Smart Roller: Soft Sensor Array for Automated Fiber Placement

Rollers and wheels are widely used in industry and transportation, but there is seldom direct information about contact forces. A smart roller is introduced which provides real‐time pressure measurements from a soft, elastomer‐coated cylinder. The roller is designed for automated fiber placement (AFP) machines, which are widely used in the aerospace industry to manufacture complex composite parts. For optimum process performance, real‐time feedback is highly desirable for detecting flaws during manufacturing. The sensor replaces the elastomer outer layer of a standard roller with 4 by 13 tactile pixels (taxels) of soft capacitive sensors, which provide more than 1 pF of change in capacitance per taxel over a pressure range of 1 MPa. Sensors are made of silicone and mounted on a flexible printed circuit board on which a microcontroller with Bluetooth‐Low‐Energy collects and transmits capacitance data. The sensor dielectric layer is composed of pillars that increase layer compliance and sensitivity while also providing the stiffness of typical industrial rollers. The ability of the roller to measure real‐time local compaction pressure at typical manufacturing speeds enables the monitoring of spatially‐resolved pressure‐time curves, which can be used to predict and control adhesion.


Introduction
Contact forces between tires, conveyors, roll-to-roll printers, and other rubber-coated wheels and rollers and their surroundings are commonplace in transportation and industry. These forces are directly related to quality in composite manufacturing. Significant advances in composites manufacturing automation have occurred in recent decades, and automated fiber placement (AFP) systems are at the heart of this progress (Figure 1). [1,2]  Pressure sensing films and digital pressure mapping sensors based on resistive arrays are two existing technologies that enable localized pressure detection by the roller. [4,5] In both technologies, the sensing films are supplied as multiple flat rectangular sheets in limited dimensions, and therefore, cannot conform well to the curved-tooling surfaces typically used in the AFP process. Furthermore, given that the sensing films are external components, they cannot be readily integrated as a part of the process to provide continuous and real-time monitoring, which considerably limits their applications.
In this paper, a novel smart roller technology is developed for AFP processing. Flexible capacitive pressure sensing technology is developed and deployed to measure real-time local compaction pressure. Smart roller data is transmitted wirelessly to the control unit enabling real-time feedback during the AFP process.
The development of the smart roller builds on advances in soft sensors. Electronic skin sensors are sensor arrays that provide a soft and skin-like interface for measuring force and displacement. [6] Such sensors typically utilize piezoresistive or capacitive mechanisms. [7,8] Although piezoresistive sensors have higher theoretical sensitivities, they tend to display noticeable nonlinearity and high hysteresis when responding to high strain. [9] In comparison, capacitive sensors are more flexible and more stable under high strain while consuming less power. Therefore, we deploy capacitive sensing technology for this application.
While flexibility is required for the AFP roller, the sensor needs to have a soft interface that matches the dimension and stiffness of industrial elastomer rollers that are used to lay down carbon fiber composites. To avoid the need for modifications to AFP machines, the smart roller is designed to be fully modularized, containing its own battery, microcontroller, and wireless transmission module. It is designed to be easily integrated into existing industrial AFP systems.
The development of the smart roller paves the way for creating an in situ process monitoring system that predicts layup outcomes using local processing conditions as the deposition process is carried out. This can eventually reduce the need to perform costly ex situ inspections post-layup. The unique ability of the smart roller to measure localized pressure distribution of complex surfaces and provide real-time feedback can also help detect signatures of defects, as well as the underlying substrate geometry. [10] The structure of the paper is as follows: Section 2 discusses the working principle of capacitive sensing technology, application requirements, and constraints on the roller sampling rate. Section 3 discusses the sensor structure of the smart roller. The subsections describe the material selection of the top and bottom electrodes, the dielectric, and its pillar structure. Section 4 discusses the readout circuit. Section 5 outlines the fabrication process with subsections describing the sensor layer fabrication and the roller assembly process. Section 6 presents the characterization results of the smart roller. Section 7 discusses how smart roller can support pattern detection and tack prediction during the AFP process.

Capacitive Sensing
The working principle of the capacitance tactile sensor is based on the property of a parallel plate capacitor. [11] Its capacitance is approximated by where 0 is the vacuum permittivity, r is the relative permittivity, A is the overlapping area of the parallel plates formed by top and bottom electrodes, and d is the vertical distance between two plates. In Equation (1), the capacitance is inversely proportional to d. When pressure is applied to the smart roller, the deformation of the dielectric layer leads to a decrease in d and thus an increase in capacitance. In the devices described here, the dielectric is composed of an elastomer that is patterned to form pillars. The force sensing range of the capacitance tactile sensor depends on Young's modulus and Poisson's ratios of the elastomers at small deformations, the viscoelastic and nonlinear mechanics at large strains, and the geometry of the pillar structures. [12] The pillar structure increases compliance by reducing cross-sectional area, greatly reducing the stiffening effect of elastomer incompressibility. The desired material and structure should have stiffness like commercial rollers and low hysteresis.

Automated Fiber Placement Compaction Roller
Compaction rollers used in AFP processing are made of a relatively compliant outer layer that is mounted around a rigid hub. A variety of compaction rollers with different structures, dimensions, and material properties are used in industrial applications to perform tow deposition, and an appropriate compaction roller is selected for a specific application. Currently, the typical selection process of a roller depends heavily on engineering knowhow. The development of the smart roller aims to further the understanding and knowledge of compaction roller design and implementation by facilitating a systematic study of the process using instantaneous pressure measurement. The smart roller is also intended to enable real-time process feedback, reduce defects, and provide a faster production rate. The smart roller developed in this study also imitates the rollers used in the industry with respect to dimensions, materials, and overall performance.
Silicone-or urethane-based rubbers with a wide range of hardness grades are commonly used to manufacture the compliant outer layer of the AFP compaction roller. Rubber grade is identified using hardness, which is measured and quantified using the durometer scale. [2] 30 to 90 Shore A durometer rubbers are typically used as the compaction roller's flexible outer layer in this process. The stiffness of the roller is not only determined by the material used in the compliant outer layer but is also dependent on the absolute and relative dimensions of the compliant outer layer and the rigid hub.
The contact characteristics of the roller, such as the magnitude and the distribution of the compaction pressure and the dimension of the contact area, under a typical processing force, are important in understanding the mechanical performance of the roller. Under a processing force, the peak pressure from 1 up to 1.5 MPa is observed in a typical industrial roller. [10] The productivity of an AFP system is determined and limited by the maximum speed at which defect-free tow deposition can be performed. High-speed AFP systems can deposit prepreg tows at rates up to 4000 in min −1 or 1.69 m s −1 (e.g., ref. [13]). The speed is considerably reduced when prepreg tows are steered, for example, when more complex and curved structures (e.g., an aircraft fuselage or a cockpit) are laminated. Deposition speeds up to 1 m s −1 can be reached in this case, for which a typical roller with an outer diameter of 40 mm, will result in ≈240 rpm or 4 Hz rate of rotation.

Sampling Rate
Discretized feedback delay and sampling rates are key parameters that affect the performance of a control system. In our case, the dimension and the speed of the roller determine the minimum sampling frequency required while the maximum sampling frequency is restricted by the readout circuit and the sensor time constant. The time constant of capacitance measurement is given by where R is the resistance in series with the capacitor. In our design, the bottom electrodes are copper traces with negligible resistance. Since the top electrodes are near the surface of the roller, it needs to have similar mechanical properties as those conventional rollers. This restricts us to use materials with higher resistance. The overall sampling frequency (150 Hz) of the sensor is related to the number of tactile pixels (taxels) on the roller. Figure 2 shows that the maximum sensor sampling frequency is reduced with an increasing amount of taxels. This is assuming that the time constant for capacitance measurement is around 100 s per taxel, which is determined by the sampling rate of the capacitance sensing microcontroller, such as Cypress CY8C6347BZI-BLD54, used in this work. Here, we also assume that a roller with a diameter of 42 mm and a rotation rate of 4 per second in this case.
A larger amount of taxels leads to the need for a higher sampling rate since each taxel remains in contact with the surface for a shorter period. Figure 2 shows that the minimum sampling frequency must increase linearly to ensure the entire roller samples  at least once during the contact time of a taxel. The plot indicates an overall sampling rate of 200 Hz can be achieved. With additional filters and communication speed limitations in the system, the actual sample rate is close to 150 Hz.
The balance between a high taxel density and an adequate sampling frequency yields a roller sensor pattern with 52 taxels in total. A layout consisting of 4 rows by 13 columns is then chosen to fit the geometry of the roller.

