## Main

Potato (Solanum tuberosum L.) crisps are manufactured either by deep-frying thin slices of fresh potato1 or by mixing dry potato flakes (PFs) with water and emulsifier to make a crumbly dough that is sheeted, cut into pieces and then deep-fried2,3. PF-based crisp-making yields products of uniform shape and size and is the topic of the present study.

PFs are produced by boiling, mashing and drum-drying steam-peeled potatoes4. Starch, which makes up ~80% of their dry matter (dm)5, consists essentially of two d-glucose polymers with (1→4)-α linear links and (1→6)-α branch points—amylose (AM; 18–21% of dm; number-average molecular weight $$(\overline M _{\mathrm{n}})$$ of ~106 g mol−1; mainly linear with a few long-chain branches) and amylopectin (AP; 79–81% of dm; $$\overline M _{\mathrm{n}}$$ of ~1.8 × 106 g mol−1; a large number of short-chain branches)6,7,8—organized in semi-crystalline granules. Heating starch in excess water at and above its gelatinization temperature causes granule swelling, crystallite melting and leaching, particularly of AM, from the granules9. On subsequent cooling of the resulting starch suspensions, (leached) molecules associate and form double helices, and a gel network is formed9. After a short time, AM crystallites are formed, and over longer time periods, retrogradation of AP takes place, mainly in the granule remnants. Consequently, the initial starch gel strength is mainly determined by the extent of AM crystallization9,10. In PF manufacturing, starch gelatinizes within the potato tissue5. Drum-drying partially disrupts the potato cell walls, which causes some gelatinized starch to enter the extracellular space4.

According to the World Health Organization (WHO), total fat should not exceed 30% of the total energy intake11,12. To bring about lower oil uptake during the manufacture of deep-fried foods, without losing their distinctive palatability, it is essential to understand the mechanisms of oil uptake during deep-frying and immediately thereafter. Three main mechanisms have been examined: the replacement of water by oil in large voids and cracks during deep-frying13; oil uptake in small voids as a result of a vacuum-induced effect during cooling14,15; oil uptake in capillaries as a result of capillary action during cooling16,17.

Both pre-frying1 and post-frying18 techniques have been developed to reduce the oil content in deep-fried potato slices. Also, the use of different potato-based raw materials leads to differences in product oil content2. As starch is the main PF constituent, it is logical to examine whether differences in its gelatinized constituents cause differences in water evaporation and oil uptake and, if so, whether the acquired knowledge can provide a basis for selecting potato cultivars based on the molecular fine structure of their starch19.

Gelation of starch depends on this fine structure20,21, as reflected in the distributions of the molecular weights and sizes of its components and of their chain lengths22. In aqueous media, shorter AM chains associate more quickly than longer chains due to the high mobility of the former23,24,25. However, over a longer time frame, longer AM chains produce stronger gels because they interact with greater numbers of neighbouring chains25.

As potato cultivation, storage conditions as well as PF production practices determine the characteristics of PFs (starch), the crisp manufacturing industry faces large variability in the functionality of its raw materials, which in turn impacts dough and final product quality. Here, we examine a sample set consisting of 47 industrial PFs, which were supplied monthly by four PF manufacturers over the course of one year. The molecular structural features of their extractable starch (E-S) were evaluated using high-performance size exclusion chromatography (HPSEC). These features were then reduced to a small number of biosynthesis-based parameters to develop statistically valid correlations with gel network formation during dough making and the texture and lipid content of the crisps. We used temperature-controlled time-domain (TD) 1H NMR to carry out an in situ study of the molecular dynamics of both starch and water during deep-frying, and X-ray micro-computed tomography (µCT) to fully understand how heat-induced changes in the starchy gel network during deep-frying affect the product microstructure in relation to oil uptake. Our work provides a new basis for reducing the calorie content of crisps. It shows that it is mainly the level of short extractable AM (E-AM) chains in PFs that dictates crisp texture and lipid content. Optimization of the molecular structure of potato starch by breeding and/or of the process conditions during PF and/or crisp manufacturing can result in crisps with a significantly lower lipid content.

