Machine learning cuts trial-and-error from thick-film resistor manufacturing, holding resistance deviation within 5%

Researchers developed a data-driven method that combines a GA-BP neural network with a chaotic particle swarm optimization algorithm to predict and optimize screen-printing parameters for thick-film resistors. The approach predicts resistance with an R² of 0.991, recommends optimal process settings in about 2.38 seconds, and cuts resistance error by up to 14.86%, keeping deviations within 5%.

Thick-film resistors are essential passive components in hybrid integrated circuits, radio-frequency and microwave modules, and power electronics, where the accuracy and consistency of their resistance directly determine circuit gain, noise performance, and overall reliability. Yet manufacturing these resistors with precise resistance values remains a persistent challenge. The screen-printing stage that forms each resistor is governed by many tightly coupled parameters—printing pressure, gap, speed, mesh count, stencil thickness, and more—that interact in strongly nonlinear ways. Traditionally, engineers tune these settings through expert experience and repeated trial-and-error, a process that drives up cost and slows production while still struggling to hit target resistance values reliably.

A research team led by Wenhua Gu at Nanjing University of Science and Technology, publishing in Interdiscipline, set out to replace this experience-driven approach with a data-driven one. Their method pairs two complementary tools. First, a backpropagation (BP) neural network is trained to capture the complex nonlinear relationship between screen-printing parameters and final resistance. Because BP networks are sensitive to their starting conditions and can settle into suboptimal solutions, the researchers used a genetic algorithm (GA) to globally optimize the network's initial weights and thresholds—producing a GA-BP model with markedly better accuracy and stability.

The second tool addresses the reverse problem: given a desired resistance, what process settings should the factory use? Here the team applied a chaotic particle swarm optimization (CPSO) algorithm, using the trained GA-BP model as its fitness function to search the full parameter space. By injecting a chaotic mapping mechanism to keep the particle swarm diverse, CPSO avoids premature convergence and locates parameter combinations that hit the target resistance.

Built and tested on 4,800 datasets drawn from a real production line, the GA-BP model achieved an R² of 0.991 and a mean absolute percentage error of just 3.09% on the test set—outperforming conventional BP, response surface methodology, and linear regression models. When CPSO was used to recommend settings for a 115-ohm target, all optimized parameter groups landed within 5% of the goal, with the best deviating by only 0.459%, compared with a 15.32% deviation for the standard empirical settings. The method reduced resistance error by up to 14.86% and produced recommendations in roughly 2.38 seconds per target. The team further validated the approach at 120, 130, and 140-ohm targets, with measured resistances tracking predictions within ±2.1%.

A feature-importance analysis revealed that height compensation and stencil thickness exert the greatest influence on final resistance—a result that aligns with shop-floor experience and lends physical credibility to the model. By turning a slow, intuition-based tuning process into a fast, automated one, the work points toward smarter, more consistent manufacturing for printed-electronics components.

This paper “Data-driven optimization of screen-printing parameters for thick-film resistors using GA-BP prediction and CPSO search” was published in Interdiscipline.
Sun Y, Wang Y, Xue J, Zhang X, Liao X, et al. Data-driven optimization of screen-printing parameters for thick-film resistors using GA-BP prediction and CPSO search. Interdiscipline 2026(1):0001, https://doi.org/10.55092/interdiscipline20260001.