Research Trajectory: From Energy Grids to Adaptive Biomedical Sensing
TL;DR — The research program described on this site did not change subjects abruptly. A related technical toolkit — probabilistic inference, constrained optimization, sequential decision-making, and laboratory control integration — was developed in the context of energy grids and is now being extended to biomedical sensing and closed-loop experimentation. This article traces that trajectory across five stages, showing how each step built on the one before.
The Common Thread
Before the stages, the unifying idea: systems that must infer, decide, and act under imperfect information.
Whether the system is a microgrid balancing renewable generation and demand, a radar sensor reconstructing a physiological signal from noisy reflections, or an experimental platform choosing which chemical combination to test next, the core challenge is the same. The system has incomplete information about its environment. It must use what it knows to make a decision. The decision changes what it observes. The cycle repeats.
This is the thread that connects energy systems to biomedical sensing. The methods — state estimation, model predictive control, Bayesian inference, optimal experimental design — are transferable. The applications change; the architecture does not.
Stage 1: Building Energy and Thermodynamic Systems (2014–2016)
The earliest work was in building energy systems: modeling heat dynamics, optimizing solar-thermal control, and understanding how buildings respond to weather and occupancy.
Key outputs:
- Research assistant at University of Stuttgart (IAR and IFK, 2014), working on solar-thermal control and automated measurement systems.
- Google Summer of Code developer at Empa (2016), building the Modular Energy Hub Framework — an open-source optimization framework in Python and Pyomo for modeling multi-energy carrier systems.
What carried forward: The experience of modeling physical systems with known governing equations (heat transfer, thermodynamics) and optimizing their operation under constraints. The Modular Energy Hub Framework encoded thermodynamic and network constraints into a structured optimization — a pattern that reappears in later work on physics-constrained machine learning.
What was new: The recognition that building and energy systems are not deterministic. Weather, occupancy, and renewable generation are uncertain. A controller that ignores this uncertainty will make suboptimal decisions.
Stage 2: Microgrids, Flexibility, and Hierarchical Control (2016–2020)
The PhD at DTU Compute (Technical University of Denmark) focused on efficient operation, modeling, optimization, and control of energy grids — specifically islanded microgrids with high renewable penetration.
Key outputs:
- Target-adjusted MPC for microgrid frequency control (2019) — a model predictive control formulation that adjusts its target trajectory online as renewable generation and demand change, rather than tracking a fixed setpoint. Published in IET Renewable Power Generation.
- Prosumer response estimation via SINDYc + MCMC (2020) — data-driven discovery of prosumer behavior dynamics with Bayesian uncertainty quantification. Published in Energies.
- Three-level hierarchical microgrid control (2020) — scheduling, frequency control, and local tracking demonstrated on a grid-connected laboratory microgrid. Published in Electric Power Systems Research.
- SINDYc and MCMC framework — open-source software released on Zenodo.
- PhD thesis: Efficient Operation of Energy Grids (2020), DOI: 10.11581/DTU.00000334.
What carried forward:
- Model Predictive Control as the core sequential-decision framework. The target-adjusted MPC formulation — optimize over a horizon, apply the first action, re-plan when new measurements arrive — is structurally identical to the adaptive sensing loop used in current biomedical work.
- Bayesian parameter inference (SINDYc + MCMC) as a way to retain distributions rather than only point estimates. The prosumer study also exposed imperfect posterior recovery and predictive checking, lessons that motivate stronger validation in physiological state estimation.
- Laboratory control integration as a step beyond simulation. The three-level hierarchical control paper integrated multiple control timescales on a physical grid-connected microgrid; the broader principle of testing interacting components also applies to closed-loop bioassay platforms.
What was new: The experience of closing the loop — connecting sensing, inference, and control in real time on physical hardware. This is the direct precursor to closed-loop experimentation.
→ See the project page on target-adjusted MPC, the SINDYc + MCMC project, and the hierarchical control project.
Stage 3: Grey-Box Modeling and Bayesian System Identification (2020–2024)
The postdoctoral period at ETH Zurich and NCCR Automation extended the probabilistic system identification work into broader automation and control contexts.
Key outputs:
- Postdoctoral research on energy systems, automation, modeling, and control (ETH Zurich, 2020–2024).
- Scientific Officer and Data Manager for NCCR Automation (2020–2024), managing research data infrastructure for a large Swiss national research centre.
- Lecturer and laboratory lead for the Quad-Rotors and Control Experiments Lab (ETH Zurich, 2023–2024), teaching undergraduate control and drone automation.
What carried forward:
- Probabilistic thinking as a default. The energy work motivated carrying uncertainty and checking models rather than relying only on point estimates. Current biomedical sensing extends that principle, with calibration treated as something to demonstrate rather than assume.
- Data management and reproducibility. The NCCR Automation role built infrastructure for reproducible research — a practice that carries into the current work on open-source software and documented experimental protocols.
- Teaching as a forcing function. Teaching control to undergraduates required distilling complex ideas into accessible explanations — the same skill that makes the Methods page and research blog posts readable to adjacent-domain readers.
What was new: Exposure to a wider range of automation problems and research data practices. The postdoctoral period did not produce a single breakthrough publication, but it consolidated the methodological toolkit and established the reproducibility standards that the current research program follows.
Stage 4: Physical Automation and Experimental Platforms (2023–2024)
The laboratory lead role at ETH Zurich and concurrent collaborations began shifting the focus from pure simulation and control to physical experimental platforms.
