Writing
Research writing extends the site’s method-led identity into accessible prose. Articles are organized by category:
- Research Explainers — accessible summaries of principal papers and methods, aimed at readers from adjacent fields.
- Methods Notes — technical notes on current methods, prototypes, and open questions.
- Project Notes — reflections on specific projects, past and present, connecting them to the research program.
- Teaching and Software — notes from courses, labs, and software development that may help students and collaborators.
- Technical Notes — infrastructure, tooling, and deployment articles (RAG pipelines, environment setup, etc.). These remain discoverable but secondary to the research writing.
Control Without Command: Using EV Prices to Support a Microgrid
How a microgrid can offer prices to electric vehicles, learn their changing aggregate response, and use that uncertain flexibility alongside fast direct frequency control.
When the Schedule Becomes a Suggestion: Microgrid Flexibility with MPC
How a fast microgrid MPC can follow a long-horizon energy schedule while retaining enough discretion to reject disturbances, respect constraints, and respond to forecast errors.
NewPINN Does Not Eliminate Loss-Weight Tuning: A Controlled Ablation on 1D Burgers
NewPINN replaces the PDE residual loss with a solver-consistency loss and claims to eliminate the loss-weight tuning that plagues standard PINNs. A controlled ablation sweeping the loss-weight ratio over three orders of magnitude on 1D Burgers refutes the robustness claim: NewPINN is ~6.8x MORE sensitive to the loss-weight ratio than the PINN baseline. The good news is that at its best weight, NewPINN is 7.6x more accurate than PINN at its best. The value is in the accuracy gain, not in eliminating tuning.
The Adaptive Acquisition Boundary: Where Active Sensing Starts to Win
Adaptive acquisition has a boundary condition. Below D=5 it does not beat uniform sampling; above D=5 with a tight budget it wins decisively. A controlled dimensionality sweep across a shared radar forward model resolves the long-standing contradiction between our 1D radar negative result and the high-D torchebm positive result — the two experiments were on opposite sides of the boundary all along.
When the Schedule Meets the Grid: Three-Level Microgrid Control in the Laboratory
How an economic scheduler, a fast MPC frequency controller, and local plant controllers were coordinated on a grid-connected laboratory microgrid.
A Knowledge Graph for Experimental Insights: Connecting Experiments, Themes, and Literature in Neo4j
Experimental insights scatter across project files, experiment logs, and paper notes — making it hard to see what the portfolio actually knows. A Neo4j knowledge graph that connects experiments to findings to research themes to the literature changes that: theme-polarity queries reveal where evidence supports and contradicts each theme, drift tracking keeps the graph in sync with org-mode control files, and the literature-experiment bridge through shared method nodes shows which paper claims your experiments test.
Sparse First, Bayesian Second: Learning Prosumer Dynamics with SINDYc and MCMC
How sparse system identification can focus Bayesian inference—and why posterior uncertainty still depends on excitation, model structure, and prior design.
Move the Equilibrium, Not the Frequency: Target-Adjusted MPC for Microgrids
How target-adjusted Model Predictive Control estimates a microgrid disturbance, moves its operating equilibrium, and regulates frequency under actuator constraints.
Research Trajectory: From Energy Grids to Adaptive Biomedical Sensing
How a shared methodological base — uncertainty-aware inference, adaptive decision-making, and closed-loop control — connects energy systems research to current work in adaptive biomedical sensing and closed-loop discovery.
Reflections on Visiting Qianhai: Innovation, Openness, and Opportunity
An HKWYSC excursion into Shenzhen's Qianhai district — impressions from company visits across smart home devices, robotics, medical devices, and memory technologies, plus a lunch with district leadership.
Co-Creating the Future: Reflections from HKWYSC 2026
Reflections from the 3rd Hong Kong World Youth Science Congress and Xiangjiang Nobel Forum 2026 — a week of conversations linking AI, biomedicine, commercialization, and cross-border collaboration.
Psyche: Distributed Transformer Training Over the Internet — A First Look
A first-impressions tour of Psyche — a protocol for distributed training of transformer models over the internet among untrusted participants, coordinated via Solana-based consensus and game-theoretic incentives.
When Adaptive Sensing Doesn’t Win (And Why That’s Interesting)
Adaptive sensing promises smarter measurements, but on two toy problems — 1D radar-like reconstruction and 2D point-source tomography — uncertainty-driven acquisition does not consistently beat naive baselines. The negative results are constructive: they tell us when adaptive helps, when it does not, and why.
Physics-Grounded ML for Adaptive Biomedical Sensing: Three Complementary Paths
Three complementary physics-grounded ML strategies for sparse, noisy biomedical sensing: PINNs as regularizers, energy-based models for adaptive measurement, and solver-coupled networks.