## Results

### Selection of PFs

The chemical compositions (on dm basis) of 47 PF samples are provided in Supplementary Table 1. Starch (77.1–82.5%), protein (6.1–9.8%) and ash (2.2–3.7%) contents were in line with previous studies5,23,26. Differential scanning calorimetry (DSC) and wide-angle X-ray diffraction (WAXD) analyses showed that starch was completely gelatinized and amorphous in all PF samples examined. Although PF starch contents differed only slightly, their E-S contents ranged from 11.7 to 22.1% of dm. Moreover, separation of debranched E-S chains by HPSEC showed that E-AM (degree of polymerization (DP) ≥ 150) and extractable AP (E-AP; DP < 150) contents ranged from 2.8 to 6.5% and from 7.0 to 15.7% of dm, respectively. That AP is abundantly present in the extracellular starch matrix results from its high solubility27 because of it being phosphorylated28. Based on the DP corresponding to the E-AM peak maximum (DP ~ 1,500, Fig. 1), the elution pattern of E-AM chains was further divided into relatively short (DP = 150–1,500) and relatively long (DP > 1,500) chains. Their contents ranged from 1.5 to 4.4% and from 0.8 to 2.8% of dm, respectively.

Two multivariate statistical methods29 were applied to present the variation in PF E-S characteristics and to select a set of PF samples to elaborate on the impact of (extractable) starch fine structure on water–oil dynamics during the manufacture of potato crisps. On the one hand, hierarchical cluster analysis (HCA), a method that explores the organization of samples in groups and among groups, grouped the PFs into four well-defined clusters. On the other hand, principal component analysis (PCA) based on three uncorrelated principal components (that is, linear combinations of the PF E-S characteristics) explained 96.7% of the total variation in E-S properties. Twelve PF samples (numbered 1 to 12) were selected to cover the variation in E-S properties (PCA) in the different clusters (HCA; Supplementary Fig. 1 and Supplementary Table 1). The distributions of E-S chains for this subset of 12 PFs are shown in Fig. 1a. Normalization of these chain distributions to the first E-AP peak revealed that the chain length distributions of E-AP branches were very similar (Supplementary Fig. 2). By contrast, the E-AM chain length distributions (Fig. 1b) varied. To develop statistically valid correlations with gel network formation during dough making and the texture and lipid content of deep-fried crisps, the E-AM chain length distributions of the 12 selected PFs were fitted to a mathematical model22,30,31 that parameterized the data in terms of a small set of biologically meaningful parameters, hE-AM,i and βE-AM,i (i = 1–2), related to the activities of the relevant starch synthase and branching enzymes, respectively. In terms of their biological meaning, higher hE-AM,i and βE-AM,i indicate higher starch synthase and branching enzyme activities, respectively31. In terms of their structural meaning, higher hE-AM,i values indicate higher levels of chains in the ith region, while higher βE-AM,i values indicate shorter chains in the ith region. Model fitting of the E-AM chain length distribution of PF 1 is shown in Fig. 1c, while all fits are provided in Supplementary Fig. 3. The E-AM fitting parameters of the 12 selected PFs are presented in Table 1. The model fittings shown in Fig. 1 and Supplementary Fig. 3 provide a good, but not perfect, fit to the data. These minor imperfections are similar to those seen in previous applications of this methodology31,32, and are regarded as having small but not qualitative implications for the validity of the structure–property relations obtained by this method.

### Dough sheet strength

The specific strengths of dough sheets made from the 12 selected PFs ranged from 1.09 N cm3 g−1 (PF 10) to 2.22 N cm3 g−1 (PF 5) (Table 2, dough sheet properties columns). Neither the PF E-S, E-AP and E-AM levels nor their protein contents were related to specific dough sheet strength. In contrast to what was the case for the absolute E-AM content in the PFs, its structural attributes (that is hE-AM,1 and βE-AM,2) were directly related to the specific strength of the dough sheets (Table 3).

### Dough proton populations

The impact of E-AM fine structure in PFs on the molecular dynamics of starch and water in dough was evaluated using TD 1H NMR. Three free induction decay (FID) proton populations (the most mobile of which was not considered33) and four Carr–Purcell–Meiboom–Gill (CPMG) proton populations were distinguished in the fresh dough sheet samples (Supplementary Fig. 4). Those that were considered were given letters A to F, in order of increasing mobility. As shown for starch model systems33,34 and bread dough and crumb samples35, the area of population A is a measure of the level of crystalline structures in a starchy gel, while the spin–spin relaxation time of population E (T2,E) relates to the rigidity of the starch gel network. Short T2,E values are indicative of rigid networks that strongly enclose the water present33.

The areas of proton population A, which ranged from 8,650 arbitrary units (a.u.) (PF 1) to 10,040 a.u. (PF 5), were positively correlated with hE-AM,1 and βE-AM,2 (Table 3). Furthermore, a strong linear relationship was observed between the specific strength of the dough sheets and the areas of proton population A. In addition, the T2,E values, ranging from 4.51 ms (PF 5) to 5.44 ms (PF 1), were negatively correlated with the areas of proton population A and directly related to hE-AM,1 and βE-AM,1.

