Context-Aware Feature Importance and Rule Filtering for Predicting Power Consumption and SNN Accuracy in Nano-Structured Metal Oxide Thin-Film Transistor Devices Vita Efelina1,2, Endah Purwanti3, Chaerur Rozikin4, Riza Ibnu Adam4*
1 Department of Physics, Faculty of Engineering, Universitas Singaperbangsa Karawang, Telukjambe Timur, Karawang 41361, Indonesia
2Doctoral Program in Physics, Faculty of Mathematics and Natural Sciences, Institut Teknologi Bandung, Jl. Ganesha 10, Bandung 40132, West Java, Indonesia
3Department of Chemistry, Faculty of Engineering, Universitas Singaperbangsa Karawang, Telukjambe Timur, Karawang 41361, Indonesia
4Department of Informatics, Faculty of Computer Science, Universitas Singaperbangsa Karawang, Telukjambe Timur, Karawang 41361, Indonesia
Abstract
Nanostructured metal oxide thin-film transistor (TFT) circuits are promising electronic devices for low-power neuromorphic hardware because they can enable efficient spike-based computation and stable operation of spiking neural networks (SNNs). However, TFT datasets often contain many heterogeneous features, which can increase model complexity and reduce interpretability. This study aims to develop predictive models for power consumption and SNN accuracy on TFT devices using context-aware feature importance and rule filtering. This research was conducted in three stages. First, a baseline model was developed using Random Forest, XGBoost, Gradient Boosting, Support Vector Regression, and K Nearest Neighbor Regression, and evaluated using MAE, RMSE, R2, and computation time. Second, context-aware feature importance was used to identify the dominant features for SNN power consumption and accuracy, as well as their joint contributions to both targets. Third, the selected features were used for rule filtering and remodeling to obtain a more concise and interpretable model. The results show that Random Forest provides the most stable performance, achieving the lowest RMSE for power consumption of 2.8741 and the best baseline performance for SNN accuracy with MAE of 6.8251 and RMSE of 8.0850. The analysis identifies cycles to failure, mobility, gate length, annealing temperature, temperature stability, and nanostructure size as influential parameters. These findings suggest that context-sensitive feature importance and rule filtering can support efficient and easily interpretable TFT design analysis for low-power neuromorphic computing applications.
Keywords: thin film transistor, power consumption, SNN accuracy, context aware feature importance, rule filtering.
