PROJECT ID: EXO-2026 // QUANTUM COMPUTING
Hybrid Quantum ML for Biosignature Retrieval
Built during the Hack4SAGES hackathon from 10 to 12 March 2026, ExoBiome explores biosignature retrieval from exoplanet transmission spectra with a hybrid quantum-classical regression stack. The project focuses on estimating five atmospheric gases - H2O, CO2, CO, CH4, and NH3 - from Ariel-style spectra while comparing the quantum branch against classical baselines and winner-model ports.
View RepositoryProject Overview
The Scientific Problem
Biosignature retrieval from exoplanet transmission spectra requires recovering subtle atmospheric signals from complex spectral patterns. ExoBiome was built as a fast research prototype to test whether hybrid quantum machine learning can support this estimation task on realistic Ariel-style data.
Project Goal
The team organized the project around a hybrid quantum branch, classical controls, and winner-model ports so the hackathon result would be both exploratory and measurable. That structure made it possible to compare approaches quickly while still producing a polished repository and presentation.
Modeling & Evaluation
Data & Model Families
The repository centers on Ariel ADC2023-style spectra and related TauREx or POSEIDON-generated variants. Around that data, ExoBiome compares a hybrid quantum regressor with classical ablations, winner-style normalizing-flow baselines, and FMPE or flow-matching experiments.
Verified Result Snapshot
The current verified reports point to the epoch-6 hybrid checkpoint as the strongest confirmed result, with validation mRMSE around 0.2936 and holdout mRMSE around 0.2994. That made ExoBiome a concise hackathon project with a measurable benchmark instead of only a concept demo.
Technical Architecture
Hybrid ExoBiome Regressor
The main research path combines quantum components with classical regression for atmospheric gas estimation and includes dedicated variants for Ariel data, TauREx-style data, and five-qubit experiments.
Winner Ports & Ablations
The repo keeps classical controls close to the quantum branch, including winner-style normalizing-flow baselines, flow-matching variants, and non-quantum comparisons for faster benchmarking.
Generators, Validators, Reports
Supporting packages prepare datasets, validate spectra, and store generated evaluation summaries so results remain inspectable after the hackathon sprint.
Repository & Presentation
The public GitHub repository documents the model map, cleanup history, and verified result snapshot, while the recorded presentation captures the team's hackathon pitch.
Contributors
- Iwo Smura
- Iwo Wojtakajtis
- Maria Płatek
- Michał Szczęsny