To provide detailed information on how the strength of the water-containing starch network relates to water evaporation and oil uptake during deep-frying, four samples of PFs resulting in dough sheets with substantially different T2,E readings (Table 2, dough sheet properties columns) were selected for further work. T2,E relates to the rigidity and water inclusion of the gel network (see above). Hence, PFs 3, 4, 5 and 10 were used to evaluate the drying characteristics of dough sheets and the dynamics of starch and water during deep-frying.

### Dough desorption isotherms

During deep-frying in oil, water is lost at a high rate. As precise monitoring of dough moisture content (MC) during deep-frying is not feasible because of the very rapid moisture loss, the drying behaviour of dough sheets in the absence of oil was approached by determining their desorption isotherms. The water activity (aw) of a food system is an indicator of the availability of water to participate in physical events and chemical reactions36. Sorption isotherms describe the relation between the MC and aw at a given temperature37. Figure 2 presents MC as a function of drying time (Fig. 2a) and also the relation between MC values and aw during drying (that is, the desorption isotherms; Fig. 2b) of dough sheets made from PFs 3, 4, 5 and 10. The values of aw for the fresh dough sheets (MCs of ~38.0%) were similar. However, during drying (Fig. 2a), dough MC decreased more slowly for dough sheets with shorter T2,E times (4.74, 5.02, 4.51 and 5.36 ms for dough sheets made from PFs 3, 4, 5 and 10, respectively; Table 2). Moreover, the aw readings differed substantially among the dough sheets analysed at fixed MCs (Fig. 2b). By way of example, when the dough MC was decreased to ~10.0%, the aw values dropped to 0.525, 0.645, 0.450 and 0.655 for dough sheets made from PFs 3, 4, 5 and 10, respectively. Hence, shorter T2,E values correspond to lower aw levels at a fixed MC during drying. Finally, the values of aw for dough sheets dried for 360 min (dough MC of ~6.5%) were low and, again, similar for all dough samples analysed (Fig. 2b).

### Proton dynamics during deep-frying

Temperature-controlled TD 1H NMR was used to evaluate changes in the molecular dynamics of both starch and water in situ (mainly proton populations A and E) during deep-frying. Sunflower oil protons appeared as a single, very mobile proton population F with T2 of 640 ms, which did not interfere with any of the different dough proton populations (Supplementary Fig. 5). Figure 3 presents the areas of proton populations A (Fig. 3a) and E (Fig. 3b) of dough sheets made from PFs 3, 4, 5 and 10 as a function of deep-frying time. The areas of proton population E, containing mobile exchanging protons of water and starch in the gel network, ranged from 5,940 a.u. (PF 5) to 6,120 a.u. (PF 10) at the onset of deep-frying (0 s). They were thus quite similar for all tested samples. During deep-frying, they decreased exponentially for all analysed dough samples, whereas, at the same time, proton population A areas strongly increased.

The area of proton population E after 32 s of deep-frying differed substantially among the analysed dough sheets (1,210 and 435 a.u. for dough sheets made from PFs 5 and 10, respectively), as did the rate of decrease of this area, with rate constants (k) of 0.062, 0.069, 0.050 and 0.080 s−1 for dough sheets made from PFs 3, 4, 5 and 10, respectively. The rate of increase of the area of proton population A also differed, for example, increasing from 2,120 to 4,895 a.u. for sheets made from PF 10 while only increasing from 1,990 to 2,430 a.u. for sheets made from PF 5 after 16 s of deep-frying. The area of proton population A decreased slightly during the initial stages of deep-frying (Fig. 3a).

### Properties of deep-fried crisps

Intensive water evaporation during deep-frying promotes expansion of the food matrix and allows oil absorption38. The texture of potato crisps is typically characterized in terms of hardness and crispness39. Brittle cereal-based breakfast flakes exhibit a high spatial frequency of ruptures (Nsr) and are perceived as crispy40. Table 2 presents the texture characteristics of crisps made from the 12 selected PFs. The value of Nsr ranged from 2.34 mm−1 (PF 5) to 3.20 mm−1 (PF 10) and was negatively correlated with crisp density. Moreover, Nsr readings were significantly related to the specific strength of the dough sheets, with stronger dough structures leading to less expanded and hence less brittle crisps. Crisp hardness ranged from 13.3 N cm3 g−1 (PF 10) to 19.7 N cm3 g−1 (PF 5) and was positively correlated with specific dough sheet strength. In the present work, hE-AM,1 and βE-AM,2 were correlated with Nsr while only the former was correlated with crisp hardness.

