(697a) Machine Learning–Enabled Design of All-Natural Plastic Substitutes | AIChE

(697a) Machine Learning–Enabled Design of All-Natural Plastic Substitutes


Chen, T. - Presenter, University of Maryland
He, S., University of Maryland
Yang, H., National University of Singapore
Shrestha, S., University of Maryland, College p
Little, J., University of Maryland
Chen, P. Y., University of Maryland
Plastic pollution emerges as one critical environmental issue, as the production of non-biodegradable plastic products has overwhelmed the world’s capability of dealing with them. One immediate solution is to develop biodegradable structural materials using all-natural components that can serve as plastic substitutes in various application scenes and break down easily in the natural environment. However, several challenges remain, including rapid recipe optimization to meet product specifications and accurate property prediction from its compositions and fabrication conditions. Herein, a multi-stage machine learning framework is realized for a high-accuracy prediction model to automate the design of all-natural plastic substitutes with programmable optical, fire retardant, and mechanical properties. First, the feasible design boundary is defined to reduce experimental failures. With the aid of an automated pipetting robot, a support vector machine classifier is trained from 351 data points. Each data point contains compositions and fabrication results of various all-natural nanomaterials (including montmorillonite nanosheets, cellulose nanofibers, gelatin, and glycerol). Second, material properties across the entire design space were explored through active learning. Light transmittances at 550, 950, and 365nm were measured for optical properties. Restrained Bézier curves were used to interpolate experimental stress–strain curves and extract mechanical data labels, and image analysis was used to estimate fire retardancy. After each active learning round, artificial neural network models were trained, and recipes for the next round were generated by maximizing uncertainty expressed in distance and variance. After 14 active learning loops, 135 all-natural plastic substitutes were stagewise fabricated, and 1350 data points (virtual and real) were generated by data augmentation to enrich the multi-degree-of-freedom dataset. The mean relative error is 17% for final models, which is similar to the standard deviation observed in experiments. The ML-enabled prediction model can execute two-way tasks of automatic plastic substitute design, including (1) predicting the performance of an all-natural structural material from its compositions and (2) recommending new material compositions for a specific plastic replacement scene. Three-dimensional heatmaps of material properties vs. compositions were constructed using the prediction model to visualize the full-map material properties. Spearman correlation factor was used to find non-linear correlations between material composition and performance on the experimental dataset. And by utilizing Shapley additive explanations, the performance contribution of each component was also exemplified. As final demonstrations, a reverse-design task targeting high-stress material was conducted using the property model. By applying clustering analysis to model-suggested composition ranges, we found two interesting groups of compositions. With additional post-processing steps such as hot-press and cross-linking, all-natural plastic substitutes with ultrahigh strength (above 500 MPa) were fabricated. Compared to the conventional approach requiring trial-and-error experiments, our data-driven approach can significantly save the research and development cost for diverse plastic replacements.