Digital Twin of Aerosol Jet Printing

TL;DR

This work presents a physics-informed digital twin for Aerosol Jet printing that fuses a macro-model of the AJ process with probabilistic state estimation and computer-vision-based measurements. The latent-state vector $x=[\bar{d}_a, V_l, \Delta r_T, \Delta r_N, \phi_A]$ evolves under transition dynamics $x_{k+1}=f_d(x_k,u_k)+\xi_k$, while outputs $y=[L_w, L_o, P_c, P_s, Q_m]$ are linked via $y_k=g_d(x_k,u_k)+w_k$, with inputs $u=[I_A, Q_c, Q_s]$. An EKF-based state estimator coupled with an EM-based parameter updater continually aligns the digital twin to real-time sensor and video data, enabling drift detection, root-cause analysis, and improved prediction of line quality under changing operating conditions. Experimental validation across three studies demonstrates accurate latent-state inference, reveals drivers of process drift (notably $\bar{d}_a$ and $\phi_A$), and shows that updating transition-parameters via EM enhances open-loop prediction, supporting future closed-loop control and broader applicability to advanced manufacturing. The framework's ISO-aligned, modular architecture supports extension to varied inks, substrates, and process configurations while offering actionable insights for reliability and scale-up of AJ-based manufacturing.

Abstract

Aerosol Jet (AJ) printing is a versatile additive manufacturing technique capable of producing high-resolution interconnects on both 2D and 3D substrates. The AJ process is complex and dynamic with many hidden and unobservable states that influence the machine performance, including aerosol particle diameter, aerosol carrier density, vial level, and ink deposition in the tube and nozzle. Despite its promising potential, the widespread adoption of AJ printing is limited by inconsistencies in print quality that often stem from variability in these hidden states. To address these challenges, we develop a digital twin model of the AJ process that offers real-time insights into the machine's operations. The digital twin is built around a physics-based macro-model created through simulation and experimentation. The states and parameters of the digital model are continuously updated using probabilistic sequential estimation techniques to closely align with real-time measurements extracted from the AJ system's sensor and video data. The result is a digital model of the AJ process that continuously evolves over a physical machine's lifecycle. The digital twin enables accurate monitoring of unobservable physical characteristics, detects and predicts anomalous behavior, and forecasts the effect of control adjustments. This work presents a comprehensive end-to-end digital twin framework that integrates customized computer vision techniques, physics-based macro-modeling, and advanced probabilistic estimation methods to construct an evolving digital representation of the AJ equipment and process. While the methodologies are customized for aerosol jet printing, the process for constructing the digital twin can be applied for other advanced manufacturing techniques.

Figures (44)

• Figure 1: Example of drift that can occur in linewidth and overspray over the course of a single print.
• Figure 2: The aerosol jet printer (a) initially aerosolizes an ink solution containing nanoparticles or polymer in the ultrasonic atomizer. The aerosol droplets are carried to the machine via the carrier gas flow and through the transport tube in (b). The flow of aerosol droplets are focused through a sheath gas flow in (c), and then deposited onto a moving platen with substrate in (d). The substrate is typically heated in order to evaporate the remaining solvent after droplet impact, leaving behind a profile of deposited material.
• Figure 3: The digital twin framework, inspired by the ISO 23247 standard iso2020automation, provides a structured approach to developing algorithms that update and control the AJ printer. The present effort developed five modules: data collection and feature extraction module, the state-based model, state estimation module, state parameter estimation module, and the simulation engine. The digital twin supports applications such as anomaly detection, process tolerance control and process optimization. Control commands are left for future work.
• Figure 4: a) Optical image of printed silver line and b) surface profilometer scan, c) frame taken from alignment camera inspection video of print surface, d) cross section of the line shows the grayscale value as a function of position along the surface, i corresponds to center of line, ii are the points of maximum slope on either side, and iii is where the linewidth is defined.
• Figure 5: a) Annotated frame from alignment camera video, green is Hough transform fit, red is linewidth, blue is overspray. Averaging window size in this example is 50. Frames shown from t = 0 and t = 90 minutes. b) Plotting linewidth and c) overspray over time (and distance) for entire inspection video over wafer in comparison to optical profilometer results.