Journal article
Enhanced iCVD reactor optimization through convexified attention networks: improving liquid repellency in fluoropolymers
Machine Learning: Engineering, v 1(1), 015009
25 Nov 2025
Abstract
Initiated chemical vapor deposition (iCVD) reactors offer a versatile, solvent-free platform for the precise fabrication of ultrathin polymer coatings through surface-initiated polymerization. Our previous work demonstrated that Convexified Convolutional Neural Networks (CCNNs) effectively model and optimize iCVD reactor parameters to maximize liquid repellency in fluoropolymers. However, standard CCNNs lack adaptive feature prioritization, limiting their performance on sparse industrial datasets. This paper introduces CCNNs with Convexified Attention Mechanism (CCNNCA), which dynamically weights input features while preserving convexity guarantees. Using the same dataset of 49 experimental batches from our AIChE Journal publication, CCNNCA achieves significantly improved prediction accuracy (R2 = 0.95) compared to previous CNN (R2 = 0.83–0.89) and CCNN (R2 = 0.92) models. The framework identifies promising process conditions that are predicted to enhance liquid repellency, with a 75.4% reduction in prediction error compared to prior methods. Through visualized attention weights, the model provides interpretable feature importance, revealing the critical roles of initiator flow rate and reactor temperature. This methodological advancement offers a more accurate predictive framework to guide future experimental optimization in data-limited, experiment-intensive engineering applications.
Metrics
1 Record Views
Details
- Title
- Enhanced iCVD reactor optimization through convexified attention networks: improving liquid repellency in fluoropolymers
- Creators
- Daniel Schwartz - Drexel UniversityAtieh Armin - Drexel UniversityAli Shokoufandeh - Drexel UniversityMasoud Soroush (Corresponding Author) - Drexel University
- Publication Details
- Machine Learning: Engineering, v 1(1), 015009
- Publisher
- IOP Publishing
- Number of pages
- 17
- Grant note
- CBET-1953176; CMMI-2132141 / National Science Foundation
- Resource Type
- Journal article
- Language
- English
- Academic Unit
- Computer Science (Computing); Chemical and Biological Engineering
- Other Identifier
- 991022145534304721