Water in the spray solvent: how a few percent of water makes spray dried ritonavir–copovidone dispersions phase separate, and what it does to release

Background and motivation

An amorphous solid dispersion (ASD) puts a poorly soluble drug into a polymer in the amorphous state, which raises the concentration the drug can reach in water. ASDs are kinetically trapped rather than stable, and they tend to separate into a drug-rich and a polymer-rich amorphous phase, most often when water is present. The introduction lists the stages where this has been seen: manufacturing (Mugheirbi 2017, ref 8; Chen 2018, ref 9), storage (Rumondor 2011, ref 10; Qian 2010, ref 11) and dissolution (Purohit 2017, ref 12), with crystallization and altered release as consequences (refs 13–15) C11.

Spray drying is the usual commercial route for drugs that dissolve in a low-boiling solvent, and the properties of spray dried dispersions (SDDs) are known to depend on solvent composition, solids loading, drying rate and dryer geometry (refs 40, 45–47). What was missing, the authors argue, is an understanding of how these process parameters change the phase behavior of the ASD; most studies of ASD phase behavior relied on calculated miscibility (refs 34–39) or on model samples made by spin coating, rotary evaporation, film casting or melt methods (refs 15, 23, 40–44) C1. The group's own earlier work had shown that trace water in an otherwise non-aqueous solvent causes substantial phase separation in spin-coated films (Li 2017, ref 23) and in rotary-evaporated dispersions (Li 2019, ref 15). The same could happen in spray drying, because water enters spray solvents from the air (methanol is hygroscopic) or is added on purpose to dissolve a polymer or surfactant; and solvent pairs used for cellulosic polymers, such as methanol–dichloromethane, change composition during drying in the same way because their boiling points and polarities differ C14·C46·C47.

A second problem is detection. DSC is the usual "gold standard" (Baird 2010, ref 16), but a single glass transition temperature ($T_g$) does not prove homogeneity (Qian 2010, ref 11): domains below about 100 nm or phases with similar $T_g$ can escape it, and heating during the scan can mix or demix the sample (Purohit 2015, ref 21) C12·R24. Atomic force microscopy and transmission electron microscopy need flat or thin model samples, and solid-state NMR is costly, slow and not very sensitive to small amounts of a second phase. Two surface-sensitive alternatives had been applied to ASDs: fluorescence probes such as pyrene (Purohit 2015, ref 29; Purohit 2017, ref 25; Tian 2016, ref 30) and X-ray photoelectron spectroscopy (XPS) of spray dried particles (refs 9, 31, 32) C13.

The paper therefore sets two goals: to understand how a binary spray solvent containing water affects the miscibility of a model ASD made by spray drying, and to develop experimental and theoretical ways to follow the phase behavior A1·A5. The model pair is ritonavir (RTV) with copovidone (PVPVA, Kollidon VA 64). Ritonavir crystallizes slowly (Baird 2010, ref 16), so amorphous phase separation can be followed without crystallization getting in the way; PVPVA is common in marketed ASDs and dissolves readily in both methanol and water C15. Water is an antisolvent for ritonavir and methanol a solvent, which is the situation that should provoke separation as methanol evaporates C47. The release side of the story builds on Indulkar 2019 (ref 67): ritonavir–PVPVA dispersions release drug and polymer together (congruently) up to 25% drug loading, and above that release slows to roughly that of the neat amorphous drug R23·C63.

Background concepts

Phase separation during drying. A spray drying feed is a solution of drug and polymer in a solvent. As the droplet loses solvent, its composition moves along a path in composition space toward the dry drug–polymer mixture. If at some point along that path the drug concentration exceeds the drug's solubility in the local mixture, the droplet can split into a drug-rich and a drug-poor phase. The authors express the driving force as the ratio of the drug concentration to its solubility in the current droplet composition, $C/S$: below 1 the drug is dissolved, above 1 it is supersaturated and separation can start E14·R20. Whether separation actually happens and leaves a trace in the dry particle depends on how high $C/S$ goes and for how long, because drying takes well under a second C54·R22.

