(124c) Multimodal Machine Learning for Predictive Structural Characterization of Plant-Based Meat Products in Food Extrusion Processes

Plant-based meat analogs are receiving growing attention as sustainable alternatives to conventional animal proteins. Among available processing techniques, extrusion is particularly effective in producing fibrous, meat-like textures from plant proteins. This thermomechanical process involves complex transformations driven by heat and mechanical energy, including protein denaturation, molecular alignment, and aggregation [1,2]. Modeling such processes remains challenging due to the complex thermo-physical transitions in proteins under high-pressure, high-shear conditions. These transitions critically influence product quality but hinder accurate prediction and control of extrusion outcomes [3,4]. Prior efforts have employed both mechanistic and data-driven models, including machine learning (ML), to simulate extrusion process behavior and capture rheological dynamics [5-8]. However, these models often rely on limited real-time data and struggle to directly link process variables to final product characteristics.

This study introduces a novel multimodal process modeling framework developed using comprehensive data from a pilot-scale co-rotating twin-screw extruder, producing plant-based meat analogs. The dataset reflects the inherent complexity of commercial production, capturing variations in material properties and process thermomechanical response. High-resolution time series data were collected across diverse operating conditions, including motor torque, screw speed, and zone-specific barrel temperatures. Experiments were conducted under different processing regimes (extruder screw speed and in-barrel moisture) and two plant protein-based feed compositions representative of commercial formulations. In parallel, extrudate images were captured at two-minute intervals to monitor the evolving surface and internal structural characteristics of the product. These images provide structure-relevant visual information—such as fibrousness and porosity—that cannot be directly inferred from process variables alone. The numerical and visual datasets offer a rich, time-aligned basis for developing a predictive multimodal model to capture operational behavior and end-product quality dynamics.

The proposed architecture consists of two integrated components. The first is a computer vision (CV) pipeline that processes the extrudate image sequence to quantify structural features. Using edge detection and surface analysis techniques, the CV module extracts interpretable metrics related to features such as fibrousness, which directly impact key textural attributes for consumer perception of meat analogs. These image-derived metrics form continuous profiles of product structure over time. The second component is a time series machine learning ensemble that combines boosting and bagging regressors to model the nonlinear relationships between process variables, intermediate outputs, and CV-derived structural scores (primarily fibrosity). By integrating visual and numerical modalities, the framework captures complex interactions not evident in either data source alone. This multimodal modeling approach enables real-time prediction of structure outcomes based on evolving process conditions, facilitating in-process adjustments to achieve targeted product quality. The framework offers a path toward structure-informed control and optimization in extrusion systems, supporting the development of intelligent, adaptive food manufacturing processes capable of consistently delivering high-quality, consumer-aligned plant-based meat products.

References:

[1] Flory, J., Alavi, S., Use of hydration properties of proteins to understand their functionality and tailor texture of extruded plant-based meat analogues. Journal of Food Science 89(1):1–14, 2023

[2] Plattner, B.J., Hong, S., Li, Y., Talavera, M.J., Dogan, H., Plattner, B.S., Alavi, S., Use of pea proteins in high-moisture meat analogs: Physicochemical properties of raw formulations and their texturization using extrusion. Foods 13:1195, 2024

[3] McGuire, C., Siliveru, K., Chakraborty, S., Ambrose, K., Alavi, S., Flow properties of coarse powders used in food extrusion as a function of moisture content. Processes 12:1246, 2024

[4] Gaspar-Cunha, A., Monaco, F., Sikora, J., Delbem, A., Artificial intelligence in single screw polymer extrusion: Learning from computational data. Engineering Applications of Artificial Intelligence 116:105397, 2022

[5] Khan, M.I.H., Sablani, S.S., Nayak, R., Gu, Y., Machine learning-based modeling in food processing applications: State of the art. Comprehensive Reviews in Food Science and Food Safety 21:1409–1438, 2022

[6] Bhagya Raj, G., Dash, K.K., Comprehensive study on applications of artificial neural network in food process modeling. Critical Reviews in Food Science and Nutrition 62:2756–2783, 2022

[7] Dahl, J.F., Schlangen, M., van der Goot, A.J., Corredig, M., Predicting rheological parameters of food biopolymer mixtures using machine learning. Food Hydrocolloids, 160:110786, 2025

[8] Kircali Ata, S., Shi, J.K., Yao, X., Hua, X.Y., Haldar, S., Chiang, J.H., Wu, M., Predicting the textural properties of plant-based meat analogs with machine learning. Foods 12:344, 2023