Roll-to-roll (R2R) manufacturing is a class of continuous industrial processes that fabricate thin-film devices by moving flexible substrates through a series of rollers. R2R techniques enable efficient, high-throughput production of advanced technologies such as lithium-ion batteries [1], flexible solar cells [2], and fuel cell membranes [3]. A key step in these processes is transferring functional materials or devices from a donor to a target substrate for device integration. Traditional batch-style transfer methods are inefficient [4], and while R2R-compatible methods exist, many have limited throughput [5] or use hazardous chemical etchants that destroy the donor substrate [6].
Recent advancements in R2R mechanical dry transfer provide an efficient, eco-friendly alternative to liquid-based or discrete transfer methods [7], [8]. As shown in Figure 1, the process involves laminating a 2D material or device between donor and target substrates, peeling them apart so the material adheres to the target, and rewinding the functional material or device for further processing and the donor substrate for recycling. This process can be repeated for a continuous series of 2D devices or material samples.
Successful R2R dry transfer requires precise control of peeling conditions, particularly web tensions and speeds [7], [8], [9]. Maintaining optimal web tension is critical but challenging due to nonlinear peeling dynamics and abrupt material property changes from device patterning [10], [11]-challenges absent in typical R2R processes without peeling. Furthermore, control design must account for practical constraints such as motor torque limits [11] and online computational capacity. Alongside these challenges, R2R dry transfer has an iterative structure that can be leveraged to improve control. Each 2D sample or device transfer represents one "transfer task" iteration, meaning data from previous iterations can be used to improve performance in future transfers within the continuous manufacturing line.
Researchers have tried to develop controllers for the R2R mechanical dry transfer process. Tuned PI controllers ignore the nonlinear process model [10], [12], while linear differential inclusion (LDI)-based optimal controllers are overly conservative [13], [14], [15]. Both methods neglect input constraints. Nonlinear MPC offers high performance but incurs significant online computational costs [11]. In addition, these existing controllers for R2R mechanical dry transfer do not take advantage of its iterative structure. Thus, a control method is needed that incorporates the nonlinear process model, considers input constraints, has minimal online computation, and uses previous task data to improve performance.
This last consideration motivates an iterative learning control (ILC) approach. ILC, illustrated in Figure 2, improves system performance by adjusting control inputs based on measurements from previous task iterations. The most common form, classical ILC, uses a constant learning gain to adjust the control based on the tracking error from the previous task but does not account for nonlinear models or input constraints [16]. Norm-optimal ILC uses a constant linear-time-varying process *Corresponding author. Email: cbmartin129@uexas.edu 2 MECC25_39, Martin
model to improve the controller after each task iteration with respect to a quadratic cost function. If there are no input constraints, this optimal update takes the form of a constant learning gain [17]. Otherwise, a quadratic program (QP) must be solved after each iteration [18]. When a constant linear approximation of the nonlinear model is insufficient, Sequential Quadratic Programming ILC (SQP ILC) becomes necessary. SQP ILC formulates the ILC problem within the SQP framework, which is a robust method for nonlinear constrained optimization offering reliable performance and superlinear convergence when initialized near a feasible solution [19]. Each iteration of SQP ILC involves performing the task, linearizing the nonlinear process model, and then using these measurements and linear dynamics to solve a QP that improves the control [19].
Given the nonlinear dynamics and input constraints in R2R mechanical dry transfer, SQP ILC is a suitable approach. We thus propose an SQP ILC approach to regulate R2R mechanical dry transfer, incorporating nonlinear system dynamics and input constraints while maintaining low online computational cost. The key contributions of this study are as follows: 1) SQP ILC is used to control an R2R system for the first time. 2) R2R mechanical dry transfer is formulated within the ILC paradigm for the first time. 3) The practicality of this approach is validated experimentally. 4) A case study suggests that the proposed method could significantly improve the R2R transfer quality of advanced 2D materials.
The paper is structured as follows: Section 2 introduces the dynamic model of R2R mechanical dry transfer, followed by the control problem formulation in Section 3. Section 4 presents the SQP ILC method, while Section 5 validates this approach through an experiment. Finally, Section 6 demonstrates the utility of the approach on an advanced manufacturing example.
This section provides an overview of the R2R mechanical dry transfer dynamic model presented in [20], [21]. The variables and parameters of the system are listed in Table 1, and the process is illustrated in Figure 3.