Four components, one tetrahedron. With methanol (S), water (W), ritonavir (D) and PVPVA (P) the composition space is a tetrahedron. The bottom face DPS holds the water-free mixtures, which are fully miscible; adding water lifts compositions toward the W apex C48. Because a tetrahedron is hard to read, the authors also cut pseudo-ternary sections in which water and methanol are lumped into one "total solvent" axis, at a fixed initial H₂O:MeOH ratio P12·R18.

Flory–Huggins free energy. The phase boundaries come from a Flory–Huggins free energy of mixing for four components (SI Eq. 12, after Horst and Wolf 1995), printed as:

In plain terms: the free energy of mixing has an entropic part, a sum over the components of their volume fraction ${\phi }_i$ divided by a relative molecular size $N$ times $\ln {\phi }_i$, and an interaction part, a sum over all pairs of components of ${\chi }_{ij}{\phi }_i{\phi }_j$, where ${\chi }_{ij}$ is the Flory–Huggins interaction parameter of the pair (the larger it is, the less readily the pair mixes) E11. As printed in the SI the two sums are not joined by an operator and $N$ carries no component index; the standard multicomponent form adds the two parts, and the card records this as a printing problem rather than a different model X15. Where the free energy along a composition line has two minima with a common tangent, two phases can coexist; these points map the binodal, the boundary of the two-phase region P12.

The six pairwise $\chi$ values (with 1 = methanol, 2 = water, 3 = ritonavir, 4 = PVPVA) were estimated in different ways. Four came from dynamic vapor sorption (DVS) fitted with the binary Flory–Huggins sorption isotherm given in an SI footnote:

In plain terms: the activity $a_{1,g}$ of a vapor (water or methanol) taken up by a solid depends on the volume fractions of sorbed vapor (${\varphi }_1$) and solid (${\varphi }_2$), on the ratio of their molar volumes $v_1/v_2$, and on their interaction parameter; fitting measured sorption isotherms gives $\chi$ E12. The values used are 0.8 for drug–methanol, 2.8 for drug–water, 0.73 for polymer–methanol and 1.13 for polymer–water; methanol–water was set at −0.002 from published heats of mixing R15.a–e. Note by the AI writer (not discussed in the paper): the drug–water value of 2.8 is by far the largest of the set, which is what makes water act as a strong antisolvent for ritonavir in the model. The drug–polymer value is the least certain: the main text says it "is likely <0.5, and possibly close to zero", while the SI says it was assumed to be 0.5; a sensitivity analysis from −0.5 to 0.5 moved mainly the drug–polymer side of the diagram R15.f·R16·C72·X10. The main text and SI also disagree on where the drug–methanol value came from (methanol solubility versus DVS), and the DVS and solubility data are not shown X11.

Reading the surface. Pyrene emits a five-band monomer spectrum between 370 and 405 nm; the ratio of the first to the third band (the "peak 1/3 ratio") is higher in a polar environment and lower in a non-polar one (Kalyanasundaram 1977, ref 66; Purohit 2015, ref 29) C25·R5. At 0.1 wt% of solids, pyrene is meant to sit dissolved in the ASD and report the polarity around it; crystalline pyrene adds a broad excimer band near 460 nm, which arises only when two pyrene molecules are close R5·C16. In a one-phase SDD the ratio should fall in proportion to the amount of lipophilic drug; if the drug separates and pyrene moves into the drug-rich phase, the ratio falls further C26·C28. XPS counts photoelectrons from the outer 0–10 nm of a particle; from the C 1s, O 1s and S 2p signals, compared with spectra of the pure components, the authors compute the drug mass fraction at the surface (ritonavir contains sulfur, PVPVA does not) P8·R24. The authors argue that solid-state fluorescence is also surface sensitive because UV light penetrates only a fraction of its wavelength into a solid C34.

