Seeing the first crystals in an amorphous drug: second harmonic generation microscopy as a crystal-only detector

Background and motivation

The solid form of an active pharmaceutical ingredient (API) decides much of how it behaves in a product. A drug can exist as polymorphs, hydrates, cocrystals, salts or an amorphous solid, and the free energy of the form sets its apparent solubility and chemical stability (Byrn 1999, ref 1). An amorphous solid has no long-range order, so it sits at a higher free energy than the crystal and typically dissolves faster, which makes it attractive for the many drug candidates with poor aqueous solubility. The same excess free energy drives crystallization, on time scales that vary between compounds and with temperature and relative humidity (refs 2–5, including Konno 2008 and Rumondor 2009). Polymeric crystallization inhibitors are the main formulation answer (refs 6–9), and the introduction states that crystallization must be detected early and, where possible, prevented over the life of the product.

The authors argue that the bottleneck is detection. A stability study of an amorphous material has to run until crystals can be definitively detected, so its length is set by the analytical method. Thermal, optical, X-ray diffraction and spectroscopic techniques are the routine tools (refs 10–18), and the paper calls crystallinity below 1% a challenging limit for them, especially in formulated products. There is a further consequence. The lag before crystals appear is usually ascribed to nucleation, but the authors write that "it is clear" its length also depends on how sensitive the method is to the fraction or size of the crystallites. A lag time is therefore partly a property of the instrument.

Each kinetic method gives part of the picture (refs 3, 19–22). Isothermal calorimetry follows the heat of crystallization and so the overall rate, but cannot separate nucleation from growth. X-ray powder diffraction (XRPD) and spectroscopy measure the lag time and the overall rate but are sensitive only to about 1% crystallinity (Taylor and Zografi 1998, ref 23; Forster 2001, ref 24). Conventional optical microscopy (OM) follows nucleation and growth of single crystals and is "a reasonable compromise between sensitivity and ease of use". But a crystal must reach several micrometres before it can be told apart, for example by birefringence, from non-crystalline particles and other scattering centres. OM therefore "can be difficult" to use quantitatively for early growth, and it works only in transparent samples.

The tool the paper brings in comes from nonlinear optics. In the intense field at the focus of a pulsed laser, second harmonic generation (SHG) converts light into light of twice the frequency, here from 800 nm to 400 nm. By symmetry, SHG is allowed only from structures that lack an inversion centre: surfaces, interfaces and all homochiral crystals. Liquids, glasses and other amorphous materials give no coherent SHG, so an amorphous drug gives almost no background. SHG had been used for surface characterization, for crystalline pharmaceutical powders (Strachan 2004, ref 27, with refs 26 and 28) and for deep tissue imaging (refs 29–33). The Simpson group had introduced an imaging method built on it, second order nonlinear optical imaging of chiral crystals (SONICC; Wampler 2008, ref 25). Because significant SHG arises only in the focal volume, SONICC shares the penetration depth of multiphoton fluorescence in turbid media.

The paper explores SONICC "as a possible sensitive and selective alternative" for detecting crystallization in model amorphous systems. Its two drugs differ in a way that matters for SHG: griseofulvin is intrinsically chiral, and chlorpropamide is not chiral and has both SHG-active and SHG-inactive polymorphs. The work compares the time of first detection by SONICC and OM in the same field, estimates the smallest crystal SONICC can detect, and, as the abstract adds, examines the crystallization kinetics of amorphous griseofulvin. No polymer or formulated product was studied; polymer-stabilized amorphous formulations are the motivation, not the test system.

Background concepts

Glass, supercooled liquid and glass transition. Below its glass transition temperature ($T_g$) an amorphous solid is a glass, in which molecular motion is very slow. Above $T_g$ it is a supercooled liquid: still below its melting point, but mobile enough to rearrange into a crystal. Griseofulvin has a $T_g$ of 88.9 °C (from Zhou 2008, ref 22) and was held at 120 °C for the detection runs and at 125 °C for the main kinetic run; chlorpropamide has a $T_g$ of 15.1 °C (Wildfong 2005, ref 34) and was held at 30 °C (Section "Monitoring API Crystallization"). With both a driving force toward the crystal and enough mobility, crystallization is "expected to be rapid", so each experiment takes minutes to hours.