The difference equations for the web velocities and unstretched lengths are given as follows:
𝑡 + 1) = 𝑣 𝑖 (𝑡) + -𝑅 𝑖 2 𝑇 𝑖 (𝑡) -𝑓 𝑖 𝑣 𝑖 (𝑡) + 𝑅 𝑖 𝑢 𝑖 (𝑡) 𝐽 𝑖 • 𝑇 𝑆 (1) 𝑙 1 (𝑡 + 1) = 𝑙 1 (𝑡) + 𝑣 1 -𝑣 𝑝 (𝑡) 1 + 𝜀 1 (𝑡) • 𝑇 𝑆 (2. 𝐴) 𝑙 𝑖 (𝑡 + 1) = 𝑙 𝑖 (𝑡) + ( 𝑣 𝑝 (𝑡) 1 + 𝜀 1 (𝑡) -𝑣 𝑖 (𝑡) 1 + 𝜀 𝑖 (𝑡) ) • 𝑇 𝑆 (2. 𝐵) where 𝑡 is the discrete timestep, 𝑇 𝑠 is the sampling time, and 𝑖 = 2, 3. Equation (2) can be interpreted as a mass conservation, where 𝜀 𝑖 = 𝑇 𝑖 (𝑡) 𝐴 𝑖 𝐸 𝑖
and 𝑣 𝑝 is the peeling front velocity. 𝑣 𝑝 is a nonlinear function of the web tensions, web speeds, and energy balance at the peeling front [20]. This energy balance can be summarized as follows [21]:
where the time dependence in Eq. ( 3) has been omitted for brevity. Thus, there is a nonlinear relationship between the three web tensions and the bending and adhesion energy at the peeling front, parameterized by 𝑈(𝑡).
Using the geometry in Figure 3, the three web tensions, 𝑇 𝑖 (𝑖 = 1, 2, 3), can be uniquely determined from the corresponding unstretched web lengths, 𝑙 𝑖 , and vice-versa. This relationship is computed through an iterative root-finding algorithm, represented as [20]: Initial Parameters Perform Task Control Improvement Measurements Updated Control Parameters = + 1 = FIGURE 3: SIMPLIFIED CARTOON OF R2R MECHANICAL DRY TRANSFER (LEFT); ZOOMED PEELING FRONT (RIGHT).
Moment of inertia roller 𝑖 (kg m 2 ) 𝑓 𝑖 , 𝑖 = 2,3
Rotational friction for roller 𝑖 (m/s) 𝑏 Width of the contact surface (m)
*Corresponding author. Email: cbmartin129@uexas.edu 3 MECC25_39, Martin
Thus, using Eqs. ( 1)-( 5), it is possible to form a state space model of the R2R mechanical dry transfer system: 𝒙(𝑡 + 1) = 𝒇(𝒙(𝑡), 𝒖(𝑡), 𝒘(𝑡)) (6) where 𝒙(𝑡) ∈ ℝ 𝑛 𝑥 = [𝑣 2 (𝑡), 𝑣 3 (𝑡), 𝑇 1 (𝑡), 𝑇 2 (𝑡), 𝑇 3 (𝑡)] 𝑇 , 𝒖(𝑡) ∈ ℝ 𝑛 𝑢 = [𝑢 2 (𝑡), 𝑢 3 (𝑡)] 𝑇 , and 𝒘(𝑡) ∈ ℝ 𝑛 𝑤 = 𝑈(𝑡). This model has been experimentally validated [20], [21], and, with small adjustments, is fast enough for online control [11].
R2R mechanical dry transfer can be modeled as an iterative discrete-time system where each iteration, corresponding to the transfer of one sample from the donor to the receiver substrate, consists of 𝑁 timesteps with a step size 𝑇 𝑠 . The values of 𝑁 and 𝑇 𝑠 are determined by the pre-set unwinding velocity and control system sampling rates. The system dynamics follow:
𝒙 𝑘 (𝑡 + 1) = 𝒇(𝒙 𝑘 (𝑡), 𝒖 𝑘 (𝑡), 𝑡), 𝑡 = 1, 2, … , 𝑁 -1 (7) where 𝒙 𝑘 (𝑡) ∈ ℝ 𝑛 𝑥 and 𝒖 𝑘 (𝑡) ∈ ℝ 𝑛 𝑢 are the state and control at timestep 𝑡 in iteration . The initial condition 𝒙 𝑘 (1) = 𝒙 0 is assumed constant, as the R2R line typically reaches steady-state operation between transfers. Note that Eq. ( 7) is a reformulation of Eq. ( 6) within an iterative framework. The time-dependence in Eq. ( 7) accounts for 𝒘(𝑡) ∈ ℝ 𝑛 𝑤 , which is assumed to be known a priori [22].