Aging in DSC. A glass held below its $T_g$ relaxes toward a lower enthalpy; on reheating the recovered enthalpy shows up as a peak in the non-reversing heat flow of modulated DSC near the glass transition of that phase. The authors use this deliberately: they age 50% drug loading SDDs for 72 h at 40 °C, low enough, they judge, not to induce separation, so that a small hidden drug-rich phase grows its own relaxation peak and becomes visible; a low-temperature peak that shrinks on aging would instead be residual solvent P5·C22.

Surface normalized dissolution and congruent release. The surface normalized dissolution rate (SNDR) is measured on a compact of constant exposed area, so it does not depend on particle size or disintegration E13·C36. It is the slope of drug concentration against time divided by the exposed area and the drug loading E13. For copovidone ASDs at low drug loading, the dissolving polymer carries the drug with it (congruent release) and the rate is high; above a drug-specific loading the drug controls release and the rate falls to about that of the neat amorphous drug (Indulkar 2019, ref 67) C63·R23. For ritonavir–PVPVA the switch lies at 25% drug loading, so the 25% DL dispersions studied here sit at the edge where small changes in the particle could matter most R23.

Methods in detail

The study has three parts that run in parallel: making SDDs from solvents with different water contents, reading each powder in bulk, at the surface and in release, and simulating the drying droplet. The flow below shows how they connect.

Part 1: Making SDDs with controlled solvent water

What was done. Stock solutions of ritonavir and PVPVA at 25, 50 and 75 wt% drug loading (DL, relative to total solids) were made at 4% solids in methanol or in water–methanol mixtures (w/w), each with 0.1 wt% pyrene P1. They were sprayed on a BLD-35 dryer with a Schlick 2.0 nozzle at 175–195 °C inlet and 50 °C outlet temperature, 120 psi atomization pressure, 450 g/min drying air and 25 g/min feed (Table 1), then tray dried at 30 °C for at least 12 h and stored at 4 °C over drierite P1. The full matrix (25, 50 and 75% DL from methanol, 10:90 and 60:40) went to every analysis; 50% DL SDDs from 20:80 and 40:60 were added for DSC only (Table 2) P2. The 60:40 solution had already phase separated before spraying, with drug-rich droplets suspended in a polymer solution, so in that case a suspension was dried R13. Note by the AI writer (not discussed in the paper): the 60:40 arm therefore tests what happens to a feed that is already two-phase, while the 10:90 arm, whose feed was one phase, is the real test of separation during drying.

Part 2: Three views of the same powder

Bulk: PXRD and DSC. PXRD (Cu Kα, 5–35° 2θ) checked that the SDDs were amorphous P3. Modulated DSC (2–5 mg in nonhermetic pans, 0–150 °C at 2.5 °C/min with ±1.5 °C modulation every 60 s) gave the midpoint $T_g$; the nonhermetic pans let residual solvent escape on heating P4. Because the first runs of the 50% DL series showed broad, ambiguous transitions in which small bumps could be either solvent or the overshoot of a second glass transition, those samples were aged for 72 h at 40 °C and rerun P5·R2.

Surface: pyrene and XPS. Powders were sandwiched between quartz slides and excited at 332 nm; the peak 1/3 ratio was taken from the emission spectrum P6. A calibration line of ratio against DL for the one-phase SDDs (data of Figure 6) converted each measured ratio into an apparent surface ritonavir content (Figure 8) P7·E15. XPS used monochromatic Al Kα radiation, a charge neutralizer and at least four spots per sample; the C 1s and O 1s envelopes were fitted with the spectra of the pure components and converted into drug mass percent, with S 2p as an independent check P8. Note by the AI writer: the card recomputed the pyrene calibration from the plotted methanol points (slope about −0.0054 per % ritonavir, intercept about 2.164); the 60:40 25% DL ratio of about 1.87 then gives about 55% surface ritonavir, as plotted in Figure 8 E15.

