A KAN-based Interpretable Framework for Process-Informed Prediction of Global Warming Potential

Process information matters as much as molecular structure: The model achieved a 25% improvement in prediction accuracy (from 61% to 86% R-squared) by incorporating process-related data like location and description alongside traditional molecular information, demonstrating that where and how a chemical is made significantly impacts its environmental footprint.

AI models can be made transparent for environmental decisions: The researchers used a Kolmogorov-Arnold Network (KAN) to convert their deep learning model into an interpretable symbolic formula, allowing chemists and engineers to understand which specific molecular and process factors drive Global Warming Potential rather than relying on unexplainable "black box" predictions.

Understanding model limitations is crucial for responsible use: The analysis revealed that predictions are reliable for common chemical processes but become less certain for high Global Warming Potential values, enabling users to know when they can trust the model's recommendations and when they should seek additional expert input.

Explainable AI can support better environmental design choices: By identifying that process title information plays a significant role in predictions, the framework provides actionable insights into how process design choices influence environmental impact, potentially guiding companies toward lower-impact manufacturing approaches.