I am a PhD candidate at the ETH AI Center and D-MATH, advised by Siddhartha Mishra and Eleni Chatzi. I build systems that combine learning, mathematical structure, and symbolic reasoning to discover interpretable solutions—and then test whether those solutions are actually correct.

Conservative verification and failure analysis for generated PDE solutions.

PDECert checks symbolic identities, initial and boundary conditions, parameter assumptions, singularities, and numerical counterexamples. Its open failure atlas collects untouched solver and language-model candidates under a reproducible, blind-review protocol.

Ways to join: contribute a difficult PDE case, a solver or model output, a verification backend, or a careful review of an existing candidate.

Code · Roadmap · Contributing · Pilot dataset

• Neuro-symbolic scientific discovery: grammar-valid candidates, latent exploration, and residual-validated refinement
• ML for PDEs and dynamical systems: operator learning and physics-informed models
• Trustworthy SciML: analytical structure and adversarial solutions as tests of surrogate credibility

• 🧠 SIGS — Neuro-Symbolic AI for Analytical Solutions of Differential Equations · ICML 2026 · paper · 5-minute talk · project site
• 🌊 ZeroFlood — Geospatial Foundation Model for Flood Susceptibility Mapping · paper
• ⌚ Cross-Domain Human Activity Recognition — self-supervised learning and enhanced fine-tuning

I welcome conversations and collaborations on verifiable scientific AI, PDE failure cases, and neuro-symbolic methods. I am especially interested in cases where a plausible-looking solution fails for a subtle mathematical reason.

Google Scholar · Bluesky · LinkedIn · ooikonomou@ethz.ch