I. Core Research Summary

• Establishing the causal correspondence between ink droplets and circuit defects.

The system categorizes typical ink‑droplet morphologies into six classes, corresponding to six types of circuit defects: broken traces, edge burrs, large‑area material overflow, and trace spreading, among others. It is demonstrated that airflow‑focusing imbalance is the primary cause of droplet‑morphology degradation: low carrier‑gas flow rates tend to produce thin droplets that break, while excessive sheath‑gas flow can lead to diffuse material overflow.

• Latin Hypercube Sampling (LHS) for Global Uniform Sampling

By abandoning conventional single-variable gradient experiments and uniformly spanning the entire parameter space of sheath gas and carrier gas, a small number of samples can fully capture nonlinear coupling patterns, thereby avoiding misjudgments arising from local optima.

• K-means unsupervised clustering for automatic partitioning

Automatic clustering is performed using droplet roughness, material overflow, and layer thickness as features, thereby partitioning the parameter space into high‑quality forming regions and subspaces corresponding to various defects, thus providing an intuitive delineation of unfavorable process windows.

• SVM classification locks the stable printing window.

The support vector machine generates the optimal process boundary, with model classification accuracy exceeding 94% for both nozzle specifications. Within the processing window, uniformly conductive traces with smooth edges and no overflow can be fabricated.

• GPR Gaussian Process Regression Quantitative Prediction Model

The model accurately captures the mapping between airflow, ink diameter, and film thickness, with an average relative error of less than 9%, enabling early prediction of droplet geometry under various airflow conditions.

• NSGA-III Multi-objective Pareto Optimization

It simultaneously balances the two conflicting objectives of “precise narrow linewidth” and “sufficient film thickness,” generating multiple sets of optimal process solutions to meet, on demand, the production requirements of both high‑density microcircuits and thick current‑carrying traces.

II. Complete Workflow for Aerosol + Machine Learning–Based Intelligent Optimization

(1) Complete Preparation Procedure

• Ultrasonic homogenization pretreatment of silver nanoparticle ink, followed by the construction of an AJ300 ultrasonic atomization printing platform;
• LHS generates Latin hypercube‑based experimental design points across the entire parameter space, and batch‑prints individual deposited ink droplets alongside standard conductive traces.
• Optical confocal scanning is used to extract four quantitative metrics: ink diameter, film thickness, edge roughness, and overflow.
• K-means clustering is used to classify ink droplet morphologies, while SVM is employed to determine the stable process operating window.
• GPR construction parameters—droplet size prediction model; NSGA-III multi-objective optimization yields Pareto-optimal process settings.
• Optimal parameters were used to fabricate a multilayer strain sensor, and the high-precision, narrow‑gap circuit‑forming performance was verified.

(II) Five Key Advantages of Intelligent Processes for Atmospheric Aerosol Systems (Compared with Traditional Manual Tuning)

• Optimize defects at the source, rather than relying solely on post‑production reverse debugging.

Focusing directly on the fundamental deposited ink droplets, we trace the root causes of ink overflow and filament breakage, refining the underlying logic for more substantial improvements in print quality.

• Global parameter coverage to avoid getting stuck in local optima.

LHS uniformly samples the parameter space, ensuring that even a small number of gradient-based experiments will not overlook high-quality airflow ratios.

• High‑precision modeling can be achieved with just a small number of experimental samples, significantly reducing costs associated with ink, substrates, and time.

No need for hundreds of iterative trial-and-error cycles—predictive modeling and optimization can be accomplished with just dozens of data sets, shortening the R&D cycle by more than 70%.

• Multi-objective balanced optimization, balancing fine-line accuracy with trace thickness.

Traditional debugging can only achieve a single, compromise‑based solution, whereas the NSGA algorithm generates multiple process‑plan alternatives, each tailored to the production requirements of different devices.

• The model boasts strong versatility and is compatible with a wide range of aerosol inks, including metal‑based and polymer‑based formulations.

The framework is independent of any specific ink; by simply replacing the conductive and dielectric pastes, only a small number of additional samples are needed to rapidly rebuild the model.

III. Five Key Technological Innovations

Innovation 1: A successful attempt to develop an AJP machine-learning optimization framework centered on the morphology of deposited ink droplets.

Innovation 2: A Comprehensive Classification System for Defects in the Morphology of Six Types of Ink Droplets

Innovation 3: End-to-End Algorithm Integration for Hybrid Machine Learning

Innovation 4: Pareto multi-objective optimization addressing the trade-offs between line width and film thickness.

Innovative 5-Stage Highly Versatile Aerosol Intelligent Modeling Paradigm

IV. Current Industry Challenges and Corresponding Solutions for Overcoming Them

Challenge 1: Strong nonlinear coupling between the sheath gas and the carrier gas, leading to blind trial-and-error approaches that are both material‑ and time‑intensive.

Challenge 2: Line defects can only be identified retrospectively and cannot be predicted at the source of deposition.

Challenge 3: The two major objectives—narrow, fine lines and high film thickness—are mutually conflicting, leaving manual processes with no choice but to make a single compromise.

Challenge 4: Traditional craftsmanship relies solely on qualitative visual assessment, lacking digital quantitative standards.