Crisps made from PFs 3, 4, 5 and 10 (Fig. 4) exhibited a wide range of structural thickness (that is, the thickness of the solid matrix between the air- and/or oil-filled pores; Supplementary Fig. 6). The average cell thickness (Table 4) of crisps from PF 10 (55.0 µm), that is, the thickness of the sample that showed both the highest Nsr readings and the lowest hardness, was significantly lower than that of crisps made from PFs 3, 4 and 5 (62.6, 61.2 and 61.1 µm, respectively).

Finally, the lipid contents of the deep-fried crisps ranged from 21.0% (PF 5) to 32.9% (PF 10) of dm (Table 2), while the crisp MC was lower than 3% in all cases (results not shown). Table 3 shows that the lipid contents were directly correlated with the T2,E readings in fresh dough sheets. For crisps made from PFs 3, 4, 5 and 10, lipid contents were also strongly related (r = 0.997, P = 0.003) to the rate of decrease in the area of proton population E (k values, Fig. 3) during deep-frying. Analysis of the porosity and pore size distributions (Supplementary Fig. 7) indicated that crisps made from PF 10, which had the highest lipid content, also had the highest porosity (70.8%; Table 4), the highest frequency of large pores and the highest average crisp pore equivalent diameter. Indeed, the latter was 70.2 µm, compared to 50.8, 57.6 and 52.5 µm for crisps made from PFs 3, 4 and 5, respectively.

Figure 4 further shows that ~50% of the total pore volume was accounted for by one large air-filled void in all the crisps analysed (Supplementary Fig. 7). The air-filled space also had some variously sized pores connected to the large central void and some small isolated pores (Supplementary Video 1). Oil was mainly located at the product surface. Indeed, it was not present in either the big central pore or the small isolated pores in any of the samples.

## Discussion

In the present work, we have examined the impact of the fine structure of PF E-S on gel network formation and water–oil dynamics during the manufacture of potato crisps. The E-S contents as well as E-AP and E-AM levels of different industrial PF samples vary strongly, but only the structural features of their E-AM fractions are significantly related to dough and crisp properties. The quantities h and β are the fitting parameters of the E-AM chain length distributions and represent the relative level and length of chains in each region, respectively. Higher levels of E-AM chains in region 1 (larger hE-AM,1 as a result of higher starch synthase activity) and shorter E-AM chains in region 2 (larger βE-AM,2 as a result of higher starch branching enzyme activity) increase the extent to which AM crystallizes during dough making (as evidenced by larger areas of proton population A), directly resulting in stronger dough sheets. This shows that the quantity of short AM chains plays a major role in gel formation of pre-gelatinized (potato) starches, rather than the E-S content itself23. Relatively short AM chains are more mobile than larger ones, which allows them to associate readily with each other on cooling21,24,25. Indeed, we recently showed that enzymatic trimming of long E-AM chains increases the strength of PF-based dough systems due to increased levels of AM crystallization41. Moreover, higher levels of E-AM chains in region 1 lead to lower T2,E values in dough sheets made from such starches. A similar reduction in T2,E readings during bread storage has been ascribed to AP retrogradation and the associated water incorporation in both the newly formed AP crystallites and the amorphous parts of the starchy gel network33. Similarly, the higher levels of short E-AM chains noted here increase the extent of AM crystallization and the water inclusion in both the crystalline and amorphous regions of the starch network3.

In contrast to the case for potato starch, potato proteins have only poor gelation properties42. They form loose aggregates in suspension instead of ordered gel structures. Moreover, patatin, the major potato protein (30–40% of the total potato protein) contains only one cysteine residue43. As a consequence, no extended protein network can be formed because of the oxidation of its thiol group. In the present work, no relation between PF protein content and dough or crisp properties was observed. That PF protein contents are not related to dough (and crisp) structural attributes probably results from their poor gelation capacity42.

