Recipient Organization
UNIV OF MARYLAND
(N/A)
COLLEGE PARK,MD 20742
Performing Department
Nutrition and Food Science
Non Technical Summary
Innovations in food packaging systems are urgently needed to meet the evolving market demands, such as consumer preference for "healthy" and high-quality food products and reduction of the negative environmental impacts of food packaging. For example, new ideas in intelligent/active packaging technologies provide a variety of creative ways to extend the shelf life of food products while also enhancing their quality and safety. New techniques have been developed to enhance the functional properties of biodegradable food packaging, such as its mechanical durability, barrier effectiveness, and thermal stability. Those intentions might be considered as the ultimate future goal for food packaging technology. However, the discovery of innovative packaging materials is often time-consuming and labor-intensive. Until now, innovative food packaging design has encountered critical challenges: (1) First, significant efforts and time are required to obtain an optimized fabrication recipe that leads to material with desired performance. 2) It is still not possible to predict the functional properties of a food package material from its composition because the recipe-structure-property correlations are too complicated to be simulated using state-of-the-art computational techniques.Because of these methodological limitations to finding materials with promising functional properties, we try to present a different yet effective approach to accelerating the discovery of innovative food packaging material. Recently, the application of artificial intelligence/active learning (AI/AL) to materials science research has attracted tremendous attention in areas such as organic/inorganic catalyst design, drug discovery, and quantum dot synthesis, where simulation tools or analytical tools can provide a significant amount of data. However, there are significant barriers to creating an AI/AL model with high prediction accuracy in the field of self-assembled materials for food packaging. These barriers are mostly caused by the dearth of high-quality data. Besides, recent reports focused on pursuing materials with either superior tensile strength/elongation/toughness or high optical transparency. Such unbalanced data are likely to produce an AI/AL model that can only predict single property, which would not benefit the programmable design with multi-property controls. For accurate AI/AL prediction at the self-assembly level, it is highly valuable to overcome these challenges using a hybrid technique (experimentally and computationally together), which can automate the discovery and design of all-natural food packaging material.In this proposal, our ultimate goal is to automate the discovery/design of a food packaging materials from all-natural ingredients through AI/AL prediction and robot-human teaming, facilitating the optimization process to meet the requirement of a variety of food products. First, we will build a prediction model via AI/ML frameworks, data augmentation, and flexible robotic technologies. Second, we will design more than 50 different innovative food package matrixes through simpler design approaches and with the usage of input all-natural ingredients. Finally, we will analyze data and apply molecular simulation tools to improve ML/AI interpretability. The successful completion of this study will provide scientific information that will support the design of food packaging materials and enhance food safety.
Animal Health Component
(N/A)
Research Effort Categories
Basic
50%
Applied
(N/A)
Developmental
50%
Goals / Objectives
In this proposal, our ultimate goal is to automate the discovery/design of a food packaging materials from all-natural ingredients through AI/AL prediction and robot-human teaming, facilitating the optimization process to meet the requirement of a variety of food products. First, we will build a prediction model via AI/ML frameworks, data augmentation, and flexible robotic technologies. Second, we will design more than 50 different innovative food package matrixes through simpler design approaches and with the usage of input all-natural ingredients. Finally, we will analyze data and apply molecular simulation tools to improve ML/AI interpretability. The successful completion of this study will provide scientific information that will support the design of food packaging materials and enhance food safety.The specific objectives are:Objective 1: Progressive construction of a prediction model via ML/AI frameworks, data augmentation, and flexible robotic technologies.Objective 2: Programmable design of all-natural food packaging materials.Objective 3: Evaluation of biocompatibility, biodegradability, and food safety impacts.Objective 4: Integration of antimicrobial materials into all-natural packaging materials.
