Cristina Cornelio, Judy Goldsmith, et al.
JAIR
Safer Materials Advisor (SMA) is an AI application designed to identify and characterize materials of concern within products described in documents such as Safety Data Sheets (SDS) or Full Material Declarations (FMDs). In the identification process, the advisor extracts from those documents information related to ingredients composing the product, mainly the name, identification number and concentration. This step is enabled by our knowledge extraction pipeline (KEP) powered by LLMs to ingest documents and extract relevant information. In the sequence, the advisor uses a set of analogue methods to find similar products and suggest possible missing ingredients. This step is facilitated by a knowledge integration framework (KIF) that provides integrated way of searching heterogenous databases and by the use of AI techniques such as embeddings comparison. SMA gathers from KIF all available information related to retrieved and suggested ingredients.
The characterization process focuses on discovering additional information for ingredients. One of SMA's target use cases is the identification and characterization of PFAS, per- and polyfluoroalkyl substances known as forever chemicals. SMA classifies ingredients being PFAS or not following several definitions and lists that the advisor has access to. In regulatory analysis step, SMA provides information about specific regulatory risks according to the geography and use cases of the PFAS compounds. Additionally, SMA offers chemical hazard analysis for all compounds, by checking public lists that describes hazard information of chemicals and by predicting risks related to toxicity, persistence and bioaccumulation. Prediction of those properties is aided by FM fine-tuned specifically for that purpose.
SMA is developed by using a framework for tools integrations called discovery workbench (DWb). With SMA, user has detailed information regarding chemical safety and regulatory compliance about each ingredient in their products and is able to generate a report or define the next steps towards replacing concerning compounds.
Cristina Cornelio, Judy Goldsmith, et al.
JAIR
Pavel Klavík, A. Cristiano I. Malossi, et al.
Philos. Trans. R. Soc. A
Erik Altman, Jovan Blanusa, et al.
NeurIPS 2023
Conrad Albrecht, Jannik Schneider, et al.
CVPR 2025