Sensor Structure
The objective of the smart roller is to keep the functionality of traditional rollers while measuring the compaction forces at the nip point of the AFP process. The traditional roller consists of a rigid shell and soft outer layer (green in Figure 3 left). The smart roller also contains a hard shell (light gray in Figure 3 right), which is wrapped by a soft capacitive sensor layer that mechanically behaves as the soft outer layer. The electronics are packed into the hollow center of the structure.
The dimension of the rigid hub and the compliant outer layer of the smart roller is chosen to replicate the 43 mm-outermost diameter of the industrial roller while allowing space for readout electronics in the center. Extensive finite element (FE) studies are performed to ensure that the mechanical performance and contact characteristics of the smart roller are similar to that of typical rollers used in industry. [10] The smart roller is developed from soft capacitance sensing technology. The soft tactile sensor layer replaces the flexible outer layer outside of the hard shell. The sensor consists of a top electrode layer, a dielectric layer, and a bottom electrode layer that is directly adhered to the top of the rigid hub.
The smart roller uses the array layout shown in Figure 4  The mutual capacitance between any pair of top and bottom electrodes will increase the most when a localized pressure is applied at their overlapping area. This area of overlap is a taxel, as highlighted in red in Figure 4. In contrast, the mutual capacitance between any pair of top and bottom electrodes is insensitive to pressure applied away from their area of overlap. Therefore, by measuring the capacitance between every pair of top and bottom electrodes, we can determine the pressure applied at each taxel, providing a spatial resolution of better than 1 cm in each dimension. This starting dimension, similar to the width of the tow, enables a first direct and real-time look at non-uniformities.

Top Electrode Material
The ideal material for the top electrode of the sensor is elastic, bondable to the dielectric (silicone rubber in this case), and electrically conductive.
The material we have chosen for the top electrode is silvercoated stretchable conductive fabric (less EMF stretch conductive fabric). [14] The resistance of the 8-mm-wide conductive fabric strip is less than 1.25 Ω cm −1 , and the total top electrode resistance is less than 100 Ω. The resistance of the fabric is also stable enough with pressure to not affect capacitance readings. The low and stable resistance allows for a sensor array refresh rate of up to 150 Hz, as detailed in Section 2.3. The conductive fabric also bonds to the silicone elastomer used for the dielectric and top electrode layer.

Dielectric Layer and Material Selection
The elastomer dielectric forms a spring-like element when deformed under pressure. We choose a 20A durometer material, Dragon Skin series (Silicone rubber from Smooth-On) for its balance between stiffness and sensitivity. An increase in stiffness will cause a reduction in sensitivity, [12] while a reduction in stiffness can produce increased viscoelastic and nonlinear responses. While Dragon Skin is a softer material than the shore 60 A elastomer used in standard AFP rollers, the smart roller is designed to have a thinner outer layer and a larger rigid hub to compensate for the softer material used.
As shown in Figure 4, the dielectric layer is constructed into a pillar structure. This is done to ensure that the deformation of the dielectric at each taxel is dependent only on the pressure applied directly on the taxel, and to increase the sensitivity of taxel capacitance to local pressure. We designed the pillar structure with enough clearance between pillars to ensure that each pillar can deform independently of other pillars up to 50% strain, which is the point where the capacitive sensor has the best sensitivity.
The pillar-based design of the smart roller introduces nonuniformity in pressure distribution mechanically, as the roller exerts greater pressure in the pillar region than in the unsupported region. [15,16] This non-uniformity can be reduced by reducing the spacing between pillars; however, the requirement that pillars deform independently of other pillars limits the minimum spacing between pillars. We now estimate this minimum pillar spacing, which is implemented in sensor design.
Given that the rubber is approximately incompressible, the minimum spacing between pillars that would allow for independent deformation of each pillar without interference from adjacent pillars can be approximated by [16] Adv. Sensor Res. 2023, 2, 2200074 As shown in Figures 5a,b, the length and width of pillars, respectively, c and d are the clearance between pillars along their length and width, and is compression factor (50%). Here, it is assumed for simplicity that the entire pillar expands uniformly. The current sensor fabrication technique allows for manufacturing pillars as small as 2 mm × 2 mm. Using Equations (3) and (4), the minimum clearance required between pillars is found to be 0.8 mm.
In practice, rubber is not ideally incompressible and is much more compliant in shear, which can further contribute to pillar deformation. FE simulations are performed in the commercial finite element package Abaqus FEA to confirm pillar sizing and examine the overall mechanical performance of the smart roller compared to a baseline industrial roller. [17] A typical industrial-grade AFP roller (Figure 6a) is simulated to establish a standard baseline of compaction roller performance. The baseline roller's outer and inner diameters are 38 and 20 mm, respectively. The material is a shore A 60 durometer polymer, with Young's modulus approximated at 5.5 MPa.
For smart roller simulations, two different geometries are discussed here. In both cases, pillar dimensions are 2 (mm) × 2 (mm), while the spacing between the pillars changes from 0.5 (mm) to 0.8 (mm) between models-this spacing results in 17 and 15 columns of pillars across the full width of the roller. To reduce the demand for computational resources, only three pillars across the roller width are simulated (Figure 6b). Moreover, the full-scale process compaction force (200 N) is scaled appropriately to consider the reduced number of pillars.
The models include rigid tooling against which the roller is applied. The 3D-printed core in the smart rollers is significantly stiffer than the flexible outer layer material and therefore is replaced by a rigid shell for FE analysis. The internal surfaces of pillars are tied to the rigid internal shell to represent the adhesive bond between the roller tire and the 3D-printed core. Hard mechanical contact is defined between all model surfaces, including the pillar and outer shell of the roller.
Shore A 20 durometer rubber is used in the smart roller's outer layer and is simulated using the Mooney-Rivlin material model to represent the hyper-elastic behavior of the rubber. Empirical relationships are used to estimate the Young modulus of this rubber to be E ≈800 (kPa). [18] At the limit of small strains, Mooney-Rivlin material parameters can be estimated based on shear (G) and bulk ( ) moduli of the rubber ( C 01 = C 10 = G/2 = 0.07 (MPa) and D 1 = 2/ = 0.033 (MPa −1 )).
Four-node hybrid tetrahedron elements (C3D4H) are used to discretize the flexible outer layer of the roller. The size of the elements in the contact region is 0.25 mm. Rigid quadrilateral elements (R3D4) are used to discretize rigid surfaces (tooling and internal roller shell). The models are solved using Abaqus implicit static solver while considering nonlinear geometries and large deformations. Figure 6c-h presents the results of the finite element study. Figure 6c-e shows the resulting contact pressure profile at the tool roller interface. It is observed from figures d and e that the presence of pillars introduces variations in the magnitude of contact pressure. In areas where the rollers' outer skin is unsupported, minimal pressure is applied. Decreasing the clearance between the pillars from 0.8 (Figure 6d) to 0.5 mm (Figure 6e) helps in decreasing contact pressure variations. Simulations show that at 0.8 mm spacing, the uniformity factor achieved in compaction pressure is 56% and 42% for the 0.5 and 0.8 mm pillar spacing, compared to 79% for the baseline roller. [16,19] The uniformity factor across a given axis quantifies the uniformity of applied pressure by considering the normalized standard deviation of the pressure distribution across that axis.
A tradeoff exists between the uniformity of pressure achieved at the contact interface and the individual pillar's ability to deform freely without interference from adjacent pillars. Figure 6f illustrates the deformation of the smart roller's outer layer, confirming pillars can deform independently with 0.8 mm spacing, and therefore, it is implemented in this work. Figure 6g summarizes compaction pressure distribution for all rollers across their width and Figure 6h compares the normalized force-displacement behavior of the baseline roller with that of the smart roller with 0.8 mm spacing between pillars (up to 200 N). Displacements are normalized by the maximum displacement of the baseline roller (1.04 mm) in this graph. The baseline roller is initially stiffer under small loads, which can be attributed to the stiffer material used in its outer layer. However, the smart roller's stiffness quickly rises with increasing the applied force, with bulk rubber deformation becoming the primary mode of deformation, providing very similar differential stiffness. Finally, the maximum deformation of the smart roller is only 14% larger than that of the baseline roller.

Bottom Electrode
The bottom of the dielectric layer is attached to a flexible PCB. We chose 50 m polyimides as the substrate and 35 m copper as the conductive layer for the flexible PCB. [20] Polyimide was chosen due to its great thermostability under high temperatures (up to 288°C). This property allows us to solder connectors onto the flexible PCB.
The length of the PCB is 113 mm, equal to the circumference of the roller. Its width is also the same as the width of the roller in order to cover the entire surface of the roller. The PCB contains 13 www.advancedsciencenews.com www.advsensorres.com

