Data-Driven Chemical Representation Learning for Property Prediction in Materials and Molecular Systems with Complex-Valued Extensions

Abstract

Recent advances in materials and molecular machine learning have highlighted the importance of chemically meaningful representations for describing complex chemical environments and predicting physicochemical properties. In this thesis, graph neural network–based approaches incorporating complex-valued representations were developed to improve the modeling of electronic interactions, local structural environments, and stereochemical relationships in crystalline materials and molecular systems. For crystalline materials, a multidimensional electronegativity-inspired representation was developed to overcome the limitations of conventional scalar or tabulated elemental descriptors. The proposed representation was learned from formation energy data while retaining the physical meaning of Pauling electronegativity through norm-based normalization. As a result, the learned descriptor captured environment-dependent elemental relationships and provided improved atomic features for crystal graph convolutional neural network （CGCNN）- based property prediction. Furthermore, a complex-valued CGCNN （CV-CGCNN） was introduced by incorporating electron affinity and ionization energy, which are closely related to Mulliken electronegativity, into the real and imaginary components of complex-valued atomic embeddings through feature-wise modulation. This complex-valued extension enabled more effective modeling of chemical bonding environments and electronic interactions, leading to improved prediction of formation energy, band gap, and Fermi energy. For molecular systems, a complex-valued message passing neural network （CV-MPNN） was developed to perform molecular graph learning using magnitude–phase representations. Instead of processing atom and bond descriptors only as real-valued concatenated vectors, the proposed model encoded molecular features as complex-valued variables and updated them through complex-valued message passing operations. Analysis of the learned representations showed that magnitude components tended to emphasize chemically important local environments, whereas phase components reflected relative structural and directional relationships within molecular graphs. In addition, a chirality-aware aggregation strategy was introduced for the Chiral Molecular Retention Time （CMRT） dataset to incorporate local stereochemical arrangements around chiral centers. The improved performance in elution order prediction suggests that phase-based representations can be useful for describing stereochemical and directional information in molecular systems. Overall, this thesis demonstrates that chemically grounded and complex-valued representations can provide expressive frameworks for graph-based property prediction. In crystalline materials, the electronegativity- inspired and complex-valued descriptors improved the representation of elemental interactions and electronic structures. In molecular systems, magnitude–phase representations and chirality-aware aggregation improved the description of local structural and stereochemical relationships. These findings suggest that complex-valued graph learning can serve as a useful extension of conventional real-valued models for representing chemical information in both materials and molecular systems.