Edit | X
Plastic waste has a serious impact on the ecological balance and human health. In recent years, materials scientists have struggled to find all-natural alternatives to plastics that can be used in packaging and product manufacturing.
However, it remains challenging to discover all-natural alternatives that meet specific properties. Current methods still rely on iterative optimization experiments.
Recently, researchers at the University of Maryland, College Park, proposed an integrated workflow that combines robotics and machine learning to accelerate the discovery and design of environmentally friendly plastic alternatives.
Professor Po-Yen Chen, co-author of the paper, said: "By combining automated robotics, machine learning and molecular dynamics simulations, we accelerate the development of eco-friendly, all-natural plastic alternatives that meet essential performance criteria, and our integrated approach combines automated robotics, machine learning and active learning loops to accelerate the development of biodegradable plastic alternatives. 」
该研究以《Machine intelligence-accelerated discovery of all-natural plastic substitutes》为题，于 2024 年 3 月 18 日发布在《Nature Nanotechnology》上。
Paper link: https://www.nature.com/articles/s41565-024-01635-z
Professor Chen said, "This study was inspired by a visit to Palau in the Western Pacific in 2019. The impact of plastic pollution on marine life – floating plastic film – 'tricking' fish and turtles into mistaking plastic waste for food – is deeply disturbing. This prompted me to apply my expertise to this environmental problem and to focus on finding solutions as I set up a research lab at the University of Maryland. 」
Finding sustainable alternatives to plastics through traditional methods is time-consuming and inefficient. And, often, poor results are produced, for example, by identifying materials that are biodegradable, but do not have the same desirable properties as plastics.
Active learning, robots and humans work together to build high-precision predictive models
The innovative approach to identifying alternatives to plastic in the study relies on a machine learning model developed by Chen.
In addition to being faster than traditional material search methods, this method can also be more effective at discovering materials that can be used in manufacturing and industrial environments. Chen applies his machine learning techniques to discover all-plastic alternatives.
Figure 1: Machine intelligence accelerates the discovery of all-natural plastic alternatives with programmable properties. (Source: Paper)
First, Chen and his colleagues compiled a comprehensive library of nanocomposite films from a variety of natural sources. This is done using an autonomous pipetting robot that can independently prepare laboratory samples.
Subsequently, the researchers used this sample bank to train Chen's machine learning-based model. During the training process, the model gradually becomes more proficient at predicting the properties of materials based on their composition through an iterative process of active learning.
Figure 2: Building high-precision predictive models through active learning loops, computer data augmentation, and robots working with humans. (Source: Paper)
Specifically, the study selected four naturally occurring ingredients that are generally recognized as safe (GRAS): cellulose nanofibers (CNFs), montmorillonite (MMT) nanosheets, gelatin, and glycerin, as building blocks for a variety of all-natural plastic alternatives.
First, an automated pipetting robot, i.e., an OT-2 robot, was commanded to prepare 286 nanocomposites with different CNF/MMT/gelatin/glycerol ratios and evaluate the film quality to train a support vector machine (SVM) classifier. Next, 135 all-natural nanocomposites were fabricated in stages through 14 active learning loops with data augmentation, and artificial neural network (ANN) prediction models were established.
The study demonstrates that predictive models can perform bidirectional design tasks: (1) predicting the physicochemical properties of all-natural nanocomposites based on their composition, and (2) automating the reverse design of biodegradable plastic alternatives to meet a variety of user-specific requirements.
By inputting specific performance criteria, the predictive model found several all-natural plastic alternatives that fit without the need for iterative optimization experiments.
"The synergy between robotics and machine learning not only accelerates the discovery of natural plastic alternatives, but also enables targeted design of plastic alternatives with specific properties," Chen said. "Compared to traditional trial-and-error research methods, our approach significantly reduces the time and resources required. 」
Discover several all-natural alternatives to plastic
The researchers used the model to predict the performance of nanocomposites. The model accurately predicted the optical transmittance, fire resistance and stress-strain curves of a variety of natural nanocomposites, which was in good agreement with the experimental results.
Figure 3: The model accurately predicts optical, flammable, and mechanical properties. (Source: Paper)
The model performs an automated reverse design of an all-natural plastic substitute with programmable physicochemical properties.
Figure 4: AI/ML-accelerated reverse design of all-natural nanocomposites for model interpretation of multiple plastic alternatives. (Source: Paper)
To investigate the strengthening mechanism between the CNF chain and the MMT nanosheets, the researchers performed MD simulations on three models under tension: CNF only, MMT only, and MMT/CNF models. In the MMT/CNF model, the tensile failure mechanism is different from the CNF-only and MMT-only models.
Figure 5: MD simulations reveal mechanisms of deformation and failure at the molecular scale. (Source: Paper)
SHapley Additive exPlanations (SHAP) model analysis was used to determine the effects of different gelatin sources and MMT sizes on all nine attribute labels. SHAP analysis showed that the gelatin source and MMT size had a considerable impact on optical properties, but limited on fire resistance and mechanical properties.
Future Research
In the next study, the researchers plan to continue working on solving the environmental problems caused by petrochemical plastics.
For example, they want to expand the range of natural materials that manufacturers can choose from. In addition, they will try to broaden the possible applications of the materials identified by their models and ensure that these materials can be produced on a large scale.
"We are now working to find suitable biodegradable and sustainable materials to package fresh produce after harvest, replace single-use plastic food packaging, and increase the shelf life of these products. Chen added.
"We are also looking at how to manage the disposal of these biodegradable plastics, including recycling them or converting them into other useful chemicals. The study makes a significant contribution to global initiatives to reduce plastic pollution. 」
References: https://phys.org/news/2024-04-machine-based-approach-nanocomposite-biodegradable.html
Note: The cover is from the Internet