Designing self-maintaining inorganic systems at the boundary between chemistry and biology.
We study whether iron sulfide minerals — the same materials found at deep-sea hydrothermal vents — can be engineered into a minimal system that maintains itself without being alive. The core idea: a bilayer pentlandite + mackinawite membrane that drives CO2 reduction to formate using only a pH gradient — no enzymes, no genes, no replication.
| Project | Description | Status |
|---|---|---|
| digital-twin | ORACLE pipeline: Sobol sensitivity, Gillespie SSA, FNO/PINN surrogates, DFT scripts, SRE monitoring for cloud compute | Active |
| sulfide-proton-barriers | DFT NEB calculations of H+ diffusion barriers in iron sulfides (pentlandite, mackinawite, pyrite, troilite) | Active |
| crystalformer-x | Post-generation analysis and screening toolkit for CrystalFormer: Voronoi channel analysis, percolation, composition-biased sampling | New |
| yt-slide-mark | Extract slides from YouTube videos and pair them with transcript text in Markdown | Released |
| exopoiesis.space | Project website (EN/RU/ZH) | Live |
- Pentlandite blocks protons: E_a = 1.1 eV (GPAW DFT), 0.9 eV (ABACUS DFT), 1.43 eV (MACE-MP-0)
- Mackinawite conducts protons via Grotthuss mechanism: E_a = 0.44 eV (MACE), 0.74 eV (GPAW DFT)
- Pyrite has low hop barrier but high vacancy formation energy: E_f = 1.95 eV
- Cross-verification across 3 DFT codes (GPAW, Quantum ESPRESSO, ABACUS) + ML potential (MACE)
- Bilayer membrane architecture enables spontaneous CO2 reduction at pH gradient >= 2.3
- ORACLE surrogate model: RF feature importance (AUC = 0.982) identifies k_d as dominant parameter
- System is transport-limited: membrane thickness is the only tunable lever
Python · ASE · GPAW · Quantum ESPRESSO · ABACUS · JDFTx · MACE-MP-0 · CrystalFormer · MatterGen · CHGNet
CI-NEB · AIMD · DFT+U · CANDLE solvation · Gillespie SSA · Sobol sensitivity · FNO · PINN · PyTorch · JAX
Docker · Vast.ai · GCP · Claude Code
Independent researcher, Ukraine. Background in computer hardware engineering. Funded by Trelis Research AI compute grant. This project is developed through human-AI collaboration using Claude Code (Anthropic).