Readout Circuit
Our readout circuit (Figure 7) was designed around the CY8C6347BZI-BLD54 (BLD54) microcontroller unit from Cypress Semiconductor. It provides mutual capacitance measurements in several picoFarad ranges. Limited by the RC time constant ( ) of the sensor, the single measurement time T is set to 100 s where T > 10 × . When the measurement condition Step-by-step procedure of sensor fabrication. In Step 1, the fabric electrodes are aligned and cut. In Step 2, the elastomer is poured into the mold, followed by a degassing process in Step 3. In Step 4, the conductive fabric is laid down on the mold and cured in an oven, after which, the structure is cured and then removed from the mold (not shown).
T > 10 × is satisfied, the capacitance measurement is robust under variations in trace resistance. Its internal high-speed analog multiplexer allows easy switching between trace measurements. Its compact footprint (24 mm × 18.5 mm × 4.5 mm), low power consumption, and wireless communication capability through Bluetooth-Low-Energy (BLE) allow the readout circuit to operate inside the compact space margin of the roller. The peak power consumption of the whole system is 20 mW. When connected to a 3.7 V, 100 mAh LiPo battery, the roller can continuously run for at least 8 h. The readout circuit is connected to the sensor via a flat flexible connector (FFC). The BLD54's capacitance to digital converter converts capacitance into an integer raw count. This process is explained in Sections SA and SB, Supporting Information.

Roller Fabrication
The roller (Figure 3) is composed of three main components: a flexible capacitive sensor, including the PCB (Figure 7), onto which the dielectric and stretchable electrodes are bonded (Figure 8, step 1), the roller body (Figure 4), and the readout electronics ( Figure 7). The fabrication of the flexible sensor is done in-house and detailed below (Figures 8 and 9).
The roller body is printed with high-temperature resin on a Formlabs Form 3 resin printer. When possible, the fabrication process uses commercially available tools, such as laser cutting, 3D printing, and external PCB sourcing, to enable reproducibility and scaling of production. Steps involved in assembling the sensor, readout circuit, and roller shell. In Step 1, the sensor is bonded on a PCB. In Step 2, the sensor is wrapped around a rigid shell. In Step 3, the microcontroller and battery are connected. The final product is shown in Step 4.

Sensor Fabrication
The two main steps for sensor layer fabrication are laser cutting the conductive fabric for the top electrode and molding the dielectric layer. These are depicted in Figure 8.
The top electrodes are prepared by laser cutting the conductive fabric. Prior to cutting, we flattened and laminated transparency sheets on both sides of the fabric at 110 ○ C and used speed one on the ProLam Photo 6 Roller Pouch Laminator. Cutting was performed using 100% power and 40% speed on the Universal Versa Laser VLS 4.60, as shown in Figure 8, step 1.
The dielectric is made using platinum-cured silicone (Smooth-On DragonSkin Shore 20 A). The monomer and crosslinker of the silicone were mixed by hand and degassed in a vacuum chamber. Molds are 3D printed with ABS plastic on the AnyCubic Chiron 3D printer. The mixture was then poured into the mold as shown in Figure 8, step 2. The mold was printed with the pillar features.
In Figure 8, step 3, the four conductive fabric strips in the sensor layer are aligned on and taped to a large transparency sheet. In Figure 8, step 4, the side of the transparency sheet with the conductive fabric is pressed onto the uncured dielectric silicone layer while it is still in the mold. The mold is then placed into a vacuum oven and then cured at 60°C for 1 h. The dielectric layer, including the bonded electrodes, is removed from the mold and cooled down to room temperature before roller assembly.

Roller Assembly
The roller assembly process is illustrated in Figure 9. The sensor layer and PCB layer are first bonded together using RTV silicone (Silicone Solutions SS6004VF+) to produce the sensor shown in Figure 9, step 1. In Figure 9, step 2, the sensor is wrapped around the rigid shell of the roller. In Figure 9, step 3, the battery connector is soldered onto the readout circuit, which is then inserted into the roller and connected to the sensor bottom electrodes via a 40pin FFC. Because the PCB has a layer of adhesive on the back, the sensor is simply wrapped around the 3D-printed roller shell with the adhesive facing the shell. Finally, the roller is inspected to ensure that the PCB layer is bonded well with the shell to form the soft outer layer of the roller, as shown in Figure 9, step 4.