Next, suppose the control law has the following form:
where 𝒖 ̅ 𝑘 (𝑡) is the feedforward control signal, 𝑲 is a stabilizing feedback gain, 𝒙 ̂𝑘(𝑡) is the measured state, and 𝒙 𝑟 (𝑡) is the desired or reference state. The control input is constrained by motor torque limits: 𝒄 𝒖 ≤ 𝒖 𝑘 (𝑡) ≤ 𝒄 𝒖 , 𝒄 𝒅𝒖 ≤ 𝑑𝒖 𝑘 (𝑡) ≤ 𝒄 𝒅𝒖 (9) where 𝑑𝒛(𝑡) = 𝒛(𝑡) -𝒛(𝑡 -1).
Let the process of obtaining measurements from one transfer iteration using the control law (8) be represented as:
where 𝑭: ℝ 𝑛 𝑢 ×(𝑁-1) → ℝ 𝑛 𝑥 ×𝑁 maps the feedforward control sequence 𝑼 ̅ 𝑘 = [𝒖 ̅ 𝑘 (1), 𝒖 ̅ 𝑘 (2), … , 𝒖 ̅ 𝑘 (𝑁 -1) ] to the measured system states 𝑿 ̂𝑘 = [𝒙 ̂𝑘 (1), 𝒙 ̂𝑘(2), … , 𝒙 ̂𝑘(𝑁) ]. The goal of each iteration is to refine the feedforward control 𝑼 ̅ 𝑘 using past measurements (𝑿 ̂1, 𝑿 ̂2, … , 𝑿 ̂𝑘-1 ) such that the input constraints (9) are met, and the following cost function is minimized:
+‖𝑑𝒖 ̅ 𝑘 (𝑡)‖ 𝑺 + ‖𝒙 𝑘 (𝑁) -𝒙 𝑟 (𝑁)‖ 𝑸 𝑓 (11) where 𝑸, 𝑹, and 𝑺 are weighting matrices that penalize tracking error, control effort, and control variation, respectively; and ‖𝒛‖ 𝑴 denotes the quadratic form 𝒛 𝑇 𝑴𝒛. This control problem formulation is summarized in Figure 4.
The key next question is: how to use the system measurements to improve the feedforward trajectory in each iteration according to the objective function (11), while ensuring compliance with the dynamic constraints (7) and the input constraints ( 9)? This endeavor presents a challenging nonconvex optimization problem where measurements from new task iterations need to be integrated into the control updates.
These considerations motivate the use of the SQP ILC framework, first introduced in [19]. Assuming a quadratic objective, SQP iteratively repeats four steps: 1) execute and measure the controlled task, 2) linearize the dynamic and input constraints around a "guess" of the optimal trajectory, 3) solve a QP to find the optimal update direction, and 4) update the control policy using an appropriate step size. The remainder of this section presents a formulation of this approach crafted to fit R2R mechanical dry transfer.
