A robotics and machine learning integrated workflow for discovering all-natural plastic substitutes

The global accumulation of non-degradable petrochemical plastic waste motivates the development of biodegradable substitutes derived from sustainable resources. However, discovering all-natural composite materials that match the optical, thermal, and mechanical performance of commercial plastics remains challenging because traditional iterative optimization is inefficient. This protocol describes an integrated workflow that combines automated robotics and machine learning (ML) to accelerate the discovery of all-natural plastic substitutes. The procedure begins by commanding an automated pipetting robot to prepare a library of 286 nanocomposites that are used to train a support vector machine classifier that distinguishes high-quality, film-forming compositions from low-quality formulations thereby defining the feasible design space. Next, through 14 active learning loops coupled with data augmentation, the robot fabricates 135 all-natural nanocomposites in a stagewise manner to construct an artificial neural network (ANN) prediction model. This protocol demonstrates how to use this model for two-way design tasks: (1) accurately predicting the physicochemical properties of a nanocomposite from its composition and (2) automating the inverse design of biodegradable plastic substitutes that satisfy user-specified property targets. This protocol further details methods for model interpretation using SHapley Additive exPlanations (SHAP), validation via molecular dynamics (MD) simulations, and model expansion to incorporate new building blocks. This hybrid approach enables the accelerated discovery of eco-friendly materials with programmable functions using components drawn from the generally recognized as safe (GRAS) database.