywhere, an ideal crystal R17.
• Grain boundary: a diagonal band of surface energy 7 (width parameter 11 in the code) across an otherwise zero map, representing the boundary between two crystallites R18·P9.
• Multiple defects: three randomly placed Gaussian peaks of height 7, each with two orthogonal lines of value 7 (width 3) at a random angle through its centre, representing point defects and the strain they induce R19·E5.

The Gaussian peaks are defined in the code as (card equation E5) E5:

Here $x$ and $y$ are lattice positions, $x_k$ and $y_k$ the random defect centres, 7 the peak height and 50 the spread (one tenth of the grid size) E5. In plain terms: around each defect the surface energy rises smoothly to 7 at the centre, and lines of high energy run out from it. Each configuration was run 10 times; Figure 15 shows the mean fraction dissolved with one standard deviation P9·R20.

The sign of the surface-energy term. Equation 1 in the paper reads $P(x)=\exp (-(E_{\mathrm{s}\mathrm{u}\mathrm{r}\mathrm{f}\mathrm{a}\mathrm{c}\mathrm{e}}+\Delta E_{\mathrm{i}\mathrm{n}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{f}\mathrm{a}\mathrm{c}\mathrm{e}\mathrm{s}}))$, which would make a high-energy site less likely to dissolve E1. The SI code subtracts $E_{\mathrm{s}\mathrm{u}\mathrm{r}\mathrm{f}\mathrm{a}\mathrm{c}\mathrm{e}}$ instead, so that high-energy sites dissolve more readily, and this is what Figures 14b and 14c show: dissolution starts along the high-energy line and at the high-energy spots E4·R18·R19. Note by the AI writer (not discussed in the paper): the printed Equation 1 appears to have the wrong sign on $E_{\mathrm{s}\mathrm{u}\mathrm{r}\mathrm{f}\mathrm{a}\mathrm{c}\mathrm{e}}$; the code form, $P=min(1,\exp (E_{\mathrm{s}\mathrm{u}\mathrm{r}\mathrm{f}\mathrm{a}\mathrm{c}\mathrm{e}}-\Delta E_{\mathrm{i}\mathrm{n}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{f}\mathrm{a}\mathrm{c}\mathrm{e}\mathrm{s}}))$, is the one consistent with the results X2. A second mismatch concerns run length: the printed code runs a fixed 30,000,000 trial steps with the defect-free configuration selected and saves a snapshot every 10,000,000 steps, whereas Figures 14 and 15 show runs up to 110 million steps and snapshots at 5, 15 and 25 million steps; the paper text says each simulation continues until no crystal units remain X3. The code is therefore a template rather than the exact script behind each figure X3.

③ Why it works. Because the model has no diffusion or chemistry, any channel or fragment it produces comes purely from the geometry of the energy map. If a uniform crystal only shrinks from the edges while a crystal with high-energy lines splits, then high-energy features are sufficient to cause fragmentation in this kind of model C46.

④ Difference from earlier work. Similar kinetic Monte Carlo models had been used for dissolution pits in minerals and for dissolution of amorphous solids (refs 63–66) C4. Here the approach is applied to an organic drug crystal in a melt and compared directly with images.

Contribution 3: Two models of crystal dissolution in a polymer melt

① Problem. The diffusion-based model (Figure 16a) is the prevailing description of how a homogeneous drug–polymer melt forms, but it predicts only even shrinkage C40.

② How. The authors propose a defect-site-driven dissolution and fragmentation model (Figure 16b) C44. Once the diffusion boundary layer forms around a crystal, diffusion is accelerated at higher-energy defect sites on the surface, "likely" because the thermodynamic driving force for dissolution is greater there, giving a steeper concentration gradient C44. The authors "propose" that these sites are defects intrinsic to the crystal or the ordinary surface-energy heterogeneity of crystal surfaces C44. Faster dissolution and the stress fields around defects make channels that connect these sites; once channels from different areas meet, the parent crystal breaks into smaller crystallites, whose new surfaces dissolve by diffusion or split again C44. The sequence is summarised below.

③ Why it matters. In the diffusion model, the surface area falls steadily as the crystal shrinks. In the fragmentation model, splitting creates new surface and new particles, so the surface area and the particle count can rise during dissolution C55·C61. The authors connect this to HME, where mixing stresses "may" also act on the defect stress fields and break particles C45.

Experimental analysis

The starting crystals already carry defects. By SEM, the bulk indomethacin crystals are plate-like single crystals with small crystals stuck to a generally smooth surface, together with visible growth irregularities and surface imperfections (Figure 2a) R1. Cast into the copovidone film and not yet heated, they keep this form; one crystal in the PLM image carries a growth irregularity that is the same feature at which crystal B later splits (Figure 2b, Figure 5) R2·R5. TEM of unheated crystals (Figure 3) shows a consistent internal structure with gradual thickness changes and growth layers in a crystal without polymer, and the same smooth thickness change in a crystal in the film R3. In another particle, grain boundaries divide a flat crystal into three regions whose diffraction contrast moved independently when the focus was changed, which the authors take as "confirming" that the particle is crystalline and that the three regions have different orientations R3·C8. A fourth particle shows discrete crystallites on its surface and bend contours, which the authors attribute to these surface crystallites and to strain at the crystal–polymer interface R3·C9.

XRD: the crystalline peaks fade over two hours. At 130 °C the film shows the same indomethacin peaks as at room temperature, slightly shifted, which the authors attribute to the temperature difference (Figure 4) R4·C10. Table A gives the extractor's reading of each peak over time.

Table A. Visibility of the three indomethacin peaks in the XRD patterns at 130 °C, read from Figure 4 (approximate; "clear", "weak" and "none" are the extractor's judgement against the noise of each trace) R4.

The text states that the last detectable peak, at 20.7° 2θ, disappeared at approximately 120 min, and the discussion says the crystalline phase was detected through 2 h R4. The authors note that detectability by XRD "may be limited" by dilution of the drug in the polymer, by crystallite size and by crystal quality C11. Two details matter when reading Figure 4. Each pattern integrates 15 min and the clock starts with the first pattern after an approximately 10 min ramp to 130 °C, so XRD times are not directly the hold times used for PLM and TEM X10. And although the authors say the 130 °C pattern "matches" the ambient one, the tallest peak changes from about 19.55° at ambient to about 20.7° at 130 °C; the match is in peak positions, not relative heights X9.

PLM: single crystals split, open channels and holes. Under the light microscope at 130 °C, the smallest crystals dissolved within 30 min, leaving larger crystals to follow (Figure 5) R5. A halo formed around each crystal, which the authors read as a refractive-index change in the film caused by the composition or thickness change as dissolved drug spreads ("This suggests") R5·C12. Table B lists the events in the three crystals shown.

Table B. Events in the time-lapse PLM series of Figure 5, as stated in the text; the last row is read from Figure 5 R5.

The PLM series is the strongest single piece of evidence in the paper, because it follows the same crystals over time: the splits are watched, not inferred from different particles R5. It also shows both routes in action. Crystal B splits at its visible growth imperfection, but holes and channels also appear where no imperfection was visible, which the authors read as "indicating that the defects do not need to be on the macroscale" to affect dissolution R5·C47·C48.

SEM and EDX: irregular surfaces and drug spreading into the melt. In backscattered-electron images the crystals look lighter than the film, which the authors attribute to the higher atomic number of chlorine ("presumably") R6·C13. After 1 h the partially dissolved crystals have irregular surface patterns, unlike the generally smooth starting crystals; a small crystallite shows channels before it comes apart; and crystal fragments lie below the film surface after 1 h and at the surface after 2 h (Figure 6) R7·C14. The chlorine maps and line scans in Figure 7 compare an unheated crystal with one that had nearly dissolved after 1 h; Table C summarises the line scans.