After obtaining measurements from the th transfer task (Step 1), the linearized dynamics are updated as follows (Step 2): 𝑨 𝑡,𝑘 ≈ 𝜕𝒇 𝜕𝒙 (𝒙 𝑘 𝑔 (𝑡), 𝒖 𝑘 𝑔 (𝑡), 𝑡) (12. 𝐴) 𝑩 𝑡,𝑘 ≈ 𝜕𝒇 𝜕𝒖 (𝒙 𝑘 𝑔 (𝑡), 𝒖 𝑘 𝑔 (𝑡), 𝑡) (12. 𝐵) 𝒅 𝑡,𝑘 = 𝒙 ̂𝑘(𝑡 + 1) -𝒙 𝑘 𝑔 (𝑡 + 1) (12. 𝐶) where 𝒙 𝑘 𝑔 (𝑡) and 𝒖 𝑘 𝑔 (𝑡) represent the "guess" of the optimal state and control. The following QP is then solved to find the update direction for 𝑼 ̅ 𝑘 (Step 3): min Δ𝑿 𝑘 ,Δ𝑼 ̅ 𝑘 ∑‖𝒙 𝑘 𝑔 (𝑡) + Δ𝒙 𝑘 (𝑡) -𝒙 𝑟 (𝑡)‖ 𝑸 𝑁-1 𝑡=1 +‖𝒖 ̅ 𝑘 (𝑡) + Δ𝒖 ̅ 𝑘 (𝑡)‖ 𝑹 + ‖𝑑(𝒖 ̅ 𝑘 (𝑡) + Δ𝒖 ̅ 𝑘 (𝑡))‖ 𝑺 +‖𝒙 𝑘 𝑔 (𝑁) + Δ𝒙 𝑘 (𝑁) -𝒙 𝑟 (𝑁)‖ 𝑸 𝑓 (13. 𝐴) subject to the dynamic constraints: Δ𝒙 𝑘 (𝑡 + 1) = 𝑨 𝑡,𝑘 Δ𝒙 𝑘 (𝑡) + 𝑩 𝑡,𝑘 Δ𝒖 𝑘 (𝑡) + 𝒅 𝑡,𝑘 (13. 𝐵) and the input constraints: 𝒄 𝒖 ≤ 𝒖 𝑘 𝑔 (𝑡) + Δ𝒖 ̅ 𝑘 (𝑡) + 𝑲Δ𝒙 𝑘 (𝑡) ≤ 𝒄 𝒖 𝒄 𝒅𝒖 ≤ 𝑑 (𝒖 𝑘 𝑔 (𝑡) + Δ𝒖 ̅ 𝑘 (𝑡) + 𝑲Δ𝒙 𝑘 (𝑡)) ≤ 𝒄 𝒅𝒖 (13. 𝐶) FIGURE 4: SUMMARY OF THE ILC PROBLEM FORMULATION FOR R2R MECHANICAL DRY TRANSFER. min 𝐽 𝑿 𝑘 , 𝑼 ̅ 𝑘 such that: 𝒙 𝑡 + 1 = 𝒇 𝒙 𝑡 , 𝒖 𝑡 , 𝑡 , input constraints in Eq. (9) hold. 𝑼 ̅ 𝑘 𝑿 ̂𝑘 𝑼 ̅ 𝑘 1 = + 1 Transfer one sample/device: 𝑿 ̂𝑘 = 𝑭 𝑼 ̅ 𝑘 . For 𝑡 = 1, 2, … , 𝑁: R2R Transfer Control law Eq. (8) Motors 𝒘 𝑡 𝒙 ̂𝑘 𝑡 𝑇 𝑆 𝒖 𝑘 𝑡 𝒖 ̅ 𝑘 𝑡 Controller with saturation Discrete-time system Improve feedforward traj.: 𝑼 ̅ 0 , = *Corresponding author. Email: cbmartin129@uexas.edu 4 MECC25_39, Martin where Δ𝒙 𝑘 (𝑡) and Δ𝒖 ̅ 𝑘 (𝑡) are the update directions, Δ𝑿 𝑘 = [Δ𝒙 𝑘 (1), Δ𝒙 𝑘 (2), … , Δ𝒙 𝑘 (𝑁)], Δ𝒙 𝑘 (1) = , Δ𝑼 ̅ 𝑘 = [Δ𝒖 ̅ 𝑘 (1), Δ𝒖 ̅ 𝑘 (2), … , Δ𝒖 ̅ 𝑘 (𝑁 -1)], Δ𝒖 ̅ 𝑘 ( ) = , and 𝒖 𝑘 𝑔 ( ) = 𝒖 0 𝑔 ( ) is a user-defined constant. Finally, the feedforward trajectory and guess optimal trajectory are updated as follows (Step 4): 𝑿 𝑘 1 𝑔 ← 𝑿 𝑘 𝑔 + 𝛼 𝑘 Δ𝑿 𝑘 𝑔 (14. 𝐴) 𝑼 ̅ 𝑘 1 ← 𝑼 ̅ 𝑘 + 𝛼 𝑘 Δ𝑼 ̅ 𝑘 (14. 𝐵) 𝑼 𝑘 1 𝑔 ← 𝑼 ̅ 𝑘 1 + 𝑲(𝑿 𝑘 1 𝑔 -𝑿 𝑟 ) (14. 𝐶) where 𝑿 𝑟 = [𝒙 𝑟 (1), 𝒙 𝑟 (2), … , 𝒙 𝑟 (𝑁)], 𝑿 𝑘 𝑔 = [𝒙 𝑘 𝑔 (1), 𝒙 𝑘 𝑔 (2), … , 𝒙 𝑘 𝑔 (𝑁)], 𝑼 𝑘 𝑔 = [𝒖 𝑘 𝑔 (1), 𝒖 𝑘 𝑔 (2), … , 𝒖 𝑘 𝑔 (𝑁 -1)], and 𝛼 𝑘 is the step size, given by: 𝛼 𝑘 = 𝜂 -𝑐 (15) where 𝜂 and 𝑐 are user-defined. Typically, 𝑐 ∈ [ ,1] and 𝜂 ∈ ( ,1]. This SQP ILC framework is summarized in Algorithm 1. Algorithm 1: SQP ILC Input: 𝑿 0 𝑔 , 𝑼 0 𝑔 , 𝑼 ̅ 0 while termination criteria not met do 1. Measurement: 𝑿 ̂𝑘 ← 𝑭(𝑼 ̅ 𝑘 ) 2. Linearize: 𝑨 𝑡,𝑘 , 𝑩 𝑡,𝑘 , 𝒅 𝑡,𝑘 ← Expand around 𝒙 𝑘 𝑔 (𝑡), 𝒖 𝑘 𝑔 (𝑡) 3. Optimal control: Δ𝑿 𝑘 , Δ𝑼 ̅ 𝑘 ← Solution of QP (13) 4. Update: 𝑿 𝑘 1 𝑔 ← 𝑿 𝑘 𝑔 + 𝛼 𝑘 Δ𝑿 𝑘 ; 𝑼 ̅ 𝑘 1 ← 𝑼 ̅ 𝑘 + 𝛼 𝑘 Δ𝑼 ̅ 𝑘 ; 𝑼 𝑘 1 𝑔 ← 𝑼 ̅ 𝑘 1 + 𝑲(𝑿 𝑘 1 𝑔 -𝑿 𝑟 ) ← + 1
