Molecular and polymer representations for machine learning

Abstract

The success of AI-driven materials discovery ultimately depends on how we represent molecular structure. Traditional approaches based on expert-defined descriptors, fingerprints, and symbolic notations have enabled advances in property prediction and molecular design. These often struggle, however, to generalise across structurally complex chemical systems. This issue is particularly striking when structural hierarchy and stochasticity are not explicitly encoded or when datasets lack access to these higher-order features. Graph neural networks and large language models offer new pathways in representation learning, enabling molecular and polymer representations to be learned directly from structural inputs. These machine-learned embeddings promise more expressive, scalable, and transferable representations, potentially transforming polymer informatics and materials discovery. In this review, we survey the evolution of molecular representations from traditional descriptors to modern embedding-based approaches, examine how these approaches have been applied to small molecules and adapted for polymer systems, and analyse the challenges posed by stochasticity, structural hierarchy and limited data availability in polymer systems. We argue that future progress will depend not only on improved model architectures but also on the development of better-curated datasets and shared benchmarks, particularly for polymers, where stochasticity and structural hierarchy complicate both representation and evaluation. Advancing representation learning in chemistry will therefore require the deliberate integration of domain knowledge, physical constraints, and data-driven methods to enable the robust and interpretable discovery of advanced materials.