Uncertainty-Informed Screening for Safer Solvents Used in the Synthesis of Perovskite via Language Models

TL;DR

This paper introduces an uncertainty-informed framework combining language models for data extraction and deep learning for toxicity prediction to improve safety assessments of solvents in perovskite synthesis.

Contribution

It presents a novel approach integrating targeted language model data extraction with uncertainty quantification for toxicity prediction in solvents.

Findings

Automated data extraction improves data quality and reduces hallucinations.

Uncertainty quantification identifies data gaps and enhances prediction confidence.

Visualization reveals key solvent interactions related to hazards.

Abstract

The challenge of accurately predicting toxicity of industrial solvents used in perovskite synthesis is a necessary undertaking but is limited by a lack of a targeted and structured toxicity data. This paper presents a novel framework that combines an automated data extraction using language models, and an uncertainty-informed prediction model to fill data gaps and improve prediction confidence. First, we have utilized and compared two approaches to automatically extract relevant data from a corpus of scientific literature on solvents used in perovskite synthesis: smaller bidirectional language models like BERT and ELMo are used for their repeatability and deterministic outputs, while autoregressive large language model (LLM) such as GPT-3.5 is used to leverage its larger training corpus and better response generation. Our novel 'prompting and verification' technique integrated with an LLM aims at targeted extraction and refinement, thereby reducing hallucination and improving the quality of the extracted data using the LLM. Next, the extracted data is fed into our pre-trained multi-task binary classification deep learning to predict the ED nature of extracted solvents. We have used a Shannon entropy-based uncertainty quantification utilizing the class probabilities obtained from the classification model to quantify uncertainty and identify data gaps in our predictions. This approach leads to the curation of a structured dataset for solvents used in perovskite synthesis and their uncertainty-informed virtual toxicity assessment. Additionally, chord diagrams have been used to visualize solvent interactions and prioritize those with potential hazards, revealing that 70% of the solvent interactions were primarily associated with two specific perovskites.