Penggunaan Deep Learning dengan Metode Long Short-Term Memory untuk Prediksi Sifat Mekanik Komposit Resin DLP Berpenguat Limbah Fotovoltaik dan Glass Powder
IKKO YUSWANDA, Muhammad Akhsin Muflikhun, S.T., MSME., Ph.D.; Yi-Chieh Wu, Ph.D.
2026 | Tesis | S2 Teknik Mesin
The exponential accumulation of photovoltaic (PV) waste necessitates sustainable recycling strategies that transcend conventional disposal methods through innovative circular approaches. This research pioneers the development of hazardous PV waste-based Digital Light Processing (DLP) composite materials integrated with Deep Learning, bridging the gap between material valorization and intelligent characterization. By repurposing glass powder from PV waste as a functional filler, this study explores the mechanical and physical potential of reinforced photosensitive resins, offering a proactive solution to the environmental impact of increasing electronic waste.
Experimental characterization reveals a critical trade-off between physical density and curing maturity, significantly influenced by particle size distribution. Although the use of fine particles (-600 mesh) yields superior packing density, this fraction triggers light scattering that inhibits comprehensive photopolymerization. Conversely, coarser particles (-200/+400 mesh) facilitate more optimal UV light transmission, resulting in peak mechanical performance with a tensile strength of 27.77 MPa, flexural strength of 33.75 MPa, and surface hardness of 65.00 Shore D. These findings are corroborated by Scanning Electron Microscopy (SEM) microstructure analysis, which identifies failure mechanisms such as micro-voids and fiber pull-out, and are further validated through Finite Element Method (FEM) simulations and Digital Image Correlation (DIC) optical strain measurements.
To overcome the cost and time constraints of extensive destructive testing, this study integrates Long Short-Term Memory (LSTM) architectures to computationally predict the non-linear stress-strain behavior of the material. Data analysis confirms that utilizing specific datasets with the Adamax optimization algorithm yields the highest predictive accuracy. The optimized model architecture (1F, 3B, and 2C configurations) demonstrates highly precise predictions with a coefficient of determination (R²) exceeding 0.9. By merging material sustainability with Industry 4.0 technologies, this research presents a high-precision predictive tool supporting digital twin applications in intelligent and sustainable additive manufacturing.
Kata Kunci : Limbah fotovoltaik, Glass powder, Komposit, Deep Learning, Long Short-Term Memory