We link the strength of the starchy gel structure in dough sheets to their drying kinetics and heat-induced transformations during deep-frying. Although desorption isotherms of the dough sheets showed that stronger gel networks release water at a lower rate during drying at ambient temperature than do weaker networks, temperature-controlled 1H NMR analysis revealed that the strength of the gel network also dictates the rate at which heat-induced transformations occur during deep-frying. Intensive water evaporation results in a loss of protons in 1H NMR population E44 (Fig. 3b) and a fast transition of the starch matrix to the glassy state45. Evidence for the latter is the pronounced increase in the area of proton population A during deep-frying (Fig. 3a). The slight decrease in the area of this most rigid proton fraction, seen after 8 s of deep-frying, can be attributed to wheat starch AP crystals melting as a result of their essentially instantaneous gelatinization during deep-frying46. The decrease in the area of population E and the increase in the area of population A observed here are thus related. Moreover, the rate at which these (related) heat-induced transformations (that is, moisture loss and transition of starch to the glassy state) occur during deep-frying is associated with the strength of the water-containing gel network. In particular, stronger networks release water at a lower rate, which results in a delay of the transition of starch to the glassy state during deep-frying. This, and the fact that strong dough structures can partly withstand expansion during deep-frying, are held responsible for the dense structure, large cell wall thickness and hard texture of crisps made from dough sheets with high levels of short E-AM chains. For sugar-snap cookies47 and cereal-based breakfast flakes40, it has been reported that product cell wall thickness directly relates to product texture, with thicker cell walls leading to harder products. In line with these findings, the lower cell wall thickness for crisps made from PF 10 shows that crisp texture is (to some degree) impacted by its cellular microstructure, with thinner cell walls leading to softer crisps. The observation that the level of E-AM chains in region 1 is directly related to Nsr and crisp hardness readings provides evidence that short AM chains, in particular, dictate the texture of potato crisps.

We provide evidence that the strength of starchy gel networks determines oil absorption and that it can be impacted by controlling AM fine structure. Indeed, Table 3 shows that crisp lipid contents are negatively correlated with the level of E-AM chains in region 1. Higher levels of short E-AM chains in PFs strengthen the dough sheet, which lowers the rate of water evaporation during deep-frying. This, in turn, limits product expansion and results in a dense matrix with small pores and, hence, reduced oil absorption. It has been shown previously for deep-fried potato slices (a food matrix that strongly differs from that of the present study both in terms of dimensions and microstructure) that oil is mainly present at the product surface48. The observation that oil is mainly located at the product surface is probably caused by the oil not being able to enter the small isolated pores. Given that most of the oil is absorbed after deep-frying15,17, we further hypothesize that the lack of oil in the central void is due to insufficient oil being present at the product surface after deep-frying to completely fill the pore network or the pore network not being fully interconnected, which hinders the oil from entering the central part of the crisp.

In the context of unhealthy weight gain, the WHO recommends that total fat consumption should not exceed 30% of total energy intake11,12. This presents the food industry with the challenge to reduce oil uptake during the manufacture of deep-fried foods. In this work, which is a good example of a triple helix approach of innovation49, the Kellogg Company, with the support of the Flemish government (VLAIO), joined forces with KU Leuven and the University of Queensland to boost developments that address the challenge of calorie reduction. Here, we contribute to the demand to produce lower-fat foods by providing a better understanding of the water–oil dynamics during the manufacture of potato crisps and how it is impacted by starch fine structure. The insights gained provide guidelines for lowering the lipid content and for tailoring the texture of potato crisps by selecting potato cultivars that contain high contents of short AM chains19 and production parameters dictating their extractability.

## Methods

### Materials

Industrial samples of PFs were supplied by Clarebout, Lamb Weston/Meijer, Agrarfrost and Farm Frites. Wheat starch was from Tereos Syral. Extruded and parboiled rice flour were supplied by Herba Ingredients. Commercial maltodextrin (Maldex 190) and (monoacylglycerol- and diacylglycerol-based) emulsifier were provided by Tereos Syral and Aarhus Karlshamn, respectively. Sunflower oil was from Cargill. All chemicals were at least of analytical grade and from VWR International, unless indicated otherwise. Two peptidase preparations (trypsin from porcine pancreas (T-4799) and papain (P-3125)) and isoamylase (Pseudomonas sp. (0-8124)) were obtained from Sigma-Aldrich. They had no α-amylase side activity (as tested with the Amylazyme method (Megazyme), results not shown) and the enzyme units (EU) mentioned are as defined by the supplier.

### Potato flake composition

The ash content and MC of PF samples were analysed according to AACC methods 08-12 and 44-15a, respectively50. Starch contents were calculated as 0.9 times the glucose contents, as determined by gas chromatography of alditol acetates obtained by prior acid hydrolysis of starch, followed by reduction and acetylation of the resulting glucose units51. DSC and WAXD analyses of PFs were performed as previously described34,52 for wheat and maize starch, respectively. Protein contents were determined with an adaptation of the AOAC Official Method53 to an automated Dumas protein analysis system (EAS, VarioMax N/CN, Elt), with 6.4 as the nitrogen-to-protein conversion factor.