Release: intrinsic dissolution and DLS. About 50 mg of SDD was compressed at 1500 psi for 2 min in an 8 mm die, backed with 50 mg ethyl cellulose, and rotated at 100 rpm in 100 mL of 50 mM sodium phosphate buffer at pH 6.5; samples were not filtered because the compacts did not disintegrate, and ritonavir was assayed by HPLC at 210 nm P9. The SNDR is the slope of concentration against time divided by the exposed area and the DL, from 6 to 9 time points, in triplicate; neat amorphous ritonavir made by rotary evaporation served as the control P9·E13. The dissolution temperature is not printed X13. Colloids that formed in the medium were sized by DLS, as were precipitates formed by adding water to a methanol solution of the solids P10.

Part 3: A droplet drying model coupled to a four-component phase diagram

The problem. Experiments show the dried particle; they do not show the path the droplet took in less than a second of drying. The authors built a model to follow the composition of a single drying droplet and to ask, at each moment, whether the drug is above its solubility P11·P12.

The drying model. A spherical droplet is divided into concentric shells, surrounded by a closed gas shell; mass and heat flow between shells as a set of ordinary differential equations, which the authors call "fit for purpose" rather than a full description of drying P11·C73. At the droplet surface mass and heat balance:

In plain terms: the volume of volatile liquid lost per unit time, times its density $\rho$, equals the mass of vapor $m_{vap}$ leaving the droplet (SI Eq. 1) E1. Each solvent evaporates at a rate given by a modified Maxwell model (SI Eq. 6):

In plain terms: the evaporation rate grows with the droplet radius $r$, the vapor diffusivity $D_v$, the formula weight $F{W}_1$ and the saturation vapor pressure $P_{sat,1}$ of the solvent, falls with the surface temperature, and goes to zero as the surrounding gas approaches saturation ($RS$ = relative saturation) E6. Vapor pressures come from Antoine coefficients of the pure solvents, so azeotropes are not treated E16·C73. Methanol, with its higher vapor pressure, therefore leaves faster than water, which is the source of the water build-up described below C49. Diffusion coefficients were fixed at $1.0\times {10}^{-5}$, $1.0\times {10}^{-6}$ and $1.0\times {10}^{-7}$ cm²/s for solvent, drug and polymer, internal transport was set fast enough that concentration gradients inside the droplet are largely absent, and the simulated droplet was 30 µm across, starting at 110 °C (140 °C for the 50:50 feed) P11. A 30 µm droplet dried in about 0.5–0.7 s and spent less than 1% of that time hot before settling near the wet-bulb temperature (below 10 °C), which the authors use to justify room-temperature $\chi$ values R22·C74.

The phase model. At each step the Flory–Huggins free energy is evaluated along the drug mass-fraction axis at the current composition; its minima give the drug solubility and tie lines along the path P12. Points of the path where the drug is above its solubility are drawn red, those below blue; the red segments of many paths together map the two-phase region P12·R17. The main text and SI describe the integration differently (Euler with 1 ms steps versus RK4 with ${10}^{-6}$ s steps, both over 1 s) X9. Note by the AI writer: the simulated conditions (a 30 µm droplet, a 110 °C start in a closed gas shell) are not the conditions of the dryer used for the experiments (inlet 175–195 °C, outlet 50 °C), so the model is best read as an explanation of trends rather than a prediction for this dryer X14.

Experimental analysis

Main result 1: the bulk looks almost clean. The methanol SDDs shown by PXRD were X-ray amorphous, and none of the SDDs showed a melting peak in DSC (Figure 2) R1·C18. In unaged modulated DSC (Figure 3) every methanol SDD had a single $T_g$, 91, 75 and 63 °C at 25, 50 and 75% DL, falling toward ritonavir's own $T_g$ of 51 °C as DL rises (PVPVA: 108 °C) R2·M1·M2·C19. From 10:90 the 25% and 50% DL SDDs also had single transitions at 92 and 75 °C; the 50% DL transition was slightly broader than its methanol counterpart, which the authors say "may indicate subtle inhomogeneity" R2.b·R2.d·C20. Only the 75% DL SDD from 10:90 showed more than one step, starting near 50 °C and with a second step near 90–97 °C (read from Figure 3, approximate), "suggesting the formation of multiple amorphous phases" R2.f·C21. Later, in the discussion, this result is attributed to the 75% DL system "prepared with 60% water in solvent (Figure 3)", but Figure 3 contains no 60:40 sample; the card reads this as a misprint for 10% water C62·X1.

