AI for Materials Science Researcher
I am a Ph.D. candidate in the Materials Theory Group at HKU, building autonomous workflows that connect atomistic simulations, scientific foundation models, and large language models for accelerated discovery.
bohrium-skills-cli to unify management of skills across the DP ecosystem.
A file-system-based agent memory that distills failures into reusable Memory and successes into reusable Skills, letting LLM agents accumulate cross-task experience. Ablations on Sol27LC and MatTools confirm that removing the memory, the sandbox feedback, or cross-conversation reuse each lowers task success.
A two-plane agent architecture — a Supervisor–Worker control plane (tool constraints, role isolation, reflective replanning) over a workflow plane that models drug discovery as composable Skill Graphs with tool governance, data contracts, and human-in-the-loop. Reaches 21.4% on the HLE drug-discovery subset and beats Biomni, SciMaster, and STELLA (e.g. +65% classification accuracy over Biomni on PharmaBench, +50% MCQ, −27% regression SMAPE).
Two benchmarks — materials Q&A and simulation-code generation — reveal that LLMs answer descriptive questions well but generate correct materials-computation code only ~20% of the time. Our agent system lifts that to nearly 60%.
A multi-agent system with depth-first search over formula-derivation paths derives an interpretable law for metallic-glass forming ability (GFA), reaching R² = 0.948 at lower formula complexity than traditional approaches.
An end-to-end, multi-task-augmented fine-tuned LLM predicts elastic constant tensors with 33.1% lower MAE than prior materials LLMs on the same training set — while also enabling inverse (property-to-material) design and retrieval-augmented (RAG) prediction.
MgBERT — MatSciBERT adapted for longer token lengths and fine-tuned on Gemini-generated compositional descriptions — raises metallic-glass binary-classification accuracy from 78.2% (GBDT) to 95.2%.
Test-Time Self-Evolution in Multi-Agent Systems for Materials Discovery
Mozi: Governed Autonomy for Drug Discovery LLM Agents
MatTools: Benchmarking Large Language Models for Materials Science Tools
DrugNav: A Benchmark Dataset of Expert Trajectories for Developing and Evaluating LLM Agents in Multi-Step Drug Discovery
Large Language Models for Material Property Predictions: elastic constant tensor prediction and materials design
Active Learning for Conditional Inverse Design with Crystal Generation and Foundation Atomic Models
A Multi-agent Framework for Materials Laws Discovery
A Prompt-Engineered Large Language Model, Deep Learning Workflow for Materials Classification
Enhancing the Efficiency and Stability of Inverted Formamidinium-Cesium Lead-Triiodide Perovskite Solar Cells through Lewis Base Pretreatment of Buried Interfaces
Surface Cleaning and Passivation Strategy for Durable Inverted Formamidinium–Cesium Triiodide Perovskite Solar Cells
The effects of organic cation rotation in hybrid Organic-Inorganic Perovskites: A critical review
Quantum dots-hydrogel composites for biomedical applications
Quantum dots-hydrogel composites for biomedical applications
Quantum dots-hydrogel composites for biomedical applications
Role of organic cation orientation in formamidine based perovskite materials
Thioacetamide-ligand-mediated synthesis of CsPbBr3–CsPbBr3 homostructured nanocrystals with enhanced stability
The latest publication list can be found on my Google Scholar.