Characterization of the Sensor
The smart sensor array is first characterized in the flat (unrolled) state using an Instron Universal Testing Machine (model 5969), with the clamp and custom indenter shown in Figure 10a. The sensitivity is investigated at two loading rates to check for consistency. To do this, a 1 N contact force is applied. The compression is then ramped up to a ≈60% strain at 0.25 mm s −1 , followed by a ramp back down at the same rate. After 2 min of rest, the experiment is repeated at 10 mm s −1 . The readout circuit records the change in capacitance of the sensor. The responses are plotted in Figure 10b-d, enabling the stress-strain curves, and sensitivities to stress and strain to be compared at the two rates. In the plots, the displacement is divided by dielectric thickness to obtain an effective strain, while the stress is the force divided by the contact area.
The stress-strain curves shown in Figure 10b show similar mechanical properties despite differences in scan rates. with the material appearing to be slightly stiffer at higher rates. In Figure 10c,d, the relative changes in capacitance, ΔC/C 0 , as functions of effective strain and stress also show very similar responses, demonstrating consistency in measurement. As in the stress-strain curve, there is some non-linearity. The "hysteresis" is expected due to the viscoelastic response. Overall, the hysteresis seen in the strain and stress responses leads to an uncertainty of up to +/− 7% of full scale. This uncertainty could be reduced using a stiffer elastomer (at the cost of sensitivity) or compensated for using a load history-dependent viscoelastic model. The time between samples is 6.3 ms. It is clear that the sensor module can record changes in load even over this short time. The 10 mm s −1 rate of compression is the most we can stably achieve with the Instron, but only corresponds to a translation rate of the roller of about 0.024 m s −1 . The sensor response is now evaluated for rolling contact, enabling faster rates to be observed. Figure 10. a) The indenter used to characterize the sensor. The probe is a 3D-printed, 4.8 mm by 8 mm rectangular head, which is equal to the size of the taxels on the sensor. b) Characterization data shows stress/strain curve (sample size 1). c) Characterization data shows ΔC/C 0 with applied strain. C 0 represents the zero-stress capacitance (normalized data, sample size 1). d) Characterization data shows ΔC/C 0 with corresponding stress (normalized data, sample size 1).

Force Distribution When Rolling
Position-dependent force distribution and speed of response are investigated in this section. As a first step, a hard roller is translated across a flat sensor at different rolling speeds, with a constant downward force. The setup is shown in Figure 11a. The sensor array is adhered to the AFP machine's platform. A hard 3D printed roller (red) indents the sensor through an 8 mm wide, 1 mm thick ridge that extends around the circumference. The ridge is aligned with one of the top electrodes in the sensor and rolls along it, sequentially compressing each taxel in the row. This inverse curvature experiment allows for local control of the force and calibration of the array prior to mounting it on a roller. In production, we envisage running the fully assembled smart roller over an array of mechanical ridges in order to perform rapid calibration on demand.
In order to examine the effects of speed under a constant normal force, the rigid roller moves at a horizontal speed of 0.05, 0.25, and 0.5 m s −1 with a constant normal force of 50 N across the sensor array. Four taxels in sequence are selected for analysis, and the data from the three test cases are plotted in Figure 11b-d. The taxels are activated sequentially in time. Each taxel shows a rise in capacitance, peak, and then decay, as the hard roller passes over it. This test shows that the maximum ΔC/C 0 of each taxel drops slightly with increasing roller speed but remains within ≈10% of the slow 0.05 m s −1 response. The drop in sensitivity with increasing speed is not surprising since the viscoelastic elastomer is expected to be stiffer on shorter timescales. If improved accuracy is desired, the time and frequency dependence of the elastomer dielectric can be measured and compensated for. The sampling rate of 150 Hz is starting to limit resolution at the fast rolling rates. If much faster sampling and rolling speeds are desired, this could be accomplished by selectively sampling taxels that are in contact with the substrate and the tow.
A constant speed of 0.05 m s −1 is applied to the roller under normal forces of 50, 60, 70, and 80 N to investigate the sensitivity (our tool cannot apply lower forces). The first taxel is chosen for analysis here, and its ΔC/C 0 values from the four test cases are plotted in Figure 11e. The maximum value of ΔC/C 0 increases with an increasing normal force. The peak values are extracted to compare the relation between force and maximum ΔC/C 0 , and the relation is found to be approximately linear as shown in figure (f).
The output of each taxel is a function of both the force and angular position of the roller. The curvature of the roller or the indenter means that interpretation is not quite as simple as with the square indenter applied to a flat surface. A simple method of mapping force is to take only peak values and create a heat map. A more sophisticated approach would make use of the relationship between force and position. Given that the taxel responses overlap, either the signal from the closest taxel plus the position information or a weighted sum of the signals from the nearest taxels, could be used in the future to create a continuous force map. Further improvement could also account for rate, variations in sensitivity of individual taxels, and viscoelastic responses. Even in the absence of these advances, the smart roller is a valuable new tool as it provides real-time and spatially resolved force information. Following temperature response characterization, the roller is applied to estimate tack based on local pressure information.