Challenge 5: After replacing the nozzle or ink, the entire experimental procedure must be repeated, resulting in high costs for process iteration.

V. Future Diversified Industrial Application Scenarios

1 Foldable-screen high-precision micro-wiring production line

2 Mass Production of MEMS Micro Strain and Pressure Sensors

3 High-Throughput R&D Platform for Laboratory Aerosol Materials

4 Roll-to-Roll Flexible Printed Electronics Mass-Production Line

5 Aerospace-shaped Micro-sensor Circuits

Figure 1. Basic operating principle of aerosol jet printing using an ultrasonic atomizer.

Table 1. Different Operating Conditions of the AJP

Figure 2. Definition and extraction of deposited droplet features. (a) Definition of the deposited droplet morphology; (b) Original droplet morphology; (c) Extracted droplet features after denoising; (d) Droplet morphology detected via overspray analysis; (e) Three-dimensional contour of the deposited droplet; (f) Cross-sectional contour of the deposited droplet.

Figure 3. Classification of deposited droplet morphologies under operating conditions of Case I: (a1–b1) Droplets with high overspray (Type I–II droplets); (c1–e1) Deposited droplets with high edge roughness (Types III–V droplets); (f1) Circular droplets with reduced overspray and more distinct edge profiles (Type VI droplet); (a2–f2) Three-dimensional topographies of the deposited droplets corresponding to (a1–f1).

Figure 4. Mass analysis of deposited droplets under the operating conditions of Case I. (a) Normalized droplet mass in the design space; (b) Clustering subspaces used to characterize distinct droplet morphologies: Subspace (I)—excessive overspray with insufficient thickness; Subspace (II)—excessive overspray but sufficient thickness; Subspace (III)—high edge roughness with insufficient thickness; Subspace (IV)—high edge roughness and sufficient thickness due to poor focusing; Subspace (V)—high edge roughness and sufficient thickness caused by excessive aerosol accumulation; Subspace (VI)—low overspray with well-defined edge profiles.

Figure 5. Classification of printed line morphology (at 1 mm/s) according to the droplet types defined in Figure 3: (a1) printed lines exhibiting discontinuities (Type I lines); (b1) printed lines with excessive overspray (Type II lines); (c1) printed lines with insufficient thickness and high edge roughness (Type III lines); (d1) printed lines with adequate thickness and high edge roughness (Type IV lines); (e1) printed lines undergoing spreading (Type V lines); (f1) printed lines with reduced overspray and more sharply defined edges (Type VI lines); and (a2–f2) cross-sectional profiles corresponding to each of the types (a1–f1).

Table 2. Detailed Information on the Measured Cross-Sectional Profile

Figure 6. Experimental validation results based on the printed three-layer strain gauge. (a–c) Morphological features of the bottom layer printed under an unbalanced focusing ratio; unnecessary overlaps (d, e) and gaps (f) occur in the soldering and sensing regions of the printed three-layer strain gauge, respectively. (g–i) An ideal strain gauge fabricated using a balanced focusing ratio, where both overlaps and gaps can be precisely controlled.

Figure 7. Shows the determined operating window for (a) Case I and (b) Case II, respectively.

Figure 8. A quantitative model has been established linking the principal process parameters to the geometric characteristics of the deposited droplets. (a, b) Droplet diameter as the output response; (c–d) Droplet thickness as the output response.

Figure 9. Optimization results of the AJP process under the operating conditions of Case I, based on the NSGA-III algorithm. (a, b) show the optimal Pareto fronts obtained for specified droplet diameters of 40 μm and 50 μm, respectively; (c, d) present the corresponding optimal trade-off points between the target droplet diameter (Obj1) and the maximum droplet thickness (Obj2), as depicted in (a, b).

Figure 10. Optimization results of the AJP process under Case II operating conditions, based on the NSGA-III algorithm. (a, b) show the Pareto frontiers obtained for specified droplet diameters of 140 μm and 160 μm, respectively; (c, d) present the optimal trade-off points between the target droplet diameter (Objective 1) and the maximum droplet thickness (Objective 2), as indicated in (a, b).

Table 3. Representative Solutions of the Optimization Results Under Case I Operating Conditions

Table 4. Representative Solutions of the Optimization Results under Case II Operating Conditions

Figure 11. Comparison of deposited droplet morphologies under the operating conditions of Case I: (a1–c1) deposited droplets obtained using the solutions listed in Table 3 at a specified droplet diameter of 40 μm; (a2–c2) deposited droplets obtained using the same solutions at a specified droplet diameter of 50 μm; (d1–d2) deposited droplets generated by the RSM‑SGD optimization algorithm at the specified droplet diameters of 40 μm and 50 μm, respectively.

Figure 12. Comparison of deposited droplet morphologies under Case II operating conditions: (a1–c1) Deposited droplets obtained using the solutions listed in Table 4 at a specified droplet diameter of 140 μm; (a2–c2) Deposited droplets obtained using the solutions listed in Table 4 at a specified droplet diameter of 160 μm; (d1–d2) Deposited droplets obtained, respectively, by the RSM‑SGD optimization algorithm at specified droplet diameters of 140 μm and 160 μm.