### Extractable starch

#### Starch extraction

Starch was extracted23,54,55 in triplicate from PFs (300.0 mg) suspended in 10.0 ml of 200 mM sodium phosphate buffer (pH 8.0) by continuous shaking (150 min−1) for 30 min at 60 °C, which is the temperature of freshly mixed dough (see below)56. After cooling to room temperature and centrifugation (2,000g, 15 min, 20 °C), the carbohydrates in the supernatants were quantified according to Dubois et al.57 using a glucose calibration curve (0–1.1 µmol ml−1) and expressed as starch (0.9 × glucose).

#### Protein degradation and starch debranching

The E-AM and E-AP contents and chain length distributions of E-AM were determined as previously described23 with slight modifications. A solution (100 µl) containing 5.88% (wt/vol) trypsin preparation and 0.65% (wt/vol) papain preparation in 200 mM sodium phosphate buffer (pH 8.0) was added to 4.5 ml of extract containing E-S. Next, 0.5 ml of 100 mM disodium ethylene diamine tetraacetate (Na2EDTA) was added and the samples were incubated for 5 h at 25 °C. After inactivation of the peptidases by heat treatment (10 min, 100 °C), the pH of the resulting samples (1.0 ml) was adjusted to 4.0 by adding 0.9 ml 1.0 M acetic acid prior to starch debranching with isoamylase, as previously described23. After enzyme inactivation (10 min, 100 °C), samples were filtered (0.45 µm, Thermo Scientific) and immediately analysed by HPSEC as detailed below. When starch extracts were analysed immediately after peptidase treatments (that is, without acetic acid addition) and after incubation with acetic acid at pH 4.023, the acetic acid had no effect on the molecular size distribution of the branched E-S molecules.

#### Chromatographic separation

Debranched E-AM and E-AP chains were separated on three columns (TSK gel G6000 PWXL, G4000 PWXL and G3000 PWXL; Tosoh) mounted in series. The HPSEC system further consisted of an injector (SIL-HTc Auto sampler), an LC-20AT pump, a CTO-20A column oven at 60 °C, a DGU-20A5 degasser and a 10-A refractive index detector at 40 °C, all from Shimadzu. Elution was performed with a 10:9 (vol/vol) mixture of 200 mM sodium phosphate buffer (pH 8.0) containing 9.8 mM Na2EDTA and 1.0 M acetic acid at 0.5 ml min−1 (ref. 23). Shodex P-82 pullulan standard aqueous solutions (1.0 mg ml−1; Showa Denko) with the following molecular weights (expressed in g mol−1) and corresponding DPs were injected to calibrate the system: 5.9 × 103 (DP 35), 11.8 × 103 (DP 70), 22.8 × 103 (DP 140), 47.3 × 103 (DP 290), 112 × 103 (DP 690), 212 × 103 (DP 1,310), 404 × 103 (DP 2,490), 787 × 103 (DP 4,860) and 1,600 × 103 (DP 9,880). The HPSEC weight distributions of debranched E-S chains with DP X, wdelog(X), were plotted against X and divided into three regions based on the DP of the eluted molecules58. Components with DP < 150 and DP ≥ 150 (the DP at which the chain length distribution shows a transition from shorter to longer chains) were considered to be AP and AM chains, respectively31,58. The E-AM and E-AP contents were calculated by dividing their respective areas in the weight distribution chromatograms of debranched E-S by the total chromatogram areas and multiplying the resulting ratios by the E-S contents31. Based on the DP corresponding to the E-AM peak maximum, the elution pattern of E-AM chains was further divided into relatively short (DP 150–1,500) and relatively long (DP > 1,500) chains. Each HPSEC analysis run included a reference PF sample for quality control reasons.

#### Model fitting to amylose chain length distributions

The E-AM chain length distributions were fitted to a biosynthesis-based model using publicly available code30,31. The E-AM chain length distributions were divided into short (region 1) and long (region 2) chain regions. The distribution of chains in each region i is dominated by (but does not exclusively consist of) chains formed by a set of starch biosynthetic enzymes, comprising starch synthase, branching and debranching enzymes31. Although the regions have no strictly defined DP ranges, their chain distributions predominantly, but not exclusively, contribute to the appropriate ranges in Fig. 1c and Supplementary Fig. 3.