Main result 2: aging reveals a second phase at 50% DL. After 72 h at 40 °C (Figure 4), the reversing heat flow of the 50% DL SDDs broadened as solvent water rose, from about 67–83 °C for methanol to about 50–90 °C for 20:80, and split into a small step near 49–55 °C plus the main step for 40:60 R3. The non-reversing signal is clearer: the methanol SDD has one relaxation peak near 70 °C, while all three water-containing SDDs show a shoulder near 50–55 °C beside the main peak near 68–69 °C R4. The authors take the two peaks as "confirming the presence of a second phase" and conclude that pre-existing phase separation "can be readily detected as the enthalpic relaxation peak grows" C23·C24. Note by the AI writer: the extra event lies close to the $T_g$ of neat ritonavir (51 °C), which fits a ritonavir-rich phase; the paper does not make this comparison explicitly.

Main result 3: the surface tells a clearer story than the bulk. The pyrene ratio of the methanol SDDs fell almost linearly with drug content, from about 2.15 at 0% to about 1.62 at 100% ritonavir (Figure 6) R6.a·C26. SDDs from 10:90 gave "nearly identical" ratios, but SDDs from 60:40 were lower at every loading (about 1.87, 1.78 and 1.65 at 25, 50 and 75% DL against about 2.05, 1.89 and 1.765 for methanol), and methanol SDDs deliberately phase separated by storage at 75% RH fell to a similar level R6.b–d·C27·C29. The authors read the drop as a less polar pyrene environment that "could be due to" drug-rich regions into which pyrene partitions C28. XPS gives numbers directly (Table A).

Table A. Ritonavir on the SDD surface by XPS (% by mass, mean ± SD over at least four spots), from Table S1 R7.

For methanol SDDs the surface matched the nominal composition (R² 0.9720 for O 1s and 0.9946 for C 1s), which the authors take as a homogeneous distribution R8·C30. The 60:40 SDDs showed "significant surface drug enrichment": 59–70% ritonavir at 25% DL and 85–91% at 50% DL by O 1s and C 1s C31·R7.g·R7.h. The authors give two reasons: the 60:40 feed was already a suspension of drug-rich droplets, whose composition is set by the solvent rather than the DL, and that heterogeneity "can be carried over" into the particle; and water, which boils later than methanol, stays behind the inward-moving drying front so that the poorly soluble drug "may precipitate, causing surface drug enrichment" C32·C33. For 10:90 the authors call the departures "slight" at 25% and 75% DL C31. Table A shows that at 25% DL the 10:90 surface carries about twice the ritonavir of the methanol surface (43 against 21% by C 1s, 46 against 25% by S 2p), a larger difference than "slight" suggests, and the abstract's statement that this SDD showed no phase heterogeneity in solid-state characterization does not hold for XPS X3. At 75% DL the 10:90 surface is drug-poorer than the methanol one X4. The 75% DL 60:40 SDD gives the least consistent numbers (74 ± 11, 82 ± 17 and 92 ± 2) X19.

Surface contents back-calculated from pyrene (Figure 8) agree with XPS for the methanol SDDs (about 21, 51 and 74% against 20–21, 45–50 and 74–82%) and the authors report "good agreement" overall R9.b–d·C35. For the water-containing SDDs the two methods differ by up to about 20 points: for 60:40 at 50% DL pyrene gives about 71.5% and XPS 85–91%, at 75% DL about 96% against 75–82%, so the two methods rank these two SDDs in opposite order R9.j·R9.k·X8. The discussion cites Figure 7 for this agreement; the comparison is in Figure 8 X7.

Main result 4: release falls at 25% DL only. Table B collects the dissolution rates.

Table B. Surface normalized dissolution rate (mg·min⁻¹·cm⁻¹, n = 3, pH 6.5) read from Figure 9 (approximate) R10.

At 25% DL the methanol SDD dissolved fastest; the authors write that 10% water "decreased the surface normalized dissolution rate by nearly half" and 60% water brought it "to about a quarter of the original value" C39. Read from Figure 9, the ratios are about 0.65 and 0.26, so the first is closer to a one-third loss than to a half R10.a·X5. Plotted against the XPS surface content (Figure 10A), the three 25% DL SDDs fall in order: about 3.5 at 21% surface ritonavir, 2.28 at 43% and 0.92 at 70% R11.a–c. At 50% and 75% DL the rates were hundreds of times lower, close to the neat amorphous drug, and not changed significantly by the solvent; the authors add that the dissolution conditions "may" not have been discriminative C37·C38·R10.b–d. In Figure 10B the labels "0%H₂O" and "10%H₂O" of the 50% DL points appear swapped when checked against Figure 9 and Table S1 X6. The authors summarize the 25% DL 10:90 case as changes that "may be subtle" but caused "remarkable changes in dissolution performance" C41.

Main result 5: colloids are a solution phenomenon. During dissolution all 25% DL SDDs, including the phase-separated ones, exceeded the amorphous solubility of ritonavir and formed drug-rich colloids of 390 ± 21 nm (methanol), 467 ± 24 nm (10:90) and 454 ± 23 nm (60:40, sampled at 130 instead of 60 min); none formed at 50% or 75% DL (Table 4) R12·R14. The precipitate formed by adding water to a 1.6% solids methanol solution to reach 60:40 measured 1626 ± 94 nm, and the authors expect larger particles in the 4% solids feed and growth rather than breakage of soft amorphous drug particles during spray drying R13·C44. They conclude that the colloids "Clearly" come from liquid–liquid phase separation as drug is released quickly, not from domains formed during drying, and that fast release with colloid formation survives even 60% water in the spray solvent C43·C45.

Main result 6: the model shows why small water differences matter. Without water, the drying path stays on the miscible DPS face C48. With water, the paths bend into parabolas toward the water apex because methanol leaves first; for a 15:85 feed the water in the droplet rises from 15% to about 26% at about 0.4 s before falling to zero, for 12.5:87.5 from 12.5% to about 16% (Figure 13B) C49·R21. Table C summarizes where the paths exceed the drug solubility.

Table C. Segments of the computed drying paths drawn red (drug above its solubility), read from Figures 11A and 12 R17·R19.

The authors place the threshold at 25% DL at "approximately 12.5/87.5 H₂O:MeOH", adding that it "may move higher or lower" because concentration gradients were not explicitly modelled and the $\chi$ values have limited accuracy C50·C51. Higher DL widens the overlap: 11.5:88.5 "completely passes through the miscible region at 25% drug loading" but "is largely phase separated at 50% drug loading", and in the pseudo-ternary section at 11.5:88.5 the 50% DL path runs along the edge of the two-phase region while the 25% DL path bypasses it (Figure 11B) C53·C55·R18.a. At 15:85 and 30:70 the two-phase region covers most of the section and both paths run through it R18.b·R18.c. Figure 13A shows the driving force: at 12.5:87.5 $C/S$ exceeds 1 only between about 0.28 and 0.42 s and peaks near 2.4, whereas at 15:85 it climbs to about 320 by 0.28 s and stays above 1 until about 0.6 s R20. The authors' reading is that "a small increase in initial water content, when near a phase boundary, can have strong effects on both the driving force and timescale available for phase separation" C61. They also explain the falling $C/S$ late in drying: water drives $C/S$ up while the rising polymer fraction drives it down, since "increased polymer content can also facilitate drug solubilization" C59.

