Selected work

Representative Papers

Property-Guided Molecular Generation and Optimization via Latent Flows

Mar 27, 2026

This work addresses the challenge in molecular inverse design of simultaneously achieving desired properties, molecular validity, structural fidelity, and optimization stability. To this end, the authors propose MoltenFlow, a unified framework that integrates property-aligned representations, flow-matching generative priors, and gradient-guided optimization within a shared latent space. MoltenFlow is the first method to unify high-quality unconditional generation with controllable multi-objective conditional optimization. Experimental results demonstrate that, under a fixed evaluation budget, MoltenFlow significantly improves the validity, diversity, and optimization efficiency of generated molecules while maintaining robust stability and practical applicability.

0 citationsRead paper

Hierarchy-Guided Topology Latent Flow for Molecular Graph Generation

Mar 27, 2026

This work addresses key challenges in unconditional 3D molecular generation—specifically, valency violations, disconnected fragments, and implausible ring structures arising from discrete bond topologies, which are particularly pronounced in drug-like molecules with long-range constraints. To tackle these issues, the authors propose a planner-executor architecture that integrates multiscale latent planning to capture global context with a constraint-aware sampler that explicitly generates bond graphs endowed with 3D coordinates. This approach uniquely unifies explicit topology generation with hierarchical planning, enforcing strong topological feasibility. The resulting Hierarchically Guided Latent Topology Flow (HLTF) model enables end-to-end joint generation, achieving 98.8% atomic stability and 92.9% valid, unique molecules on QM9. On GEOM-DRUGS, it attains 85.5% validity without post-processing, improving to 92.2% after relaxation—approaching state-of-the-art baselines.

0 citationsRead paper

AI-Guided Discovery of Novel Ionic Liquid Solvents for Industrial CO2 Capture

Jan 02, 2026

This study addresses CO₂ capture from industrial flue gas by proposing an AI-driven, end-to-end design framework for ionic liquid solvents as sustainable alternatives to energy-intensive and corrosive amine-based systems. The approach generates candidate molecules through combinatorial pairing of cations and anions, employs graph neural networks to predict CO₂ solubility and viscosity, and utilizes the Van’t Hoff model to estimate working capacity and regeneration energy. Integrated Pareto-based multi-objective optimization and synthetic feasibility analysis enable a closed-loop pipeline encompassing molecular generation, property prediction, performance optimization, and synthesizability validation. The method successfully identifies 36 high-performance ionic liquid candidates, projected to reduce operational costs by 5–10% and capital expenditures by up to 10%.

0 citationsRead paper

GAIA: A Foundation Model for Operational Atmospheric Dynamics

May 15, 2025arXiv.org

This study addresses the challenge of learning semantically rich, atmosphere-dynamics-focused representations from geostationary satellite imagery—decoupled from diurnal texture variations—to enhance downstream atmospheric modeling. We propose the first geospatial foundation model architecture that synergistically integrates masked autoencoding (MAE) with label-free self-distillation (DINO), jointly capturing local spatiotemporal details and global dynamical dependencies. The model demonstrates robust reconstruction capability under high masking ratios and achieves high-accuracy precipitation estimation with minimal labeled data: a false alarm rate of 0.088 and structural similarity index of 0.881. Our core contribution lies in deeply adapting self-supervised learning paradigms to physical atmospheric process modeling, establishing a scalable, low-label-dependency framework for global meteorological representation learning.

0 citationsRead paper

Recent publications

Latest Papers

Property-Guided Molecular Generation and Optimization via Latent Flows

Mar 27, 2026

This work addresses the challenge in molecular inverse design of simultaneously achieving desired properties, molecular validity, structural fidelity, and optimization stability. To this end, the authors propose MoltenFlow, a unified framework that integrates property-aligned representations, flow-matching generative priors, and gradient-guided optimization within a shared latent space. MoltenFlow is the first method to unify high-quality unconditional generation with controllable multi-objective conditional optimization. Experimental results demonstrate that, under a fixed evaluation budget, MoltenFlow significantly improves the validity, diversity, and optimization efficiency of generated molecules while maintaining robust stability and practical applicability.

0 citationsRead paper

Hierarchy-Guided Topology Latent Flow for Molecular Graph Generation

Mar 27, 2026

This work addresses key challenges in unconditional 3D molecular generation—specifically, valency violations, disconnected fragments, and implausible ring structures arising from discrete bond topologies, which are particularly pronounced in drug-like molecules with long-range constraints. To tackle these issues, the authors propose a planner-executor architecture that integrates multiscale latent planning to capture global context with a constraint-aware sampler that explicitly generates bond graphs endowed with 3D coordinates. This approach uniquely unifies explicit topology generation with hierarchical planning, enforcing strong topological feasibility. The resulting Hierarchically Guided Latent Topology Flow (HLTF) model enables end-to-end joint generation, achieving 98.8% atomic stability and 92.9% valid, unique molecules on QM9. On GEOM-DRUGS, it attains 85.5% validity without post-processing, improving to 92.2% after relaxation—approaching state-of-the-art baselines.

0 citationsRead paper

AI-Guided Discovery of Novel Ionic Liquid Solvents for Industrial CO2 Capture

Jan 02, 2026

This study addresses CO₂ capture from industrial flue gas by proposing an AI-driven, end-to-end design framework for ionic liquid solvents as sustainable alternatives to energy-intensive and corrosive amine-based systems. The approach generates candidate molecules through combinatorial pairing of cations and anions, employs graph neural networks to predict CO₂ solubility and viscosity, and utilizes the Van’t Hoff model to estimate working capacity and regeneration energy. Integrated Pareto-based multi-objective optimization and synthetic feasibility analysis enable a closed-loop pipeline encompassing molecular generation, property prediction, performance optimization, and synthesizability validation. The method successfully identifies 36 high-performance ionic liquid candidates, projected to reduce operational costs by 5–10% and capital expenditures by up to 10%.

0 citationsRead paper

GAIA: A Foundation Model for Operational Atmospheric Dynamics

May 15, 2025arXiv.org

This study addresses the challenge of learning semantically rich, atmosphere-dynamics-focused representations from geostationary satellite imagery—decoupled from diurnal texture variations—to enhance downstream atmospheric modeling. We propose the first geospatial foundation model architecture that synergistically integrates masked autoencoding (MAE) with label-free self-distillation (DINO), jointly capturing local spatiotemporal details and global dynamical dependencies. The model demonstrates robust reconstruction capability under high masking ratios and achieves high-accuracy precipitation estimation with minimal labeled data: a false alarm rate of 0.088 and structural similarity index of 0.881. Our core contribution lies in deeply adapting self-supervised learning paradigms to physical atmospheric process modeling, establishing a scalable, low-label-dependency framework for global meteorological representation learning.

0 citationsRead paper