Dual-Axis RCCL: Representation-Complete Convergent Learning for Organic Chemical Space

Machine learning is profoundly reshaping molecular and materials modeling; however, given the vast scale of chemical space (10^30-10^60), it remains an open scientific question whether models can achieve convergent learning across this space. We introduce a Dual-Axis Representation-Complete Convergent Learning (RCCL) strategy, enabled by a molecular representation that integrates graph convolutional network (GCN) encoding of local valence environments, grounded in modern valence bond theory, together with no-bridge graph (NBG) encoding of ring/cage topologies, providing a quantitative measure of chemical-space coverage. This framework formalizes representation completeness, establishing a principled basis for constructing datasets that support convergent learning for large models. Guided by this RCCL framework, we develop the FD25 dataset, systematically covering 13,302 local valence units and 165,726 ring/cage topologies, achieving near-complete combinatorial coverage of organic molecules with H/C/N/O/F elements. Graph neural networks trained on FD25 exhibit representation-complete convergent learning and strong out-of-distribution generalization, with an overall prediction error of approximately 1.0 kcal/mol MAE across external benchmarks. Our results establish a quantitative link between molecular representation, structural completeness, and model generalization, providing a foundation for interpretable, transferable, and data-efficient molecular intelligence.

💡 Research Summary

The vastness of chemical space, estimated between $10^{30}$ and $10^{60}$, poses a fundamental challenge to machine learning: how can we ensure that a model’s predictive behavior stabilizes as the explored space expands? This paper addresses this “open scientific question” by introducing the Dual-Axis Representation-Complete Convergent Learning (RCCL) framework. The core innovation lies in quantifying “representation completeness” to achieve convergent learning, where the model’s performance becomes robust and predictable across expanding chemical domains.

The researchers propose a dual-axis approach to capture the essence of chemical diversity. The first axis focuses on local valence environments. By utilizing Graph Convolutional Networks (GCN) integrated with modern valence bond theory, the authors developed a single descriptor, $\xi$, which compresses atomic-level electronic structures, including core, $\sigma$, $\pi$, and anti-bonding orbital energies. Their statistical analysis demonstrates that $\xi$ is highly sensitive to 1-hop neighbor changes but stabilizes beyond 2-hops, effectively identifying the minimum unit of chemical diversity. The second axis addresses global topology through the concept of No-Bridge Graphs (NBG). By defining NBG0—the minimal topological unit that remains connected even after removing certain vertices or edges—the framework systematically enumerates complex ring and cage structures.

To implement this theory, the authors developed the FD25 dataset, comprising 2.1 million molecules composed of H, C, N, O, and F. FD25 achieves near-complete combinatorial coverage of organic chemistry, encompassing 13,302 local valence units and 165,726 ring/cage topologies. This coverage is more than ten times greater than existing benchmarks like GDB-9/11/17 or PC3M, providing a practical realization of “representation completeness.”

The empirical results are highly significant. Graph Neural Networks (GNNs) and Transformer-based Large Language Models (LLMs) trained on FD25 demonstrated exceptional out-of-distribution (OOD) generalization. Even when faced with unseen complex ring systems, non-standard electronic configurations, and larger molecules (>13 atoms), the models maintained a high level of precision, achieving an overall prediction error of approximately 1.0 kcal/mol MAE across multiple external benchmarks, including QM9, ANI-1x, and OpenFF.

Furthermore, the paper outlines three strategic pillars for achieving chemical space completion: exhaustive enumeration of small molecules ($N \le 6$) to cover fundamental building blocks, the construction of global minimum energy subsets to capture high-energy states, and the implementation of elemental uniformity to mitigate bias. In conclusion, this research provides a quantitative link between molecular representation, structural completeness, and model generalization, establishing a foundational methodology for developing interpretable, transferable, and data-efficient molecular intelligence.