Table C. EDX line scans of Figure 7, read from the figure (approximate) R8·R9.

The authors read the plateau as the footprint of the original crystal, about 30 µm wide, and the sloping edges as the gradient of molecularly dissolved drug, about 8 µm on each side after 60 min; the gradient stays narrow because diffusion in the viscous melt is slow, and it is driven by concentration, "or (more accurately) activity", gradients C41·C42. Because the electron beam also samples the matrix, the chlorine level at the crystal appears lower than that of a pure indomethacin crystal C17. Note by the AI writer (not discussed in the paper): the reference composition for indomethacin in Figure 1a lists 10.4% chlorine as "atomic %" without hydrogen, but the values in that table reproduce mass percentages; on the atomic basis that EDX reports, pure indomethacin is 4.0% chlorine (1 of 25 non-hydrogen atoms). The unheated crystal (about 3.4–5.1 atomic %) is therefore close to the theoretical value, and the plateau after 1 h (about 2.8–3.1) is below it X1·R8·R9.

TEM: channels, splits and rotated fragments. TEM can image only crystals thin enough after partial dissolution, so each time point samples a different population; the authors do not expect the mechanism to change with time, only what can be captured (Section "Time Progression of Crystal Dissolution") C29. Table D collects what each image set shows.

Table D. TEM and SEM observations of dissolving crystals, from Figures 8–13 and the text R10·R11·R12·R13·R14·R15·R16.

The 30 nm crystallite of Figure 11c–f is the clearest nanoscale case. Its four reconstructed fragments lie at 47.3° (red), 53.6° (blue), 62.0° (yellow) and 66.1° (green); the blue spot is the sharpest, so the authors take it as the original orientation, with red turned by roughly 6° and yellow and green by about 8° and 12° R14·C25. The red/blue pair and the yellow/green pair cover about the same area in the image plane, so the authors describe them as delaminated layers, and they count three fragmentation sites R14·C25. The rotation "presumably" comes from slow motion of the crystallites in the matrix and from the diffusion of drug out of, and polymer into, the crystal region, which pushes the fragments apart C26. All TEM images were taken after cooling to room temperature, and for Figure 12 the authors themselves allow that the movement "may also" have happened during cooling C28·X11.

Monte Carlo: defects give channels, fragments and faster dissolution. Without defects, the simulated crystal loses sites only at its edges and rounds towards a circle by 70 million steps; its later departure from a circle, ending as a small diamond at 110 million steps, is "most likely" an artifact of building a small circle from square sites (Figure 14a) R17·C32. With a grain boundary, dissolution runs along the high-energy diagonal: a channel appears by 10 million steps, widens by 20 million and splits the crystal in two by 30 million, after which each half dissolves from its edges (Figure 14b) R18·C33. With multiple defects, dissolution starts at the high-energy spots and lines, forms extensive channels by 15–20 million steps and fragments the crystal by 20–25 million, after which the fragments shrink from their edges (Figure 14c) R19·C34. Table E gives the fraction dissolved.

Table E. Fraction of the lattice dissolved versus number of Monte Carlo trial steps, mean of 10 runs, read from Figure 15 (approximate; selected rows) R20.

At 30 million steps, where Figures 14a-2, 14b-3 and 14c-6 are taken, about 6%, 23% and 72% of the crystal had dissolved without defects, with a grain boundary and with multiple defects R20. Using step count as a surrogate for relative rate, the authors conclude that defects and high-energy sites shorten the simulations and are "thought to" increase the rate of dissolution, and that channel formation and fragmentation is rapid compared with diffusion-based dissolution alone C35·C36·C37. They describe the agreement between the multiple-defect simulation (Figure 14c) and their conceptual model (Figure 16b) as "excellent" and say it supports the usefulness of the defect-site model C46.