Thermal Stability
During the AFP process, the workpiece is often heated to between 25 and 50°C. This means that the roller must have good temperature stability to deliver accurate measurements within the operating temperature range.
The experimental results in Figure 12 show the roller's capacitance measurement versus force at 25, 35, and 50°C ( Figure 2). The experiment was performed by holding the temperature of the workpiece and the temperature of the contact point of the roller constant at 25, 35, and 50°C. The roller is pressed against a flat workpiece while the force is then incrementally increased from 0 to 200 N with an increment of 20 N.
The maximum measured error is 5%, which occurs at 200 N. This variation may occur because silicone is known to change its dimension and dielectric constant with temperature. When the temperature is increased from 25 to 50°C, the dielectric constant of PDMS decreases by 2.1% while its dimension expands by 1%. [21] For the smart roller, volume expansion and a decrease in dielectric constant both reduce capacitance. As a result, the roller would expect to register a slight decrease in capacitance at higher temperatures. This change is also seen in the value of C o , which drops by 4.1% under the same applied force.

Pattern Detection
It is important in AFP processing to detect the location and shape of defects or features in the substrate over which prepreg tows are deposited. If untreated, layup defects create resin-rich areas and/or porous areas, as well as deviations in fiber orientation, which ultimately has a detrimental effect on the mechanical properties of final cured parts such as tensile, compressive, and interlaminar properties (for instance, refs. [22,23]).
One application of smart AFP rollers is detecting the underlying geometry over which prepreg tows are dispensed. Any defect or geometric feature that constitutes a height difference with respect to the base geometry creates local variations in the distribution of compaction pressure. [24,25] This difference can be measured by the smart roller. The variations in the local pressure can be used to detect the underlying surface shape and identify differences in applied pressure, which lead to variations in bonding or tack.
An experimental case study is designed to demonstrate the application of the smart roller in pattern detection. The roller was tested on a table-top AFP demonstrator, as shown in Figure 13a. [11,12] The base layer is a single layer of AS4/8552 UD prepreg from the Hexcel Corporation. [26] Six layers of 1-in-wide (25.4 mm), 0.2 mm thick prepreg strips are cut from the same material and are manually placed at a 45°angle with respect to the smart roller's trajectory of motion. To perform the experiments, 50 N force was applied to the smart roller, and the roller was automatically moved over the feature using the AFP demonstrator. The smart roller's response is recorded and visualized for further analysis. Figure 13b shows a visualization of the smart roller's response as a heatmap, wherein each square corresponds to a taxel, and the intensity of the color is proportional to the intensity of change in capacitance (corresponding to local pressure measurements).
It can be observed in Figure 13a that as the roller reaches the diagonally placed strip of prepregs, the right-most taxel that comes into contact with the strip first measures elevated levels of capacitance change due to the height difference (time 1). As the roller Figure 13. a) Pattern detection setup with the roller attached to an AFP demonstrator. The base layer is comprised of a single layer of UD prepreg. Six layers of 1-in-wide (25.4 mm) prepreg, with an approximate total thickness of 1.2 mm, are placed diagonally with respect to the smart roller's motion trajectory. The roller is lowered and then moved over the diagonal strip. b) Time sequence data obtained from the roller is visualized as a series of four heatmaps, wherein each square corresponds to a taxel, and increasing darkness indicates an increase in mutual capacitance and displacement. The darker horizontal strip corresponds to the region of the roller in contact with the surface. When the roller first meets the diagonal strip (1), the rightmost taxel shows the strongest signal. As the roller proceeds over the strip, the high-force taxel moves to the left (2-4), enabling the ridge to be identified. moves forward (time 2-4), other taxels come into contact with the strip of prepreg layers. In turn, they measure elevated levels of capacitance compared to the baseline value, as expected from a diagonal geometric pattern.

