LubZero: An active molecular generation framework for multi-objective properties of lubricant molecules

The screening of lubricant candidates with tailored multi-property profiles is a foundational challenge in tribology, driven by increasing demands for energy efficiency and operational safety. However, the vast discrete chemical space and the sparse-reward nature of molecular design often hinder efficient optimization. In this work, we present LubZero, a graph-based molecular generation framework that adapts the AlphaZero algorithm for the de novo design of lubricant ester materials. LubZero reformulates fragment-based molecular assembly as a single-player combinatorial game, integrating a Graph Isomorphism Network with Edge features (GINE) and Monte Carlo Tree Search (MCTS) to enable principled forward planning. Crucially, we implement a high-throughput parallel self-play architecture that utilizes multi-process inference on a shared GPU, significantly enhancing exploration efficiency and hardware resource utilization compared to serial implementations. This framework enables LubZero to navigate complex property trade-offs efficiently within a constrained molecular-design protocol. We applied LubZero to the multi-objective optimization of ester lubricants, specifically targeting high flash points and low pour points. The results show that LubZero identifies high-ranking surrogate-predicted screening candidates that extend the predicted flash-point/pour-point trade-off space beyond the region covered by the available experimental data. By integrating artificial intelligence planning with a scalable parallel computing strategy, LubZero offers a surrogate-guided workflow for screening and prioritizing candidate lubricant molecules under explicitly defined objective and applicability-domain constraints.