Project Methods
Objective 1: Progressive construction of a prediction model via ML/AI frameworks, data augmentation, and flexible robotic technologies. First, a library of natural, food-grade, and generally recognized as safe (GRAS) materials (as listed in Table 1) is adopted to fabricate various all-natural food packaging with tunable properties.To initiate the model construction, several recipes will be randomly selected with varying influential parameters to initiate the active learning loop. For instance, we plan to (1) alter the loadings of each natural component, (2) adjust the concentrations of ionic/covalent cross-linkers, and (3) conduct hot pressing at different temperatures. Afterward, the physicochemical properties of all-natural food packaging will be characterized, including stress-strain profiles, absorption spectra, and oxygen permeability.The loadings of natural materials will serve as the "composition" labels, and the characterized properties (e.g., ultimate stress, fracture strain, optical transparency) will serve as the "property" labels. For each kind of all-natural packaging, we will bundle its "composition" and "property"labels into ONE datum. As a number of plastic substitutes will be tested in the initial round, multiple data will be collected and input into the algorithms of artificial neural network (ANN) and decision tree (DT) to train a leaner model. The leaner model will classify the design space and rank various regions based on their uncertainty and diversity. The most uncertain and diverse data will be suggested for the next active learning loop. In objective 1, we propose to perform >50 loops of active learning (60 samples per loop and 2 loops per week) to explore the multi-component design space and develop a champion prediction model with a total of >3,000 all-natural packaging fabricated. PI Chen has conducted similar active learning loops in his recent work focusing on the ML-enabled fabrication of soft robotic sensors.To accelerate data acquisition rates and ensure data quality, we will integrate multiple flexible robotic technologies (all available in PI Chen's lab) into the active learning loops. After the learner model suggests the "composition" labels, the automated pipetting robotwill prepare the designated mixtures, and a collaborative robotic arm (UR5e from Universal Robots) will bring the mixture in ovens for evaporative self-assembly. Once the all-natural packaging is assembled, its stress-strain profile and optical transparency can be characterized by using multiple automated testing platforms, including a tensile tester with a robotic arm (CT6 from Instron) and a high-throughput spectrophotometer.Objective 2: Programmable design of all-natural food packaging materials. The champion prediction model can perform multiple AI/ML prediction tasks that are unachievable in state-of-the-art works: (1) accurate prediction of multiple physicochemical properties of an all-natural food packaging from its fabrication recipe, (2) suggestion of optimal recipes to fabricate adequate all-natural packaging for specific food product. The prediction power of the champion model will be harnessed to program various all-natural packaging materials with customizable properties for diverse food protection. Through clustering analyses, the champion model can suggest a distinct recipe that leads to the packaging material with high transparency (>90%) and high tensile strength (>100 MPa) for different fruits (the bag can hold 2-lb apples). In addition, the model-suggested recipe can be easily scaled up through evaporative assembly. We aim to collect more data for the champion model to enhance its predictive capabilities, and our goal is to come up with 30 all-natural packaging materials and demonstrate their viable replacement within this year.Objective 3: Evaluation of biocompatibility, biodegradability, and food safety impacts. The biocompatibility of model-suggested all-natural packaging will be examined by conducting a series of cytotoxicity experiments on L929 cells, including LDH assays, MTT assays, and live/dead cell assays. We will monitor (1) whether the LDH and MMT levels of all-natural packaging extracts are similar to the negative control and (2) whether >95% of live cells are observed in the live/dead cell assays. We expect that the all-natural food packaging will not induce cytotoxic activities on cell survival and proliferation. The biodegradability of all-natural food packaging will be evaluated in the natural environment. We propose to bury multiple model-suggested all-natural packaging in soil and assume that the natural macromolecules gradually broke down by various microorganisms while the plastic-based packaging remained intact.Objective 4: Integration of antimicrobial materials into all-natural packaging materials. Active packaging employs technology that intentionally releases or absorbs compounds from the food or the headspace of food packaging, which extends the shelf life of products by stalling the degradative reactions of lipid oxidation, microbial growth, and moisture loss and gain better than traditional food packaging.To evaluate the function of AI/AL design food packaging, antibacterial compounds, e.g., metal nanoparticles, phytochemical compounds, will be loaded in the predicted matrix. The candidate compounds loaded in the active package are listed in table 2. To assess the antimicrobial and antifungal efficacy of developed fabrication receipt, food package sample will be put into tight contact with E. coli (~5 log colony-forming unit (CFU) per sample), L. innocua (~ 5 log CFU per sample) and A. fumigatus (~ 3 log CFU per sample) for 5 minutes, 1 hour, and 24 hours using a direct contact assay method. We expected that AI/AL designed package will achieve all bacterial population reduction at contact time longer than 1 hour. In addition to antimicrobial and antifungal efficacy tests, the package will also be applied to preserve vegetable, e.g., avocados that has high nutrients but easy to be spoilage. We will measure the number of rotten avocados and weight loss between different preserve methods and compare them with uncoated avocados, AI/AL designed package is expected to reduce the weight loss and protect avocado against microflora on the exocarp of avocado.