Tsai, Chih-YunChih-YunTsaiChang, Yu-TangYu-TangChangHung, Shu-PingShu-PingHungLu, Chun-MingChun-MingLuChen, Shih-FangShih-FangChenPeng, Chia-HungChia-HungPeng2026-03-312026-03-312025https://www.scopus.com/record/display.uri?eid=2-s2.0-105015444356&origin=resultslisthttps://scholars.lib.ntu.edu.tw/handle/123456789/7368982025 American Society of Agricultural and Biological Engineers Annual International Meeting, ASABE 2025, Toronto, Ontario, Canada, 13 July 2025 - 16 July 2025Coffee blending is a key strategy in coffee production, often employed to address unstable quality caused by environmental variability and diverse processing methods. It enables producers to ensure consistent quality while offering opportunities for product customization. However, the process is complicated by unpredictable flavor interactions and the vast number of potential ratios. This study introduces a deep learning-based recommendation system designed to optimize coffee blending by predicting sensory attributes and recommending blending ratios that meet user-defined objectives. The system leverages near-infrared (NIR) spectra from pre-blend samples, along with sensory data such as scoring criteria (rated 6-10) and flavor intensity (scored 0-5), to predict coffee sensory attributes effectively. A dataset of 530 coffee samples was collected, encompassing both pre-blend and post-blend data, with NIR spectra measured from 700 to 2500 nm at a resolution of 2 nm. An additional 32 blended samples were used to assess the ranking performance of the system for blending optimization. The deep learning model employs a transformer-based architecture, chosen for its scalability and ability to handle varying sample inputs effectively. Preliminary results demonstrated that the model achieved a mean absolute error (MAE) of 0.110 for scoring criteria and 0.768 for flavor intensity predictions, highlighting its potential to predict sensory changes after blending. In the ranking test, which evaluates the consistency between predicted and actual orderings using Kendall rank correlation, the system achieved a coefficient of 0.255, reflecting moderate ranking performance and demonstrating its potential as a tool for the coffee industry. The proposed recommendation system shows promise for supporting more data-driven blending decisions and enabling targeted product development to meet evolving consumer preferences and address challenges posed by climate change. Future work will focus on expanding the dataset to include a wider variety of blending ratios and sensory attributes, as well as refining the model performance. These efforts aim to further enhance the system's applicability across diverse coffee production scenarios.falsecoffee blending optimizationDeep learningnear-infrared spectroscopyrecommendation systemsensory predictionconference paper10.13031/aim.2025006112-s2.0-105015444356