Fernanda Véliz

Automated analysis of the scientific literature using natural language processing (NLP) can accelerate the identification of potentially unexplored formulations that enable innovations in materials engineering with fewer experimentation and testing cycles. This strategy has been successful for specific classes of inorganic materials, but their general application in broader material domains such as bioplastics remains challenging. To begin addressing this gap, we explore correlations between the ingredients and physicochemical properties of seaweed-based biofilms from a corpus of 2000 article abstracts from the scientific literature since 1958, using a supervised word co-occurrence analysis and an unsupervised approach based on the language model MatBERT without fine-tuning. Using known relations between ingredients and properties for test scenarios, we discuss the potential and limitations of these NLP approaches for identifying novel combinations of polysaccharides, plasticizers, and additives that are related to the functionality of seaweed biofilms. The model demonstrates a valuable predictive ability to identify ingredients associated with increased water vapor permeability, suggesting its potential utility in optimizing formulations for future research. Using the model further revealed alternative combinations that are underrepresented in the literature. This automated method facilitates the mapping of relationships between ingredients and properties, guiding the development of seaweed bioplastic formulations. The unstructured and heterogeneous nature of the literature on bioplastics represents a particular challenge that demands ad hoc fine-tuning strategies for state-of-the-art language models for advancing the field of seaweed bioplastics.