Machine learning framework for prediction of plasma-based carbon dioxide conversion: Balancing computational efficiency with experimental efforts

• ML framework integrating literature and local data for CO 2 splitting prediction. • Hybrid model fine-tuned via random search/active learning achieves high accuracy. • Required experiments reduces by up to 38% using only 45–55 local data points. • Optimal experiment-to-literature data ratio (0.52–0.63) balances data efficiency. Plasma-based carbon dioxide (CO 2 ) conversion often involves interacting parameters and remains resource-intensive for experimental research. Machine learning (ML) presents a potential solution, but is limited by the scarcity of datasets. In this work, we develop an ML framework that integrates literature data with targeted local experiments to predict CO 2 conversion and energy efficiency in a dielectric barrier discharge (DBD) reactor. An artificial neural network (ANN) pre-trained on literature data showed good general alignment but poor accuracy for local experimental data. Subsequently, we iteratively refined a hybrid ANN model, by incorporating our local experimental data through via random search (RS) and active learning (AL). Although the RS-based hybrid model only selected 45 experimental data to achieve high predictive accuracy (R 2 > 0.95), thus reducing the planned experiments by 38%, AL-based model demonstrated superior robustness, achieving comparable accuracy with 55 data points and producing unbiased predictions. Such hybrid models also quantify the optimal ratio of adding experiment-to-literature data (0.52 – 0.63) to balance leveraging broad literature patterns and adapting to specific experimental features. This work provides new insights from coordinated datasets for balancing computational efficiency with experimental efforts, advancing ML-driven optimization in plasma-based CO 2 conversion.