Discussion

How the authors join the pieces. The chain they propose runs from the solvent to the release rate, and it splits according to how much water the feed holds and how much drug the dispersion carries C49·C58·C40·C67.

At 50% and 75% DL the same solvent puts more of the path into the two-phase region, so these dispersions are "more prone to phase separation during spray drying", but their release is already incongruent and close to that of the neat drug, so "the impact of phase separation on drug release rates may be minimal in these systems" C71·C66. Hence the authors' central practical claim: the effect of phase separation on dissolution, faster or slower, "is expected to be more profound in low drug loading systems, where the drug release is congruent or close to congruent to that of the polymer" C67. They place this within a broader view in which the population, composition and location of the separated phases decide the outcome: polymer-rich phases can speed release above the congruent limit if the dissolving front is not blocked (Li 2019, ref 15), while at or below the limit drug-rich phases leave less drug for fast release, and "if the dissolving front is covered by slow releasing drug-rich phases, a decline in drug release rate is also likely to occur" C64.

Which method sees what. The paper is in part a comparison of detection methods. For the 60:40 SDDs every method used for them shows a change: the pyrene ratio, three XPS signals and a roughly four-fold slower dissolution at 25% DL R6.c·R7.g·R7.h·R10.a. For the 10:90 SDDs the signals are weaker: unaged DSC shows two transitions only at 75% DL, aged DSC adds a second phase at 50% DL, pyrene sees nothing, XPS shows surface enrichment at 25% DL that the authors call slight, and dissolution at 25% DL falls by about a third R2.f·R4.b·R6.b·R7.d·R10.a. The authors conclude that "the release test showed the highest sensitivity to subtle differences in SDD homogeneity for the low drug loading systems", that DSC with carefully designed sample preparation also helps, and that although more powerful techniques "may be needed", "routine dissolution tests and DSC runs can serve as quality indicators" C68·C24·C69. They also propose solid-state pyrene fluorescence as a fast, low-cost way to estimate surface composition C35.

Model and experiment. The abstract says the "calculated phase diagrams and drying trajectories were consistent with experimental observations, suggesting that small variations in solvent composition may cause significant changes in ASD phase behavior during drying" C9. The agreement is in trend: more water gives more separation in both, and in the model higher DL brings separation at lower water, as in the experiments where only the 75% DL 10:90 SDD showed two transitions unaged C70·R19·R2.f. The model also explains why the 60:40 surfaces are so drug-rich: paths deep inside the two-phase region produce drug-rich phases of almost pure drug, while grazing paths give phases mixed with solvent or polymer C56. Note by the AI writer (not discussed in the paper): at the 10:90 composition itself the model and the experiments do not meet. The computed 10:90 path stays below the drug solubility at both 25% and 50% DL, yet aged DSC finds a second phase at 50% DL and dissolution drops at 25% DL; the authors' own caveat about concentration gradients and $\chi$ accuracy is the natural place to look, and the drug–polymer $\chi$, described as both "likely <0.5" and assumed 0.5, is the most uncertain input X12·C51·X10.

What it suggests for formulation. The authors' own recommendations are to consider water in spray solvents, whether added or picked up by hygroscopic solvents, and to use phase diagrams with drying trajectories to assess the risk of phase separation for mixed solvents and other solvent-based processes C46·C10. Note by the AI writer: for a low-loading SDD that relies on congruent release, the water content of the spray solvent behaves like a critical process parameter, and a constant-area dissolution test is a more sensitive release check than unaged DSC; a cloudy feed, as at 60:40, is a cheap warning sign that the particles will have drug-rich surfaces.

Limitations

Limitations stated by the authors. The model threshold "may move higher or lower depending on concentration gradients which were not explicitly considered in these simulations, and the limited accuracy of χ parameters" C51. The model is "fit for purpose": it does not treat azeotropes or convective flow, uses temperature- and concentration-independent diffusion coefficients and densities, and uses room-temperature $\chi$ values, which "likely" misrepresent the hot start of drying C73·C74. The lack of a solvent effect at 50% and 75% DL "may also be because the dissolution conditions used in this work were not discriminative" C38. More sensitive techniques for direct miscibility characterization "may be needed" C69.