Nucleation, growth, lag time and seeds. Crystallization from an amorphous phase has two stages. Nucleation creates crystalline clusters stable enough to grow; growth then enlarges them until they meet. A nucleation rate counts new crystals per unit volume and time; a growth rate is a change of size per unit time. An instrument registers crystallization only after crystals pass its size or fraction threshold, so an apparent lag time contains both the time to nucleate and the time to grow to a detectable size. Seeds, crystals present at the start, skip nucleation. The paper cites Bhugra 2008 (refs 36, 37) for the point that preparation conditions strongly affect crystallization rates and that seeds are "one of the major factors", especially early.

Why chirality matters for SHG. SHG is coherent: the laser field drives a nonlinear polarization that radiates at twice the frequency. In a structure with an inversion centre each contribution is cancelled by its inverted partner, and in an isotropic glass the random orientations cancel in the same way, so no net second harmonic leaves the sample. Crystals whose unit cells are not superimposable on their mirror image, which the paper calls homochiral crystals, have no inversion centre and give SHG. Griseofulvin, named in the paper as the (2S,6′R) stereoisomer, "strongly favors" an SHG-active space group because of its intrinsic chirality (Section "Chlorpropamide"). Chlorpropamide is not intrinsically chiral and can crystallize either way: of its four polymorphs known under common conditions (Drebushchak 2007, ref 41, with refs 42 and 43), two are SHG-active and two are not (Table A). In α and γ the chirality is conformational, arising from the orientation of the molecules in the unit cell.

Table A. The four chlorpropamide polymorphs named in the paper, with their space groups and SHG activity (from the text of Section "Chlorpropamide", citing refs 41–43). The inversion-centre column restates the SHG column in symmetry terms.

Forward and backward SHG and the coherence length. The 400 nm waves generated at different depths in a crystal stay in step only over a limited distance, the coherence length, because the fundamental and the second harmonic travel at different speeds. For SHG leaving forward (with the beam, collected in transmission), the phase slip comes only from the small index difference between 800 and 400 nm. For SHG leaving backward (collected back through the objective, the epi direction), the generated wave runs against the driving wave and the slip is large. The paper gives the two coherence lengths as eqs 1 and 2, from LaComb 2008 (ref 38):

Here $L_C^f$ and $L_C^b$ are the forward and backward coherence lengths, and ${𝐤}^{\omega }$ and ${𝐤}^{2\omega }$ are the wave vectors of the fundamental (frequency $\omega$, 800 nm) and of the second harmonic (frequency $2\omega$, 400 nm), each of magnitude $2\pi n/\lambda$, with $n$ the refractive index and $\lambda$ the wavelength. In plain terms: the forward length is set by the difference of the two wave vectors, which is small, so it is long; the backward length is set by their sum, which is large, so it is short. The card's check reduces them to $L_C^f=400\mathrm{n}\mathrm{m}/(n_{2\omega }-n_{\omega })$ and $L_C^b=400\mathrm{n}\mathrm{m}/(n_{2\omega }+n_{\omega })$. With griseofulvin's index of 1.65 (refs 39, 40) and an assumed 5% index difference between 800 and 400 nm, the authors obtain $L_C^f$ = 4.8 µm and $L_C^b$ = 120 nm; for chlorpropamide (index 1.51, estimated by the Vogel method) they obtain 5.3 µm and 136 nm. The backward length is about 40 times shorter, and this is what later turns the forward-to-backward ratio into a size gauge.

Overall crystallization kinetics. The paper describes the rise of the crystallized fraction with the Johnson–Mehl–Avrami (JMA) equation (refs 22, 47–49), printed unnumbered in Section "Application of SONICC to Study the Kinetics of Crystallization":

Here $\alpha$ is the fraction crystallized, $t$ is time, $t_0$ is an induction time, $k$ is an apparent rate constant (min⁻¹), and $n$ is an exponent that the paper calls the order of reaction "associated with the mechanism of nucleation". In plain terms: nothing crystallizes before $t_0$; afterwards the fraction rises slowly, then quickly, then levels off at 1, an S-shaped curve whose time scale is $1/k$ and whose steepness is set by $n$. The authors fix $n$ at 2 throughout.