The SQP ILC framework for R2R mechanical dry transfer is validated on an experimental testbed, shown in Figure 5. The experiment involves peeling Trimaco masking tape, chosen for its suitable adhesion properties, from a PET transfer web. R2R peeling of an adhesive strip shares key characteristics with R2R mechanical dry transfer, including nonlinear peeling dynamics and motor torque constraints [14]. Therefore, this experiment serves as a practical test of SQP ILC's feasibility and effectiveness for R2R mechanical dry transfer. The physical parameters of this R2R peeling testbed are given in Table 2. All system states were measured online using loadcells and encoders, and the motor torques were determined using state feedback, as described in Eq. ( 8).
The testing process followed the sequence outlined in Algorithm 1. Each th iteration began with measuring the system's response to a step increase in reference tension, a standard approach for evaluating control regulation performance. State measurements from the th peeling task were then used to update the control parameters according to Eqs. ( 12)- (15). The new feedforward control was then applied in the ( + 1) th peeling task. Thus, the experiment aligned with the proposed ILC flow for an industrial R2R line: each peeling task generated new measurements, which were used to improve the control policy before the next iteration.
The feedback gain 𝑲 from Eq. ( 8) was chosen to be a linear quadratic regulator (LQR) [14], while the initial feedforward trajectory was derived from a linearized system model. The first peeling task iteration, which used this LQR feedback and naïve feedforward trajectory, served as a benchmark for evaluating subsequent iterations. The 𝑸, 𝑹, and 𝑺 matrices in the cost function (13.A) were determined using Bryson's rule [23], where 𝑸 𝑖,𝑖 ~𝑥𝑖 -2 , 𝑹 𝑖,𝑖 ~𝑢𝑖 -2 , and 𝑺 𝑖,𝑖 ~𝑢𝑖 -2 . Additional weight was given to the elements of 𝑸 associated with 𝑇 As can be seen in Figure 7, the tension RMSE in the first iteration, using LQR feedback and a naïve feedforward trajectory, is 2.36 N. After 20 iterations of the peeling task within the SQP ILC framework, this RMSE converges to 0.67 N, a 1.69 N or 71% improvement. Notably, 82% of this improvement comes in the first 5 iterations. These results demonstrate that the *Corresponding author. Email: cbmartin129@uexas.edu 5 MECC25_39, Martin SQP ILC method is both implementable and useful for R2R mechanical peeling dry transfer.
With the practicality and effectiveness of the proposed method demonstrated experimentally, its applicability to advanced manufacturing is further evaluated. Specifically, SQP ILC is utilized to control the R2R mechanical dry transfer of chemical vapor deposition (CVD)-grown graphene, where the donor is the copper growth substrate [7], [8]. This R2R transfer setup, illustrated in Figure 8, consists of regularly spaced graphene-copper samples sandwiched between two PET carrier films. The adhesion and bending properties, parameterized by 𝑈, of the PET-PET and graphene-copper interfaces are also detailed in Figure 8. Additional physical parameters, sourced from previous experiments [20], [21], are listed in Table 3. To ensure minimal damage to the transferred graphene, the objective is to maintain an optimal web tension setpoint during peeling despite abrupt changes in 𝑈 when the peeling interface changes.
As a comparison controller, the nonlinear MPC scheme from [11] was also used on this task. While this MPC achieves high performance without requiring iterative learning, it must solve a nonlinear optimization over a prediction horizon at each timestep, resulting in drastically higher online computational cost than SQP ILC, which only computes state feedback. N-m] T for both controllers. For the proposed SQP ILC scheme, 𝜂 = 1, 𝑐 = 0.5, and 𝑁 = 60. Additionally, 𝑲 will once again be synthesized using an LQR approach. For the nonlinear MPC, the prediction horizon, 𝑁 𝐻 , was 20 timesteps.
A total of 100 graphene transfer tasks were simulated, corresponding to 100 iterations of the SQP ILC framework. The R2R mechanical dry transfer model, used for these simulations, was developed in [20]. Figure 9 gives a representative selection of the web tension and control input trajectories of the SQP ILC method as well as the trajectory of the nonlinear MPC for comparison. Note that the nonlinear MPC is only run for one graphene transfer task, since it does not update from iteration to iteration. Figure 10 gives the average tension RMSE signals for each of the 100 iterations controlled using SQP ILC.
As can be seen in Figure 9, the tension RMSE in the first iteration, using LQR feedback with a naïve feedforward trajectory, is 2.24 N. After 100 iterations of the transfer task within the SQP ILC framework, this RMSE converges to 0.411 N, a 1.83 N or 82% improvement. Notably, 78% of this performance improvement comes in the first 10 iterations. This reduction in tension RMSE corresponds to an estimated 1.5 kΩ decrease in electrical sheet resistance, a key metric of graphene transfer quality [7]. These results suggest that the SQP ILC method can significantly improve R2R mechanical dry transfer quality for advanced applications like CVD graphene.
Next, the performance of the SQP ILC scheme on this task is compared to that of the nonlinear MPC. The average tension RMSE of the MPC was 0.415 N. The SQP ILC surpassed this performance after 93 iterations, got within 20% of this performance after 25 iterations, and got within 50% of this performance after 10 iterations. Importantly however, a key advantage of SQP ILC is its minimal computational burden, requiring only a single "offline" optimization step between graphene transfer tasks, as defined in Eqs. ( 12)-( 15). This computation converged within a few seconds on a Dell G7 7500 (Intel i7-10750 CPU, 16GB RAM), making it feasible for highspeed industrial R2R processes with throughputs of 10-100 devices per minute. In contrast, nonlinear MPC must solve a nonconvex optimization problem at every timestep of each peeling task, making it computationally impractical for most real-world industrial applications. A quantitative summary of this performance comparison is presented in Table 4. Tape PVA Graphene Copper Tape PET PET-PET Graphene-Copper 𝑈 = 6.5 J m -1 Graphene-Copper 𝑈 = 10 J m -1 PET-PET 𝑈 for each peeling section: Sample Sample +1 FIGURE 8: SELECTED TRAJECTORIES OF THE R2R MECHANICAL DRY TRANSFER OF CVD GRAPHENE SIMULATIONS. The grey shaded region corresponds to the time segment when the graphene sample is being peeled. *Corresponding author. Email: cbmartin129@uexas.edu 6 MECC25_39, Martin
This study presents an SQP ILC method for controlling R2R mechanical dry transfer. This method achieves excellent performance after about 5-10 iterations by accounting for the nonlinear dynamics, input constraints, and measurements from previous transfer tasks. Moreover, since the control policy updates are calculated between iterations of the transfer task, the method has minimal online computation. Experimental results show that the proposed method is both practical and effective for a lab-scale R2R transfer task. Over 20 iterations, the tension RMSE was reduced by 71%, with most of the improvement occurring in the first several iterations. In addition, a case study indicates that the method can significantly improve the transfer quality of CVD graphene relative to a baseline optimal controller, and that it can achieve equivalent regulation performance compared to an advanced nonlinear MPC with essentially no online computation cost.