Projects

D–A organic photocatalyst design & screening

Molecular Machine Learning · Quantum Chemistry · High-Throughput Screening

• For charge separation and exciton binding energy in donor–acceptor organic photocatalytic systems, built a research pipeline of "molecular structures → quantum chemistry → machine learning → candidate screening".
• A tiered PM6 → DFT → TDDFT/Multiwfn computation pipeline; the data-production and screening scale is expanding to ~26,000 molecular systems.
• Explored a donor / acceptor two-tower 3D molecular model fused with low-cost electronic-structure information (PM6 HOMO, LUMO, frontier-orbital gap); the best model reaches a held-out test MAE of 0.0702 eV.
• Built the ExcitationNexus research platform, integrating molecular registration, task management, result parsing, model prediction, and candidate ranking into a trackable workflow.

Read the full case study →

Polymer self-assembly NanoPUF chip identity verification & intelligent verification platform

BCP Self-assembly · SEM · AI Pairwise Verification

• Uses the random nanostructures of block-copolymer self-assembly as a physical fingerprint on chip surfaces, building a research chain of "SEM nano-images → feature representation → pairwise verification → audit".
• Current validation data: 2 chips, 20 PUF regions, 126 SEM images; 44-dim single-image features and 88-dim pairwise representations.
• Held-out test: AUC 0.977, balanced accuracy 0.9067; an observed FAR of 0 is a pilot-dataset observation only and must not be extrapolated to deployment-grade security guarantees.
• Built the NanoPUF Research Platform integrating samples, metadata, verification, and audit, supporting model iteration and result traceability.

Research details →

Electrolyser-array optimisation model & control system for hydrogen production

Python · Electro-thermal Coupled Modeling · Scheduling Optimisation

• For power allocation, temperature divergence, and hydrogen-production efficiency of electrolyser arrays under off-grid wind volatility, built an electro-thermal coupled model of alkaline electrolysers and a control interface in Python.
• Proposed a rotation-based scheduling strategy balancing dispatch uniformity, equipment lifetime, and hydrogen economics, incorporating start-stop characteristics and temperature-gradient constraints.
• Led the project through a national-level undergraduate innovation programme to completion; registered the Cup-AWE V1.0 alkaline-electrolyser simulation software as student first inventor.

Yuanju AI short-video mini program for storefront marketing

FastAPI · FFmpeg · Object Storage · Agent Orchestration

• For brick-and-mortar stores in catering, beauty, and retail, explores turning storefront photos, filmed footage, and business info into short-video marketing material automatically.
• Two content-production chains — image-to-video and smart editing of filmed footage — supporting transitions, subtitles, voice-over, BGM, post copy, and asset persistence.
• Cloud Alpha engineering validation built on FastAPI, FFmpeg / PIL, and object storage, with agent-style task orchestration organising asset processing, generation, checks, and failure repair.
• The current focus is validating the real content-production flow and product experience for merchants; no unverified user or revenue metrics are used as promotion.

ViolinSense AI · AI music × violin practice assistant

Acoustic Analysis · AI Music Generation · Practice Feedback

• With a collaborator, explores combining generative music with real violin learning: lyrics / style-driven generation, reference audio, stem separation, and audio editing.
• For violin practice, designs pitch curves, cents deviation, rhythm analysis, score recognition, transposition, and AI-assisted practice.
• The project is at the prototype and technical-validation stage (V0); the goal is to help learners understand performance while keeping real instrument training and musical expression central.