Methods in detail

The study has three contributions. The first compares SHG and optical images of the same field during isothermal crystallization of melt-quenched films, to test whether SHG detects crystals earlier and sees crystals that OM misses. The second uses the coherence of SHG, through the forward-to-backward ratio, as an internal length scale for the smallest detectable crystal. The third turns the same image stacks into nucleation rates, growth rates, JMA kinetics and an activation energy for griseofulvin, and compares the activation energy with a DSC value from the literature.

Contribution 1: Watching the same field with SHG and with optical microscopy

① Problem. To show that SHG detects crystals earlier, the comparison must be made on the same crystals at the same time; otherwise differences between samples would be confused with differences in sensitivity. The starting film must also pass the usual test for an amorphous sample, so that any early SHG can be read as crystals that formed later or that the usual test missed.

② What they did. About 1–3 mg of griseofulvin (Hawkins Inc.) or chlorpropamide (Sigma-Aldrich) was melted onto a precleaned cover glass, held molten for 2–4 min and quenched on a cold surface (Section "Amorphous API Preparation"). The melt temperature, cold-surface temperature and film thickness are not printed. A nitrogen purge excluded moisture throughout. A film was accepted as amorphous when polarized light microscopy showed no birefringence. Films were then held at 120 °C (griseofulvin) or 30 °C (chlorpropamide); the heating stage is not described. The custom SONICC system scanned the beam with a resonant mirror (about 8 kHz) and a galvanometer. The source gave 100 fs pulses at 80 MHz and 800 nm with 40 mW average power (about 0.5 nJ per pulse), focused by a 10× objective of numerical aperture (N.A.) 0.3. The measured depth of field was about 60 µm and the spot about 2 µm; the field of view (FOV) was 600 × 600 µm. A frame took 45 s for griseofulvin and 25 s for the faster chlorpropamide. SHG at 400 nm was filtered from the laser and photon-counted both in transmission (forward, F) and in epi (backward, B); the two channels were cross-calibrated with two-photon fluorescence from standard slides, and SHG was normalized to it. Bright-field OM images of the same field were recorded at the same time with a CCD camera. Displayed SHG images were thresholded at a signal-to-background ratio (S/B) of 3.

The authors argue that imaging did not change crystallization. Both drugs are transparent at 800 and 400 nm, so absorption heating "can be assumed" negligible; µJ ultrafast pulses are known to influence nucleation (Nakamura 2007, ref 35), but "no such effects are expected" at about 0.5 nJ; and fast scanning spreads any local heating. No laser-off control or unimaged region was compared (card note X12). The number of films per condition is not printed; Figure 1 shows two griseofulvin films and Figure 2 one chlorpropamide film.

③ Why it works. The gain comes from the contrast mechanism. In OM a small crystal in a glass looks much like a dust particle or bubble until it is large enough for its birefringence to identify it. In SONICC the glass gives no coherent SHG, so the background is set by detector noise, and any pixel above threshold contains ordered matter without an inversion centre. A crystal far below the 2 µm resolution can still be detected, because detection needs only enough signal, not a resolved image. Earlier detection follows directly, provided the crystal is SHG-active.

④ How it differs from earlier work. XRPD, Raman spectroscopy and DSC average over the whole sample and, by the paper's account, need 0.1–10% crystallinity. OM resolves single crystals but needs several micrometres and a transparent sample. Earlier pharmaceutical SHG work (Strachan 2004) analyzed crystalline powders. Here SHG imaging targets the onset of crystallization in an amorphous drug, with OM recorded in the same field as a built-in comparator.

Contribution 2: The forward-to-backward ratio as a ruler for the detection limit

① Problem. A detection limit normally needs a crystal of known size or a sample of known crystallinity. The first crystals SONICC sees are below the optical resolution, and the paper had no size standard. The authors therefore looked for a length scale built into the signal itself (Section "Estimated Detection Limit of SONICC").

