Generative AI-Based Inverse Design of Sustainable Biodegradable Polymers with Target Mechanical and Thermal Properties

Abstract

The escalating global plastic pollution crisis has intensified the urgent need for sustainable biodegradable polymer alternatives that can match or exceed the performance of conventional petroleum-based plastics while minimizing environmental impact. However, traditional polymer discovery approaches are severely constrained by high experimental costs, protracted development cycles spanning years, and fundamental inability to simultaneously optimize multiple conflicting material properties such as mechanical strength, thermal stability, and degradation kinetics. This study presents a rigorously validated Generative Artificial Intelligence (GenAI) framework for the inverse design of biodegradable polymers that concurrently satisfies mechanical, thermal, and environmental sustainability requirements through an integrated computational pipeline. The proposed framework combines molecular graph representations that capture atomic-level structural information with denoising diffusion probabilistic models (DDPM) operating in continuous embedding space for novel polymer generation. A graph attention network (GAT) serves as the property predictor with multi-task learning capabilities, while NSGA-III multi-objective optimization drives the search toward Pareto-optimal candidates across competing performance criteria. Life-cycle carbon footprint assessment is embedded as a formal optimization objective rather than a post-hoc filter, enabling sustainability-driven molecular design from inception. Comprehensive computational experiments conducted on the PolyInfo and PI1M polymer databases demonstrate that the proposed framework achieves molecular validity rate of 94.8%, novelty rate of 88.3%, and tensile strength prediction RMSE of 2.87 MPa—representing substantial improvements of 13.4%, 19.4%, and 43% over the best competing baseline methods, respectively. The generated polymer candidates exhibit exceptional property profiles with tensile strengths ranging from 60 to 78 MPa, glass transition temperatures of 220 to 255°C, and biodegradation completion within 6 to 12 months, substantially outperforming conventional biodegradable polymers in all performance dimensions. This framework addresses five critical research gaps including single-property optimization limitations, absence of sustainability constraints, low molecular validity rates, dataset scarcity challenges, and interpretability deficits.