Physics AI Scientist · Professor · UN AI Research Expert Abu Dhabi, UAE
I build physics-informed AI systems that get deployed — not benchmarked and abandoned. My work puts the governing equations inside the network: five families of physics-informed models, built around sound, light, brain signals, continuum flow and the chemistry of the cell.
- Director of Research Projects, Abu Dhabi Maritime Academy — AD Ports Group
- Professor, Kyrgyz National University
- AI Research Expert, United Nations (UNODC)
- Science Advisor, Multiple startups
| Area | What I work on |
|---|---|
| Acoustic PINN | Sound reaches where cameras cannot — under water, inside a machine, through a forest at night (acoustic wave equation, impedance and transmission-loss relations). |
| Spectral PINN | Light carries what the eye misses: water, pigment and structure inside a leaf (radiative transfer equation, Beer–Lambert absorption). |
| Neuro PINN | Electroencephalography shows the brain under real load, where movement drowns the signal (bioelectric field equations, cable equation, artefact models). |
| Navier PINN | Measurements are sparse; the rock, water and air between them are not (Navier–Stokes, Navier–Cauchy, advection–diffusion). |
| Biochem PINN | Yield tells you what happened, chemistry tells you why (Michaelis–Menten kinetics, binding thermodynamics, Farquhar photosynthesis). |
- 122 publications on Google Scholar · 907 citations · h-index 13 · i10-index 15
- ~30 international patents
- 8 books in 5 languages
- Springer volume on PINNs for industrial applications (in progress)
Every repository here started as real work: a paper, a field deployment, a client problem, often all three. The cough model goes back to the COVID-19 work published in IEEE OJEMB in 2021. The cricket-farm audio came from farms in Vietnam and Uzbekistan and was published in 2023. The EEG work with elite athletes stands next to a Sensors paper from 2024. The UAV spectrometer flew over hidden plantations with a United Nations agency and was published by Springer in 2025. The hippo warning started with a request from mining companies in Africa.
I publish the repositories years after the work, once the papers are out and client obligations allow it. Client data stays with the clients, so each repository is rebuilt on open datasets, with the physics written into the model and tests that tie the numbers on the page to result files. Where the physics did not beat a simpler method, the README says so. That is how I want my work to be checked.
Every repository is archived with a citable DOI. The DOI below always resolves to the latest version.
- LinkedIn — https://www.linkedin.com/in/dmitry-mikhaylov
- Substack — https://dmitrymikhaylov.substack.com
- Google Scholar — https://scholar.google.com/citations?user=tfdBDF8AAAAJ
- ORCID — https://orcid.org/0009-0009-2108-6820
- Wikidata — https://www.wikidata.org/wiki/Q88500243
- ResearchGate — https://www.researchgate.net/profile/Dmitry-Mikhaylov-9
- Scopus — https://www.scopus.com/authid/detail.uri?authorId=56203422600
- OpenAlex — https://openalex.org/A5063625880
I write about physics-informed machine learning: where physical constraints make a neural network work, where they do not, and what breaks in between. These are field notes from deployments — acoustics, remote sensing, geophysical inverse problems, biosignal validity — rather than method papers.
The recurring argument is that a model which fits data without obeying the governing equations will be right on the test split and useless in the field. Physics-informed neural networks are one answer to that, and solving inverse problems — recovering a field or a source you cannot measure directly, from sparse and noisy observations — is where they earn their cost. The failure modes matter as much as the successes, which is why several of the pieces below are about things that did not work.
- A Ghost Town Is a Geological Signal. So Is a Standing Stone. — abandoned settlements and megaliths as geophysical field data
- AI and an Oyster: The Best Water Sensor We Will Ever Deploy — biological sensing versus instrumented water-quality monitoring
- Can AI Hear a Hippo Before It Kills You? — underwater acoustic early warning, and its detection limits
- How I Visited Socotra and Started Worrying About AI's Power Bill — the energy cost of inference, seen from an off-grid island
- I Put Wires in a Mushroom — fungal electrophysiology and what counts as a signal
- A Data Centre in Orbit — thermal and power constraints on orbital compute
- I Once Built an AI That Listened to Crickets — acoustic monitoring, and session-holdout versus random-split evaluation
- Teaching a Model to Scare an Elephant — human–wildlife conflict mitigation from bioacoustics
- Can Fish Predict Earthquakes? — precursor claims, and how to test one honestly
- A Loss Teaches the Model More Than a Win — why negative results carry the information
- AI Is Learning to Read a Grape — spectral sensing in agriculture
- Sound-isolating speech masks for talking to AI — acoustic isolation for speech interfaces
Full archive: dmitrymikhaylov.substack.com