The AI writer's own judgement (1): statements that the figures do not bear out. Several sentences in the text do not match the paper's own data. The 10% water effect at 25% DL is described as "nearly half", but Figure 9 gives a ratio of about 0.65 X5. The abstract says that 25% DL from 10:90 showed no phase heterogeneity in solid-state characterization, but its XPS surface content is about twice that of the methanol SDD X3. The abstract says that heterogeneity at 50% and 75% DL from 10:90 was seen only by DSC, but XPS also departs from the methanol SDD at 75% DL X4. The abstract says 60:40 SDDs showed heterogeneity "with all analytical methods used", but no DSC of a 60:40 SDD is shown, and PXRD is shown only for methanol SDDs X2·X18. The discussion assigns the two-$T_g$ result of Figure 3 to 60% water and the fluorescence–XPS comparison to Figure 7, and Figure 10B appears to swap two labels X1·X7·X6. None of these changes the main direction of the results, but each matters when the paper is used as a source of numbers.

The AI writer's own judgement (2): what the evidence can and cannot carry. The 60:40 result is robust but answers a narrower question than the title suggests: its feed was already two-phase, so it shows that a suspension feed yields drug-rich surfaces, not that drying itself split the drug R13. The 10:90 result is the true test of separation during drying, and there the evidence is one aged-DSC signal at 50% DL, one two-step DSC trace at 75% DL and one dissolution comparison at 25% DL with n = 3 R4.b·R2.f·R10.a. The mechanism that links surface enrichment to slower release, a drug-rich layer at the dissolving front and fewer low-DL domains, is plausible and consistent with the monotonic fall of SNDR with surface drug content, but neither the domain structure inside the particles nor the polymer release was measured C40·R11. The fluorescence–XPS agreement holds within about 20 percentage points, which is enough to rank strongly enriched from unenriched surfaces but not to replace XPS X8. Finally, the model is internally inconsistent in its description (${\chi }_{34}$, integration settings, printed SI equations) and was run under conditions that differ from the dryer used; it supports the trend argument, not a numerical threshold for this process X9·X10·X11·X14·X15.

Conclusion

The paper makes three contributions. First, it shows experimentally that water in the spray solvent changes ritonavir–PVPVA SDDs: from 60:40 H₂O:MeOH the particles have drug-rich surfaces (59–70% ritonavir at 25% DL, 85–91% at 50% DL by XPS) and the 25% DL dispersion dissolves about four times slower; from 10:90 the change is subtle but still slows the 25% DL dispersion by about a third and leaves a second phase visible in aged DSC at 50% DL R7·R10·R4·C31. Second, it compares detection methods and finds that the release test and DSC after aging are the most sensitive for low and intermediate loadings, with pyrene fluorescence and XPS giving similar surface compositions within about 20 points C68·C24·C35·X8. Third, it couples a droplet drying model with a four-component Flory–Huggins phase diagram and shows that, because methanol evaporates first, a small increase in initial water near a phase boundary raises the peak supersaturation from about 2.4 to about 320 and lengthens the time spent above the solubility R20·C61.

The authors' conclusion is that water addition "can lead to phase separation of amorphous solid dispersions during spray drying even for an initially one phase stock solution", that both experiment and model "suggested" more separation with more water, and that high-loading ASDs are more prone to separate while the effect on release "may be minimal" there C2·C70·C71. These findings are offered to aid the rational design of spray drying conditions and the risk assessment of solvent-based processes with mixed organic solvents C10. Within the tested system (one drug, one polymer, one dryer setting, one dissolution condition), the practical lesson is to control spray solvent water for low-loading dispersions and to check them with a release test rather than with DSC alone.