Tack Prediction
In AFP processing, prepreg tack is the primary mechanism that holds layers of composite together and resists the formation of layup defects. Tack is significantly influenced by the pressure history of the material. [27] Furthermore, multiple prepreg tows are deposited on tooling as a "course" during layup in a single pass. [2,28] The existence of gaps, overlaps, and ply drops in the underlying substrate and lamination over curved and complex tooling geometries create considerable variations in local pressure under the AFP compaction roller. This leads to differences in the bonding strength and can result in defects, delamination, and mechanical failure. Development of the smart roller allows individual local pressure histories for each tow delivered in a course to be considered. During the process development stage, local pressure measurements can be used to simulate the resulting prepreg tack between individual prepreg tows and substrate. [3,[29][30][31] Furthermore, during manufacturing, the combination of in situ local measurements from the smart roller and physics-based tack models can lead to the development of online process monitoring systems that continuously screen predicted tack levels as tows are deposited.
A case study is designed to demonstrate the application of the smart roller in process development. Figure 14a shows the experimental setup. The base substrate is a single ply of UD prepreg. In order to demonstrate the impact of height difference, an 8ply thick substrate (substrate 1) and a 4-ply thick substrate (substrate 2) are placed on top of the base substrate. Three 1/4-in-wide prepreg tows are cut from the same AS4/8552 prepreg material and placed on top of each substrate. Figure 14b schematically shows the overall geometry and the column of taxels compacting each tow.
Using the AFP simulator, 50 N of compaction force is applied to consolidate the prepreg tows at a speed of 10 mm s −1 . The roller seen in Figure 14a is rolled along and tows 1-3 and is applying the specified compaction force. The experiment is performed at room temperature (22°C). Throughout the experiment, raw capacitance data from the smart roller are recorded for further analysis ( Figure 14c).
As shown in Figure 14b, only taxel columns a, c, and d correspond to prepreg tow locations in this experiment. As expected, a taxel shows the largest signal, followed by d and then c. Using the characterization curve of Figure 10d, local pressure can be estimated from the normalized change in capacitance presented in Figure 14c.
RAVEN simulation software [29] offers an implementation of a state-of-the-art physics-based prepreg tack model developed in refs, 3,30,31], and is used to predict the prepreg tack obtained for each tow during the experiment.
The following presents a summary of simulation results featuring tack Energy of Separation per unit area (EoS), as a key indicator of tackiness, predicted using the pressure, time, and temperature for each tow. The prepreg tow corresponding to the location of taxel a is placed on the highest substrate and therefore experiences the largest amount of local pressure. A higher value of compaction force results in a superior quality of contact, and therefore, the highest EoS (42.6 N m −1 ). On the other hand, the prepreg tow corresponding to the location of taxel c is placed on the base substrate. The presence of substrates 1 and 2 does not allow the compaction roller to consolidate tow c very well, and therefore the lowest EoS (9.4 N m −1 ) is predicted from simulation. The prepreg tow, corresponding to the location of taxel d, experiences a comparatively moderate amount of local pressure, and therefore, a moderate EoS (26.1 N m −1 ) resulted from the simulations. Videos S1-S3, Supporting information, confirm that, as expected, the bonding is best with higher pressure applied, leading to more resistance to peel-off.

Conclusion and Future Work
Rubber-coated wheels and rollers are used in vehicles, conveyor belts, and printers, as well as other roll-to-roll processes. We demonstrate a method of integrating sensors into the elastomer to create real-time feedback. The sensor arrays are applied to AFP. The leading technology in the automated lamination of composite materials that have enabled manufacturers to fabricate highquality, complex structures at higher rates and with more consistent quality. Compaction pressure is well known to significantly impact the development of intimate contact, consolidation, and tack between prepreg layers deposited using AFP. Quality of consolidation and tack determines layup quality outcomes, yet currently, there is little to no real-time knowledge of the state of compaction under the roller. The present work develops a novel smart roller technology capable of measuring compaction pressure locally at the process nipping point, in situ of the deposition process.
The smart roller retains the mechanical properties of the industrial roller while adding real-time pressure-sensing functionality. The design, fabrication, material selection, and characterization of the smart roller are discussed in this paper. A readout cir-cuit that balances a compact form factor with fast, robust sensing capabilities was designed and implemented for the smart roller.
With the current design, the smart roller is shown to measure up to 1 MPa pressure at the process nipping point. The system can measure at the rate of 100 s per taxel (≈150 Hz for the entire system), corresponding to a maximum traveling speed of 1 m s −1 for the AFP system, while ensuring five measurements for each taxel during its duration of contact. Furthermore, two applications of the smart roller in the detection of the underlying layup geometry, as well as in predicting prepreg tack, are presented.
The sensing is made possible using a dielectric design consisting of elastomeric pillars, which enables increased sensitivity, stretchable conductive traces, and flexible PCB connections that effectively connect with the electronics. The miniaturized circuitry is designed to be embedded in the roller itself to form an autonomous, stand-alone system that can readily be mounted on AFP machines. The next step is to explore the full integration of the smart roller into an AFP machine and to develop algorithms that inform quality control and enable process optimization.

Supporting Information
Supporting Information is available from the Wiley Online Library or from the author.