XRD against microscopy. Putting the techniques side by side, the XRD peaks were gone from about 120–135 min, while PLM still showed birefringent crystals at 120 min and TEM showed crystalline fragments after 3 h R4·R5·R11·R12·R15. The authors argue that fragmentation "may also" hinder XRD detection, because both the falling crystalline content and the peak broadening of very small crystallites reduce the signal, and that methods such as XRD "may lead to an incorrect conclusion that the phase transformation is complete" C54.

Discussion

What the paper establishes. The authors write that, besides diffusion, "a fragmentation process has been clearly identified", first suggested by TEM of extrudates and now confirmed under quiescent isothermal conditions C43. In the Conclusion they state that a fragmentation-based mechanism for dissolution of organic crystals into a polymer melt "was identified" from microscopy at multiple length scales C59. The observation itself is well documented: channels, holes and splits appear in single crystals followed by PLM, in partially dissolved crystals by SEM, and in nanometre-scale crystallites by TEM R5·R7·R10–R16. Both routes coexist: some crystallites shrink evenly, others split, and the fragments then shrink in turn C24·C50.

What remains a hypothesis. The cause of the channels is stated more cautiously. Defects intrinsic to the crystal, or induced mechanically, "are hypothesized" to be the initiation sites, and their stress fields are said to accelerate dissolution, form channels and fragment the parent crystal C60. The evidence for this is indirect: visible growth imperfections and grain boundaries in unheated crystals, one crystal (B) that split at its imperfection, and simulations in which high-energy lines and spots produce channels R1–R3·R5·R18·R19. No defect density, defect type or stress was measured in the crystals that fragmented X5.

Generality. The authors suggest that a fragmentation-based mechanism "may be universally found in other solute−solvent systems" but goes unseen because particles disperse in ordinary experiments; they point to atomic force microscopy studies of single crystals in unstirred water that showed surface roughening and accelerating dissolution (Unwin and colleagues, refs 2, 3) C52·R24. Alternatively, viscous solvents such as polyethylene glycols, lipids and polymers "may have unique properties" that bring the mechanism about C52.

Implications for hot melt extrusion. The authors draw three consequences. First, a recent HME process model assumes a constant number of particles and a matching fall in surface area during dissolution; because fragmentation changes the particle number and surface area, a population balance-type model with a changing number of particles "is therefore likely necessary to accurately capture the dynamics of the dissolution process" R25·C55. Second, mechanical damage in the extruder is a complementary source of defects and breakage, and its contribution "would be expected to have a significant impact", particularly with dispersive mixing or kneading elements C56. Third, since particle size is already known to matter kinetically, the identification of a defect-driven mechanism "further suggests" that crystal quality (defect density) must also be controlled for adequate control of an HME process; citing Burt and Mitchell (ref 48), who found that crystals of higher defect density dissolve faster, they add that their work "would further suggest" these defects are where fragmentation starts C57·C58·R23. Their closing line is that "crystal quality may be an unexplored critical material attribute" C61.

Implications for testing. The comparison of XRD with microscopy carries a practical warning for anyone checking extrudates for residual crystals: a featureless XRD pattern does not mean the crystals are gone C54. Note by the AI writer (not discussed in the paper): in this study the crystals still visible by PLM at 120 min were tens of micrometres across, far too large to be hidden by size broadening, so the loss of the XRD signal is also consistent with too little crystal remaining to detect, which is the dilution factor the authors themselves list; no peak widths were reported to separate the two causes X7·C11.

Limitations

Limitations stated by the authors. The authors note several limits. XRD detection "may be limited" by dilution, crystallite size and crystal quality C11. In EDX, the beam samples the matrix as well as the crystal, so the chlorine level at the crystal appears lower than for pure indomethacin C17. In TEM, crystal lattices can be seen only where the sample is thin enough, so thickness variations obscure parts of the crystals and each time point shows a different population C18·C29. The defect-free simulation departs from a circle late in the run, "most likely" an artifact of the square lattice C32. The step count of the simulations has no direct time scale C35. And the orientation changes seen in Figure 12 could come from rotation after fragmentation, from a parent crystal that already had several orientations, or partly from the cooling process C28.

