Optimasi Gemini 3 Flash Menggunakan Parameter Deterministik dan Prapemrosesan Citra sebagai Media Pembelajaran Fisika

Abstract

Physics education at the secondary school level is frequently constrained by the high abstraction of textbook materials, which decouples natural laws from students' surrounding empirical reality. The emergence of multimodal large language models offers interactive solutions, yet pure generative models are prone to producing inconsistent answers and excessively long scientific narratives that trigger cognitive fatigue in learners. This study aims to design, implement, and evaluate SmartFisika, a web-based AI Vision Assistant application that leverages Gemini 3 Flash to analyze photos of natural phenomena deterministically. The systems engineering method focuses on locking the model's degree of freedom parameters at a temperature value of 0.0 and a seed value of 42, combined with a digital image pre-processing module using a sharpening convolution matrix kernel at the backend level. The black-box testing results demonstrate that all functional components of the system operate successfully. Output consistency testing through repeated upload experiments of the same photo sample 10 times shows absolute textual similarity, with text explanation length profiles stably remaining below the cognitive upper limit of 1200 characters. The concise explanation structure, divided into four visual components, proves effective in facilitating students' scientific knowledge reconstruction. This study concludes that combining upstream image manipulation and strict language parameter control significantly dampens the stochastic nature of generative models while creating a reliable, consistent, and learner-friendly applied physics socialization medium.

Keywords: Gemini 3 Flash; Deterministic Parameters; Image Pre-processing; SmartFisika; Contextual Learning.