② What they did. They used the coherence lengths as that scale. Within the slowly varying field approximation, which they state is valid for low-N.A. objectives (LaComb 2008), they simulated F and B as a function of crystal length (Figure 3c). They then followed F and B over time in single pixels (Figure 3a for griseofulvin, Figure 3b for chlorpropamide). For griseofulvin, F/B stayed close to 1 until 20 min, after which forward SHG was favoured. At that moment the crystal is taken to be about $L_C^b$ thick, "no greater than" about 120 nm, and the measured S/B was 30. Back-calculating to S/B 3, and "assuming for simplicity" that the crystal is equally thick in all directions, they estimated the smallest detectable domain as 90 nm, which they call an upper limit on the detection limit. For chlorpropamide, with $L_C^b$ = 136 nm and S/B 5 at the transition, the estimate is 125 nm. The relation used to scale S/B down to 3 is not printed, so neither size can be recomputed (card note X3). The size was converted into a percent crystallinity with eq 3:

Here $V_{\mathrm{s}\mathrm{m}\mathrm{a}\mathrm{l}\mathrm{l}\mathrm{e}\mathrm{s}\mathrm{t}\mathrm{c}\mathrm{r}\mathrm{y}\mathrm{s}\mathrm{t}\mathrm{a}\mathrm{l}}$ is the volume of the smallest detectable crystal, and $V_{\mathrm{F}\mathrm{O}\mathrm{V}}$ is the probed volume, 600 × 600 × 60 µm³, from the field size and the measured 60 µm depth of field. In plain terms: the detection limit as a volume fraction is the volume of one just-detectable crystal divided by the volume the microscope samples in one frame. This gives about 3.4 × 10⁻⁹ % by volume for griseofulvin and 9 × 10⁻⁹ % for chlorpropamide, "one part in 10 billion by volume". The card reproduces the griseofulvin figure: (0.090 µm)³ divided by 2.16 × 10⁷ µm³, times 100, is 3.4 × 10⁻⁹ %.

③ Why it works. While the crystal is much thinner than $L_C^b$, waves from all depths are nearly in phase in both directions, so F/B is close to 1, which the authors describe as "predicted and observed". Once the crystal grows past $L_C^b$, backward waves from different depths begin to cancel and B oscillates between zero and a low maximum, while F, in phase over the much longer $L_C^f$, keeps rising (Figure 3c). The time at which F/B leaves 1 therefore marks a known thickness, about 120 nm, far below the 2 µm spot. The authors add that F/B does not measure nucleation, since the critical nucleus has no physical relation to $L_C^b$, but it "does provide a calibration point" for size below the diffraction limit. They state that the uncertainty in the index mismatch affects $L_C^b$ "only minimally", which follows from eq 2: the backward length depends on the sum of the indices, not their small difference.

④ How it differs from earlier work. Crystallinity detection limits are commonly set with physical mixtures of known content (general practice, not discussed in the paper). Here the calibration is internal: a length fixed by the refractive index, applied to the crystal under one pixel. Because most APIs have refractive indices between 1.4 and 1.8, the authors expect submicrometre detection "routinely" in other APIs (Section "Chlorpropamide"); no other API was measured.

Contribution 3: Nucleation, growth and overall kinetics from one image stack

① Problem. A full kinetic description needs the nucleation rate, the growth rate and the overall conversion. Calorimetry gives only the last, and OM gives the first two only once crystals reach several micrometres. The question is whether background-free SHG images can supply all three from one sample.

② What they did. Amorphous griseofulvin was held at 125 °C and imaged every 45 s (Figures 4 and 5). ImageJ particle analysis found each contiguous region above 3 times the noise and recorded its area and mean SHG intensity in every frame. Plotting each particle's area against time gives a growth map (Figure 4a): each trace is one crystallite, shaded by its mean SHG intensity, which for dendritic crystallites the authors take as proportional to growth in the vertical direction. When two crystals touch, the algorithm counts one larger crystal and the trace jumps (Figure 4b, at 138 min). Growth rates came from the slope of the square root of area against time in windows without coalescence (Figure 5a). For the nucleation rate (Figure 5b), each particle was assumed to come from one nucleation site, counts per area were divided by the 60 µm depth of field, and any fall in the count was attributed to coalescence and corrected; the steady-state rate is the slope of a linear fit. The SHG integrated over the field, relative to its value at the end of the experiment, gave the fraction crystallized α (Figure 5c), fitted with the JMA equation at $n$ = 2. Repeating this at five temperatures (not printed; about 115, 120, 125, 128 and 130 °C read from Figure 6) gave an Arrhenius plot:

Here $k$ is the JMA rate constant, $T$ the absolute temperature, $\Delta E$ the activation energy and $R$ the gas constant. In plain terms: on a plot of ln k against 1/T the points fall on a line whose slope, multiplied by −R, is the activation energy. The fitted slope in Figure 6 is (−1.7 ± 0.2) × 10⁴ K, which the paper converts to 140 ± 20 kJ/mol (one standard deviation of the slope); the card's check with R = 8.314 J mol⁻¹ K⁻¹ gives 141 kJ/mol.

③ Why it works. Because the amorphous matrix gives no SHG, a simple threshold separates crystals from background, with no shape model or background subtraction. The low magnification gives a large field, so many crystallites are followed at once. The coalescence correction would miss a nucleus appearing in the same frame as a merger, which the authors judge "unlikely" because 45 s frames are short compared with the interval between nucleation events. Integrating the SHG over the field turns the same images into an ensemble measure of conversion. In the authors' words, quantitative real-time monitoring is "readily achievable" because of the low background and high contrast.

④ How it differs from earlier work. Zhou 2008 (ref 22) measured isothermal crystallization of amorphous griseofulvin by DSC and reported an activation energy of 167 kJ/mol. SONICC gives the same kind of overall rate and, from the same images, the growth of individual crystallites and the nucleation rate. The activation-energy comparison is the paper's check that the integrated SHG measures the same ensemble process as DSC.

Experimental analysis

Experimental setup. All experiments used melt-quenched films of pure drug on cover glass under nitrogen. Griseofulvin was studied at 120 °C for first detection (Figure 1), at 125 °C for the kinetics (Figures 4 and 5) and at five temperatures for the Arrhenius plot (Figure 6); chlorpropamide at 30 °C (Figures 2 and 3b). The temperature of the griseofulvin pixel trace in Figure 3a is not stated (card note X5). Each figure shows one film or one run; replicates are not reported apart from error bars on two points in Figure 6.

Table B. First crystal detection in the same field by SONICC and by OM (times from the text of Section "Detection of Nascent Nanoscopic Crystals by SONICC"; domain counts and later-frame descriptions read from Figures 1 and 2, approximate).

Main result 1: SONICC sees crystals first, and sees crystals that OM misses. In griseofulvin film (a), both channels were blank at 0 min, so the glass and substrate give no SHG above threshold. One SHG-active domain appeared at 7.5 min (circled in Figure 1a), while OM showed nothing until after 39 min. The same order held for crystals that started later; the authors write that SONICC was "always providing much earlier detection". Chlorpropamide crystallizes within about 30 min, and the 25 s frames show SHG from several centres after 125 s against first OM crystals after 500 s. These comparisons are direct, because both channels watched the same field at the same time, and they depend on no model.

Film (b) matters most for formulation practice. Prepared and monitored under comparable conditions, it showed no crystals in OM at 0 min, but SONICC showed several submicrometre crystals scattered across the field. By 12 min the SHG field was densely covered and OM showed crystal texture, whereas film (a) still held only about 10 SHG domains at 75 min. The authors write that the images "suggest the presence of residual nanocrystalline material" in the melt, formed "presumably" through incomplete melting or entrainment of nuclei during cooling (Conclusions). They conclude that SONICC "provides a selective way to verify" that a preparation has destroyed all crystalline structure, and that quenched-in nanocrystals "can profoundly influence" kinetics when crystallization is slow with long lag times. The Conclusions also state, without a hedge, that crystallization with seeds was faster. That rests on one film of each kind, imaged over different spans (12 versus 75 min), with no rate constant or crystal count compared (card note X10).

Note by the AI writer (not discussed in the paper): the SONICC–OM gap for griseofulvin is larger than growth alone would explain. At 125 °C single crystallites grew at about 1.2–1.5 µm per minute (Table D). Even at a third of that rate at 120 °C, a crystal would need roughly 10–13 min to grow from about 0.1 µm to the 5 µm at which, according to the paper, OM can reliably identify a crystal, yet the gap was more than 30 min. With a 10× objective and a low-contrast crystal in a glass film, the practical OM threshold was probably well above 5 µm, or early growth was slower than later growth.

Table C. The chain from refractive index to crystallinity detection limit (values from the text of Section "Estimated Detection Limit of SONICC"; F/B times and values read from Figures 3a and 3b, approximate). The Conclusions round both size limits to "<150 nm".

