PUFFIN: A path-unifying feed-forward interfaced network for vapor pressure prediction

Abstract

Accurate vapor pressure prediction is crucial for various applications, but obtaining precise measurements for certain compounds is resource- and labor-intensive. This challenge is amplified when a temperature-dependent relationship is required. To address this, we introduce PUFFIN (Path-Unifying Feed-Forward Interfaced Network), a machine learning approach that combines transfer learning with a specialized inductive bias node (inspired by the Antoine equation) to enhance vapor pressure prediction. PUFFIN outperforms alternative strategies that lack inductive bias or use generic compound descriptors by leveraging inductive bias and transfer learning using graph embeddings. The framework's incorporation of domain-specific knowledge overcomes data limitations and shows promise for broader chemical compound analysis applications, including the prediction of other physicochemical properties. An important aspect of our proposed approach is its partial interpretability, as the inductive Antoine node yields network-derived Antoine equation coefficients.