### Production of potato crisps

Potato crisps were produced as previously described3, with minor modifications. The ingredients were PFs (144.0 g dm; MCs ranging from 4.0 to 9.4%), wheat starch (36.0 g dm; MC 11.1%), parboiled rice flour (22.5 g dm; MC 11.8%), extruded rice flour (22.5 g dm; MC 9.6%), deionized water (139.8 g), maltodextrin (11.5 g dm; MC 6.0%) and emulsifier (3.2 g). The PFs, wheat starch, parboiled rice flour and extruded rice flour were first blended in a Hobart N50 5-Quart Mixer for 2 min. Maltodextrin was dissolved in deionized water and the solution heated to 68 °C. The emulsifier was then added to this solution and the mixture was homogenized by vigorous stirring. Next, the emulsion was mixed with the starchy blend for 90 s in a Kitchen Aid mixer (5KFP0925) to form a crumbly dough (MCs ranging from 39.5 to 40.8%) that was then transformed into a continuous dough sheet (thickness 0.5–0.6 mm) with a single pair of counter-rotating rolls (diameter 9.02 cm, roll gap 100 µm) at 6.0 r.p.m. (Model C280 Capitani). The sheeted dough was then cut into 40 oval dough pieces (86 × 50 mm) using a metal cutter. Individual dough pieces were placed in a perforated stainless-steel mould (40% porosity) and deep-fried for exactly 12 s at 180 °C. After draining (12 s) and cooling to room temperature, crisp density was determined from the weight and volume of 40 stacked crisps. The latter was calculated from the dimensions of the metal cutter used to cut the dough sheet and from the height of 40 stacked crisps, which was determined using calipers. Dough sheets and crisps were made in triplicate and all analytical measurements were performed for each technical replicate.

### Proton NMR analysis

#### Offline analysis of proton distributions in sheeted dough

TD 1H NMR analysis was used to study the molecular mobility of water and starch polymers in the dough sheets, based on a published method34. Measurements were carried out with a Bruker Minispec mq 20 spectrometer with an operating resonance frequency of 20 MHz for 1H (magnetic field strength of 0.47 T). T2 relaxation times were measured at 25 ± 1 °C. The relaxation curves for rigid and more mobile protons were obtained by using a single pulse (90°, FID) and the CPMG pulse sequence, respectively. The lengths of the 90° and 180° pulses were 2.86 and 5.42 µs, respectively. For the FID signal, the acquisition period was 0.5 ms and 500 data points were collected. The CPMG 90°- and 180°-pulses were separated by 0.1 ms and 2,500 data points were acquired. For both measurements, 32 scans were performed to obtain good signal-to-noise ratios and the recycle delay between two separate scans amounted to 3.0 s. T2 relaxation curves were transformed to continuous distributions of T2 relaxation times with an inverse Laplace transformation using the Contin algorithm of Provencher59 (Bruker software). The different proton populations were labelled A to F in order of increasing mobility. Assignment of proton groups to specific populations was based on previous work34. Here, we focused solely on FID proton population A (containing CH protons of starch in crystalline or rigid amorphous structures) and CPMG proton population E (containing mobile exchanging protons of water and starch in the gel network), because they dominate the NMR profiles. The areas under the curve are proportional to the relative amounts of protons in the respective populations and their T2 relaxation times are dictated by the mobility of their environments.

Dough samples (~0.3 g, accurately weighed) in NMR tubes (internal diameter 7.0 mm) were gently compressed to remove air bubbles. The tubes were then sealed to prevent moisture loss during analysis and the accurately determined weights of the dough samples were used for further calculations. The areas of the different proton populations are expressed in a.u. per g of dough sheet sample. Three subsamples of each technical replicate were analysed.

#### Online monitoring of proton distributions in sheeted dough during deep-frying

Proton distributions were evaluated during deep-frying at 180 °C with the above Minispec spectrometer, essentially as previously described44. The temperature of the probe head (180 °C) was controlled with a BVT3000 tempering unit (Bruker) with nitrogen gas as the tempering medium. T2 relaxation curves were acquired by performing FID and CPMG pulse sequences as outlined above, with the time lengths of the 90° and 180° pulses being 2.84 and 5.84 µs, respectively. For the FID signal, the acquisition period was 0.5 ms and 1,250 data points were collected. The CPMG 90°- and 180°-pulses were separated by 0.2 ms, and 3,200 data points were acquired. Each FID and CPMG measurement consisted of a single scan.

Dough samples (~0.12 g) were prepared as follows. The laminated dough sheet was folded twice to obtain a layer of ~2.2 mm thickness and put in an NMR tube containing ~0.3 g sunflower oil at 180 °C. The accurately determined weights of the fresh dough samples were used to calculate the areas of the different proton populations (areas are expressed in a.u. per g of fresh dough sheet). The tubes were not sealed, to allow water evaporation during analysis. Dough samples were then deep-fried for 40 s at 180 °C in the NMR device while either FID or CPMG measurements were performed after 0, 8, 16, 24 and 32 s. Fourteen subsamples of each technical replicate were analysed (seven FID and seven CPMG measurements were performed for each technical replicate).