Main result 2: the F/B switch and the detection limit. Both drugs show the pattern the simulation predicts: F and B about equal at first, then forward SHG pulling ahead (Table C). For comparison the authors state that conventional methods (Raman spectroscopy, DSC, XRD) routinely detect 0.1–10% crystallinity, and that in pristine samples OM can definitively tell a crystal from a scattering centre only above 5 µm, more than 6 × 10⁻⁴ % crystallinity. They report an "approximate 2.9 × 10⁷ to 2.9 × 10⁸-fold" gain over conventional methods and at least 5 orders of magnitude over OM; the abstract gives 8 and 5 orders.

Two of these numbers need care. The printed upper bound of the fold gain does not follow from the stated range: 0.1% divided by 3.4 × 10⁻⁹ % gives the printed 2.9 × 10⁷, but 10% gives 2.9 × 10⁹, not 2.9 × 10⁸; the printed upper value, and the "8 orders" headline, correspond to a 1% reference (card note X1). And the 90 and 125 nm sizes cannot be recomputed, because the S/B scaling is not printed (card note X3). Note by the AI writer (not discussed in the paper): for a crystal much smaller than the coherence length, SHG intensity is usually expected to scale with the square of crystal volume, the sixth power of size. That relation reproduces the chlorpropamide value (136 nm × (3/5)^1/6 ≈ 125 nm) but gives about 82 nm, not 90 nm, for griseofulvin (120 nm × (3/30)^1/6); the paper does not say which relation it used. The Conclusions call the limit "measured to be approximately" one part in 10 billion, while the Results call the limits "approximate" and say they "collectively suggest" a vastly lower limit than routine methods. The Results wording fits the evidence: these are model estimates, not measurements on crystals of known size (card note X2).

Chlorpropamide does not grow steadily in SHG. Unlike griseofulvin, chlorpropamide's F and B levelled off or even fell before rising again. In the pixel of Figure 3b, both sat at about 50 counts from about 6 to 17 min, and B fell to about 20–30 counts at about 18–20 min before both rose; at about 29 min F was about 850 and B about 170 counts (read from Figure 3b). The authors write that this "cannot be easily explained" by growth of a single polymorph and is "tentatively attributed" to conversion between polymorphic forms (Table A). No structural method identified the growing crystals.

Table D. Kinetic quantities for amorphous griseofulvin at 125 °C from one SHG image stack (values from the text of Section "Application of SONICC to Study the Kinetics of Crystallization"; uncertainties are standard errors of the fits; quantities marked "about" read from Figures 4 and 5, approximate).

Main result 3: one image stack gives three kinds of kinetics. Table D shows what one run provides: growth of individual crystallites, a nucleation rate from corrected counts, and an S-shaped α(t) curve with its JMA constant. The merged region b grew more slowly than crystallite a alone, which "may be described by" the loss of total surface on which growth occurs; differences between separate crystallites (a and c) "may arise from" intrinsic variability and crystallite orientation. The text cites Figure 5b for the JMA fit, but the α data are in Figure 5c (card note X7). The uncertainties are fit errors, not run-to-run errors. The steps in Figure 5b are about 46 mm⁻³, one crystallite in the 0.0216 mm³ probed volume, so the nucleation rate rests on about 13 crystallites, and counting statistics alone give about 28% uncertainty, far above the ±1% fit error (card note X9). Note by the AI writer (not discussed in the paper): the read α curve is consistent with the text k. With $n$ = 2 and k = 3.35 × 10⁻³ min⁻¹, α reaches 0.5 about 250 min after $t_0$; with α ≈ 0.5 at about 290 min this implies $t_0$ of about 40 min, in line with α staying near 0 until about 50 min.

Table E. Arrhenius points for the JMA rate constant of griseofulvin (read from Figure 6, approximate; temperatures converted from the 1/T axis by the card's extractor, since the paper does not print them). Fitted slope (−1.7 ± 0.2) × 10⁴ K.

Main result 4: the activation energy is close to the DSC value. Across the five temperatures, ln k rises from about −6.5 to about −5.0, and the slope gives 140 ± 20 kJ/mol (Figure 6). Zhou 2008 reported 167 kJ/mol by DSC, and the authors write that the comparable values "suggests both measurements are reliably probing similar ensemble-averaged phenomena", while SONICC adds information on individual crystallites. The DSC value lies about 1.4 standard deviations from the SONICC value, and the temperature ranges and kinetic models of the two studies are not compared, so the agreement is approximate (card note X14). There is also an internal mismatch: the Figure 6 point at about 125 °C reads ln k ≈ −5.4 (k ≈ 4.5 × 10⁻³ min⁻¹), whereas the text gives 3.35 × 10⁻³ min⁻¹ (ln k = −5.70) at 125 °C; the point carries an error bar and may be a mean of several runs, but this is not stated (card note X6).

Comparisons and controls. Both drugs give earlier detection and the F/B switch, the parts of the method that need only an SHG-active crystal; their size limits are similar because their refractive indices are similar. They differ where polymorphism enters: griseofulvin's signal rises steadily, which makes the particle analysis and integrated kinetics possible, whereas chlorpropamide's signal is non-monotonic and no kinetics were attempted for it. The simultaneous OM channel is the main control, ruling out sample-to-sample variation as the reason for earlier SHG detection; the blank frame of film (a) at 0 min rules out SHG from glass or substrate; the literature DSC value checks the integrated kinetics. There was no size standard, no laser-off control, no structural identification of the chlorpropamide forms and no replicate films.

Discussion

Where the approach applies. The crystal must lack an inversion centre, so SONICC applies to intrinsically chiral drugs, and to achiral drugs only in a polymorph with a chiral unit cell; the authors state this limit in the Conclusions. The tested samples were thin films of pure drug on glass under nitrogen, with nothing else that might give SHG. The crystallinity limits also depend on the probed depth: the 60 µm depth of field enters eq 3 directly, but the film thickness is not printed, so the limits hold only if the film was at least that thick (card note X4). Within these conditions the earlier detection is robust, the seed observation is a single but clear image pair, and the kinetics are a demonstration with one run per temperature.

What it suggests for formulation. The authors propose rapid assessment of crystallization tendency, efficient screening of crystallization inhibitors, and checking that a sample meant to be amorphous carries no quenched-in nanocrystals that would reduce its physical stability. None of these was tested with a polymer. The seed observation has a direct message: the usual acceptance test for a melt-quenched amorphous sample, no birefringence, did not exclude submicrometre crystals in film (b). Note by the AI writer (not discussed in the paper): stability studies that rank polymers or formulations by onset time may partly measure differences in residual seeds rather than in nucleation tendency. Onset times from XRPD or DSC, near the 1% limit, also include a long period of growth, so lag times from methods of different sensitivity are not directly comparable, which is the point the introduction makes.

Note by the AI writer (not discussed in the paper): in an amorphous solid dispersion the polymer is itself amorphous and should add no coherent SHG, so the approach seems compatible in principle with polymer films, and the authors' expectation that SONICC "should be easily extendable" to powders and other scattering media would, if confirmed, remove the main weakness of OM. Two cautions follow from this paper's own results. A sample with no SHG is not proven fully amorphous, because SHG-inactive polymorphs and crystals below roughly 100 nm stay invisible (card note X11). And for an achiral drug, a falling SHG signal may mean a change of polymorph rather than a loss of crystals, so SHG alone cannot follow the crystallinity of such a drug quantitatively.

What remains unresolved. The origin of the seeds in film (b) was not tested, and the melt temperature that would bear on it is not printed. The chlorpropamide polymorph conversion needs a structural method. The paper does not discuss whether integrated SHG is proportional to crystallized volume. Note by the AI writer (not discussed in the paper): Figure 3 shows that the SHG of a crystal depends on its thickness relative to the coherence lengths, and the authors suggest that orientation "may" affect growth rates, so taking integrated SHG relative to its final value as the fraction crystallized is an assumption that the DSC comparison supports only indirectly.

Further work. The gaps point to a sequence of experiments: image crystals of known size, or mixtures of known crystalline content, to test the limit; run laser-off and unimaged-region controls; repeat the seeded and unseeded comparison on several films with rate constants; identify the chlorpropamide forms by Raman spectroscopy or XRPD while recording SHG; fit $n$ freely in the JMA analysis; and then move to polymer films, amorphous solid dispersions and powders, where the formulation questions behind the paper arise.

