TideGuard AI
The Project
TideGuard AI TideGuard AI is an open-source platform that forecasts where and when floating plastic will accumulate on the coast, so volunteers can clean it up before it breaks into microplastics. THE PROBLEM Around 11 million tons of plastic enter the ocean every year. In the Black Sea region there is no public forecast of marine debris, so communities always react after the plastic has already washed ashore. The 48–72 hour window between a river-mouth storm and the moment plastic washes onto Anapa or Sochi beaches is wasted. THE SOLUTION A Physics-Informed Neural Network (PINN) solves the advection-diffusion-beaching equation and learns three physical parameters — windage (α), diffusion (K), and beaching rate (λ) — directly from data. The platform includes: a public web map (Next.js + MapLibre) a mobile app for citizen plastic reports (Flutter) a Telegram bot for early warnings 10 open educational lessons in English, Russian, and Mandarin a municipal B2G dashboard (PDF / GeoJSON / Excel export) RESULTS The 5-seed PINN ensemble was validated on a synthetic Black-Sea benchmark (14-day horizon, OSF-preregistered): RMSE = 0.0929 (vs persistence baseline 0.1006 — a 7.6% reduction) NSE = 0.913 Diebold–Mariano test vs persistence: p < 3×10⁻¹¹² The real-data run is scheduled for summer 2026. Code and benchmark are reproducible from the public GitHub repository with one command. WHAT WAS DIFFICULT The hardest part was making the PINN loss converge — the data-loss and the PDE-residual loss live on different scales, so the network kept ignoring the physics. I had to normalize the PDE residual and switch to adaptive loss balancing (ReLoBRaLo) before the physics constraint actually mattered. Reading the original Raissi 2019 paper a dozen times finally made it click. PILOT First deployment is targeted at the Russian Black Sea coast (Anapa, Sochi, Gelendzhik). No paid pilots are signed yet — I am actively recruiting one. WHAT'S NEXT Looking for a marine-biology mentor, a machine-learning peer reviewer, and one or two co-leads my own age. Everything is open-source (MIT license) and reproducible with a single command. Pre-registration and model card are published in the GitHub repository.