Key outputs:
- Quad-Rotors and Control Experiments Lab — designing and running undergraduate experiments on real quadrotor hardware, connecting control theory to physical implementation.
- Initial collaborations on closed-loop experimental design, exploring how the same sequential-decision framework used for microgrid control could apply to laboratory experiments that must choose what to measure next.
What carried forward:
- Hardware is not optional. The microgrid work had already demonstrated that simulation results do not guarantee real-world performance. The quadrotor lab reinforced this lesson: a controller that works in simulation can fail on real hardware due to unmodeled dynamics, sensor noise, and actuator limits.
- Human supervision of autonomous systems. The quadrotor lab required safety pilots and emergency stop procedures — a model for the expert review gate in the current closed-loop discovery work.
What was new: The recognition that the same closed-loop architecture — measure, infer, decide, act — applies to experimental platforms that are not energy systems. A bioassay platform that chooses which chemical combination to test next is structurally identical to a microgrid controller that chooses which generator to dispatch next. Both face uncertainty, both must act before full information is available, and both benefit from a human review gate for high-stakes decisions.
Stage 5: Adaptive Biomedical Sensing and Closed-Loop Discovery (2026–present)
The current research program applies the full methodological toolkit to biomedical sensing and closed-loop experimentation.
Key outputs (ongoing):
- Adaptive radar-based physiological sensing — using mmWave radar for contactless monitoring of respiration and cardiac motion, with adaptive waveform design to maximize information gain under real-time constraints. Conducted as Guest Researcher at Hangzhou Institute of Technology, Xidian University.
- Closed-loop discovery with AntiSyn-AI — an expert-reviewed autonomous platform for antibacterial combination screening, where active experiment selection prioritizes informative conditions under assay cost and safety constraints. Manuscript under review at npj Antimicrobials and Resistance.
- Virus-host entity mining — a related biomedical NLP collaboration, manuscript under review at Genomics, Proteomics & Bioinformatics.
- Physics-grounded ML for adaptive sensing — a three-thread research program combining physics-informed neural networks (PINNs), energy-based models (EBMs) for adaptive measurement, and solver-coupled training (NewPINNs). See the baseline blog post and the scaling roadmap.
What is new (and what is not):
| Component | New? | Precedent |
|---|---|---|
| Sequential decision under uncertainty | No | Target-adjusted MPC (2019) |
| Bayesian parameter inference and model checking | No | SINDYc + MCMC (2020) |
| Closed-loop hardware validation | No | Hierarchical microgrid control (2020) |
| Human supervision gate | No | Quadrotor lab safety protocols (2023) |
| Biomedical application domain | Yes | — |
| Physics-grounded ML (PINNs, EBMs, NewPINNs) | Yes | — |
| Adaptive measurement selection | Yes | — |
The methodological base is established. What is new is the application domain (biomedical rather than energy) and the specific ML techniques (physics-grounded neural networks, energy-based models) that the new domain requires.
What is not claimed: The biomedical work is ongoing, not completed. The radar sensing is a research testbed, not a clinical device. The closed-loop discovery platform is under review, not published. The physics-grounded ML results are from controlled toy problems, not clinical validation. Every claim on this site is labeled with its status — see the status vocabulary for the full convention.
→ See the Research page for the current research program, the Applications page for the application domains, and the Publications page for the complete output record.
The Trajectory in One Diagram
2014─2016 2016─2020 2020─2024 2023─2024 2026─present
│ │ │ │ │
▼ ▼ ▼ ▼ ▼
Building Microgrids Grey-box Physical Adaptive
Energy → & MPC → Modeling → Automation → Biomedical
Systems Control & Bayesian Platforms Sensing &
(PhD, DTU) ID (ETH, NCCR) (ETH Lab Lead) Closed-Loop
Discovery
(Xidian, HZGIFT)Each stage inherits the methodological toolkit from the previous one and extends it to a new class of problems. The arrow is not a change of subject — it is a transfer of capability.
What This Trajectory Means for a Reader
If you are an academic collaborator: the site demonstrates a coherent research program with published evidence across multiple domains. The methods are transferable; the applications are specific. If your problem involves inference under uncertainty, sequential decision-making, or closed-loop experimentation, the relevant expertise is documented here.
If you are a funding reviewer: the trajectory shows methodological development, not topic-hopping. Each stage produced published outputs. The current biomedical work is grounded in a decade of related technical work, not an abrupt pivot.
If you are a prospective student or postdoc: the research program has a clear direction and a documented track record. The open questions are named — see the open questions in the physics-grounded ML post and the scaling roadmap.
If you are an industry R&D partner: the methods have been validated on physical hardware, not just in simulation. The closed-loop experimentation framework includes a human supervision gate, making it suitable for applications where safety constraints matter.
Related Reading
- Research page — the current research program overview
- Methods page — the four methodological pillars
- Applications page — where the methods are applied
- About page — biography and career timeline
- Physics-grounded ML for adaptive biomedical sensing — the three-thread research program
- Roadmap: Scaling physics-grounded ML toward biomedical scenarios — the planning document for next steps
- When adaptive sensing doesn’t win — honest boundary conditions for adaptive acquisition
This article is a research trajectory overview, not a peer-reviewed publication. All claims about published work link to DOIs or stable project pages. Ongoing and proposed work is labeled with its current status. See the status vocabulary for the full convention.