Limitations

Limitations stated by the authors. The Conclusions name the scope limit: SONICC is limited to molecules that are "either intrinsically chiral or adopt a polymorph with a chiral unit cell". The size estimates rest on "assuming for simplicity that the crystal thickness is comparable in all dimensions"; the griseofulvin figure is "an upper limit", and the limits are "approximate". F/B "does not directly provide a measurement of crystal nucleation". The coalescence correction will undercount nucleation if a merger and a new crystal fall in the same frame, which the authors judge "unlikely". The chlorpropamide polymorph explanation is "tentatively attributed", and growth-rate differences "may arise from" variability and orientation. Non-perturbation by the laser rests on argument ("can be assumed", "no such effects are expected").

The AI writer's own judgement (1): the detection limit is an estimate, not a measurement. The headline of about one part in 10 billion by volume is the paper's most quoted result and its weakest. It is back-calculated from one S/B value per drug, at one pixel where F/B leaves 1, through an unprinted scaling, with an assumed equal-sided crystal and a 60 µm probed depth for a film of unstated thickness (card notes X2, X3, X4); no crystal of known size was imaged. The coherence lengths carry about 5% uncertainty, because the sign of the assumed index mismatch is not stated: chlorpropamide's 136 nm is reproduced only if the index at 400 nm is 5% below that at 800 nm, and the opposite sign gives 129 nm (card note X15). For a material transparent at both wavelengths the index is normally higher at the shorter wavelength, which would favour 129 nm (the AI writer's background knowledge). The inputs of Figure 3c are not given, and its backward-SHG zeros, at about 0.04 and 0.08 $L_C^f$ (read), do not sit at $L_C^b/L_C^f$ ≈ 0.025 as eq 2 implies for griseofulvin, so the figure supports the F/B argument only qualitatively (card note X16). The threshold is called both signal-to-background and signal-to-noise without defining either (card note X17). None of this weakens the direct result that SONICC sees crystals earlier than OM; the 90–125 nm and 10⁻⁹ % figures should be read as order-of-magnitude estimates.

The AI writer's own judgement (2): the kinetics are a single-run demonstration. Each kinetic number comes from one image stack with fit-only uncertainties; the nucleation rate is uncertain by about ±30%, not ±1% (card note X9). The JMA exponent was fixed at 2 with no fit of $n$ or goodness of fit. The card's extractor notes that the paper's own data show constant nucleation and linear growth, for which classical JMA theory gives $n$ = 3 (two-dimensional) or 4 (three-dimensional) (card note X8), so k and the activation energy are comparable only with fits that also fix $n$ at 2. The Arrhenius temperatures are not printed and one point disagrees with the text k (card notes X5, X6). Whether the kinetic samples held seeds is not stated, so k and the activation energy may mix seeded and seed-free behaviour (card note X13). The DSC agreement is approximate (card note X14).

The AI writer's own judgement (3): seeds and perturbation. That a film passing the birefringence test held submicrometre crystals is direct and clear for film (b). That seeded crystallization was faster rests on one film of each kind with no rate constant compared (card note X10), so it agrees with, rather than establishes, the literature view that seeds accelerate crystallization. Non-perturbation by the laser is untested (card note X12). The card's overall rating is moderate: the earlier detection is direct and holds for two drugs, but the detection limits are model estimates and the kinetics come from single runs.

Conclusion

The paper shows that SHG microscopy can act as a crystal-only detector for amorphous drugs. Because a glass gives no coherent SHG, SONICC detected the first crystals in melt-quenched griseofulvin at 120 °C after 7.5 min, against more than 39 min by optical microscopy in the same field, and in chlorpropamide at 30 °C after 125 s against 500 s. It revealed submicrometre crystals at time zero in a griseofulvin film that passed the birefringence test, which the authors read as residual seeds. Using the short backward coherence length as an internal ruler, they estimated the smallest detectable crystal at 90–125 nm and the crystallinity limit near 10⁻⁹ % by volume. One image stack at 125 °C gave growth rates, a nucleation rate and a JMA rate constant, and five temperatures gave an activation energy of 140 ± 20 kJ/mol, near the 167 kJ/mol reported by DSC.

For a formulation scientist the robust lessons are qualitative. Absence of birefringence does not prove a sample free of crystals, and a crystallization lag time reflects the method's sensitivity as well as the material. The quantitative claims are weaker than their wording: the detection limit is a model estimate, the kinetics come from single runs, and nothing was tested with a polymer. SONICC is limited to chiral drugs or chiral polymorphs, and for achiral, polymorphic drugs its signal needs a structural method beside it.