The decrease in the areas of CPMG proton population E during deep-frying was fitted to an exponential decay:

$${N}_{\mathrm{t}} = {N}_0{\mathrm{e}}^{ - {kt}}$$

where Nt is the proton population E area, t the deep-frying time (s), N0 the proton population E area at the onset of deep-frying (0 s) and k the rate constant (s−1). The fitting was performed for each technical replicate with the statistical software JMP Pro 14 (SAS Institute).

### Desorption isotherms

The drying of dough sheets was monitored by determining their desorption isotherms as previously described37. Oval dough pieces (86 × 50 mm) were dried at 48 °C and 5% relative humidity for 150 s, 5 min, 450 s, and 10, 15, 20, 30, 40, 60, 120, 240 or 360 min, respectively. The aw value for fresh and dried dough sheets was determined with a Novasina aw Sprint. After the measurement, the equilibrium MC of the (dried) samples was analysed according to AACC method 44-15a50. Single measurements were performed after each drying step.

### Specific dough sheet strength

Twelve pieces (82 × 41 mm) were cut from a dough sheet and weighed. Their thickness was determined using calipers and their strength was analysed as described by Souza et al.60 using an Instron 3342 system equipped with a self-tightening roller grips fixture from Stable Micro Systems. The initial distance between the grips was set at 35 mm and the test speed was 10 mm s−1. The maximum load is a measure of the strength of the dough sheet and was corrected for its density (henceforth termed specific strength, expressed in N cm3 g−1)3.

### Crisp properties

#### Instrumental texture analysis

The texture of deep-fried samples was evaluated as described by Salvador et al.38 using a TA-XT plus Texture Analyzer equipped with the Crisp Fracture Support Rig and corresponding platform from Stable Micro Systems. A test speed of 0.5 mm s−1 was applied. The maximum load, the number of force peaks (n) and the probe travel distance (d) were obtained from force–deformation curves as previously described40. The maximum load is a measure of the hardness of the crisps and was corrected for their density (further referred to as specific crisp hardness, expressed in N cm3 g−1)3. Nsr was calculated as nd−1. Twenty subsamples of each technical replicate were analysed.

#### Cellular structure

Crisp microstructure was investigated by µCT. Each technical replicate was analysed in triplicate by manually extracting three subsamples from the middle and external parts of the crisps. X-ray projections were generated with a Phoenix Nanotom system (General Electric Sensing and Inspection Technologies/Phoenix X-ray). For each scan, 1,600 radiographic projections with a resolution of 3.5 µm were acquired every 500 ms over 360° at a source voltage of 50 kV and a current of 275 µA. A filtered back projection algorithm implemented in Octopus Reconstruction (Octopus Imaging Software) was used to reconstruct the sample imaged volume. Further image processing was conducted in Avizo 9.7 (VSG).

Image processing was used to develop a workflow to define and segment three structural components, the solid matrix, pores filled with air and pores filled with oil. Reconstructed images were denoised through nonlocal means filtering to preserve interface sharpness. A mask was then generated and applied so as to only consider air zones internal to the crisps. Finally, solid matrix, air and oil were segmented by means of interactive thresholding based on the peaks present in the image greyscale histogram.

The segmented solid matrix was then exported as stacked two-dimensional images to CTan (Bruker microCT). A structural wall thickness analysis was performed using a skeletonizing and maximal sphere-fitting algorithm within the software package. Values of average thickness were obtained for each subsample and used to compute the average crisp thickness. Oil- and air-filled pore images were subjected to pore size distribution analysis, which was performed using Avizo 9.7 following a separation procedure performed by applying a watershed transform. Crisp porosity and pore equivalent diameter were computed for every subsample, and average values for each sample are presented.

#### Moisture and lipid contents

Crisps were ground in a mortar before determining their MCs with AACC method 44-15a50. The lipid contents of the (ground) crisps were analysed gravimetrically as previously described3, following sequential extraction with hexane and water-saturated butan-1-ol (Chem-Lab).

### Statistical analysis

All statistical analyses were conducted using JMP Pro 14. Data were analysed with one-way analysis of variance and Tukey multiple comparison tests (significance level P < 0.05) to verify whether mean values were significantly different. PCA and HCA were used to select representative PF samples. Pearson correlation coefficients (r) were calculated to determine correlations between the starch structural characteristics and the dough sheet and crisp textural properties.

### Reporting Summary

Further information on research design is available in the Nature Research Reporting Summary linked to this Article.