The AI writer's own judgement (1): what is observed and what is inferred. Fragmentation as an event is directly observed, most convincingly in the PLM series where single crystals are followed in time R5. The defect origin is not: no defect density, defect type, local surface energy or stress was measured for the crystals that split, only one crystal visibly split at a growth imperfection, and others formed channels without any visible imperfection X5·C48. The Monte Carlo energy maps (a peak height of 7 on a baseline of 0, an interface penalty of 4.0) are chosen, not derived from indomethacin, so the agreement between Figure 14c and Figure 16b shows that high-energy features are sufficient to cause fragmentation in such a model, not that they cause it in these crystals X5. Likewise, the claim that fragmentation speeds dissolution rests on simulation step counts and on a cited study; no dissolution rate or time to complete dissolution was measured for crystals of different defect content X6. The nanoscale sequence from channel to split to fragment field is assembled from different particles imaged after cooling, except for the particle of Figure 9 X11.

The AI writer's own judgement (2): inconsistencies in the source. Three points in the paper do not agree with each other. First, printed Equation 1 carries the surface energy with the opposite sign to the published code, and only the code's sign reproduces the behaviour shown in Figures 14 and 15 X2. Second, the published code runs a fixed 30 million steps with the defect-free configuration and saves snapshots every 10 million steps, whereas the figures show runs to 110 million steps and snapshots at 5, 15 and 25 million; the settings and seeds behind each figure are not given X3. Third, the relative elemental composition table in Figure 1a is labelled "atomic %" but reproduces mass percentages without hydrogen, except indomethacin carbon, which matches the mass percentage with hydrogen (63.8 instead of 66.8); the indomethacin column sums to 97.0 X1. A smaller point: 130 °C is described as about 20 °C above the copovidone Tg, but with the printed Tg of 104 °C it is 26 °C above X4. None of these changes the qualitative conclusion that crystals fragment, but the first two matter for anyone who wants to rerun the simulations.

The AI writer's own judgement (3): scope. The study covers one drug, one polymer, one composition (1:5) and one temperature, with no shear P1·A1. At 130 °C the melt is far above the cited 78 °C solubility temperature of this composition, so the drug is never close to its solubility limit, which is the regime where residual crystals are most likely in practice R21. The indomethacin polymorph is not stated, and the film thickness, amount of drug per film and drying conditions are not reported M1·P1.

Conclusion

The paper makes three contributions. First, by holding indomethacin crystals in molten copovidone at 130 °C without mixing and following them by XRD, PLM, SEM-EDX and TEM, it shows that crystals dissolve both by diffusion, with a chlorine gradient of about 8 µm around a nearly dissolved crystal after 1 h, and by channel formation and fragmentation, from single crystals tens of micrometres across down to domains of about 6 nm R9·R5·R15·C59. Second, a two-dimensional lattice Monte Carlo model shows that high-energy lines or spots are enough to produce channels and fragments and to raise the fraction dissolved at 30 million steps from about 6% to about 72% R20·C37. Third, the authors propose a defect-site-driven model in which defects, intrinsic or mechanically induced, "are hypothesized" to start the channels C44·C60.

For formulation and process work, the authors' message is that HME dissolution models should allow the particle number to change, that the defect state of the input crystals "may be an unexplored critical material attribute", and that XRD alone can miss residual crystals C55·C61·C54. What the paper does not yet provide is a measurement linking defect density to where and how fast crystals split, or a real-time dissolution rate showing how much fragmentation speeds the process X5·X6. Those are the experiments that would turn the mechanism into a design rule.