A deep learning-based quantitative prediction model for the processing potentials of soybeans as soymilk raw materials.

Current technologies as correlation analysis, regression analysis and classification model, exhibited various limitations in the evaluation of soybean possessing potentials, including single, vague evaluation and failure of quantitative prediction, and thereby hindering more efficient and profitable soymilk industry. To solve this problem, 54 soybean cultivars and their corresponding soymilks were subjected to chemical, textural, and sensory analyses to obtain the soybean physicochemical nature (PN) and the soymilk profit and quality attribute (PQA) datasets. A deep-learning based model was established to quantitatively predict PQA data using PN data. Through 45 rounds of training with the stochastic gradient descent optimization, 9 remaining pairs of PN and PQA data were used for model validation. Results suggested that the overall prediction performance of the model showed significant improvements through iterative training, and the trained model eventually reached satisfying predictions (|relative error| ≤ 20%, standard deviation of relative error ≤ 40%) on 78% key soymilk PQAs. Future model training using big data may facilitate better prediction on soymilk odor qualities.