Abstract
Integrating equipment data with computational models supports decision-making in industry, but reliable decisions require models that remain valid as operating conditions change. When process responses deviate from the nominal relationship, rebuilding the model for each new condition requires substantial additional measurements, making efficient model updating with limited online data a critical challenge. Digital twin technology provides a promising paradigm for addressing this problem by integrating online model updating within a physical–digital loop. This study develops an adaptive digital twin system with small-sample model updating and validates it in industrial inkjet manufacturing. In this application, droplet volume is treated as the process response and driving-waveform parameters as decision variables. A digital twin-driven similarity-blended residual adaptation (DT-SBRA) method is proposed as the model-updating mechanism, using only a few measurements from the current condition to preserve the nominal mapping while correcting condition-dependent prediction deviations. The updated model is then reused for virtual process evaluation and waveform-parameter optimization. Validation under actual drift conditions showed that droplet-volume prediction accuracy exceeded 96% and target-volume control accuracy exceeded 95% over the range of 3.0–5.0 pL. Data-transmission latency remained below 50 ms, and defect-free deposition on organic light-emitting diode (OLED) substrates further demonstrated practical applicability. The results indicate that integrating small-sample model updating into the physical–digital loop can sustain model-based process evaluation and parameter feedback under operating-condition drift while reducing additional data acquisition and physical experimentation.

This publication has 31 references indexed in Scilit: