Physics-informed neural surrogates, simulation-based inference, and international evaluation benchmarks for physical systems and space missions.
NeurIPS & Kaggle Benchmark
The Ariel Data Challenge
Principal Investigator · 2022 to present
Directing the international machine learning benchmark on weak-signal extraction and uncertainty calibration across 70+ countries. Explore our datasets: ADC 19/21, ADC 22, ADC 24, and ADC 25.
Open-Source Python Framework
TauREx Framework
Co-developer
Modular open-source Python ecosystem for forward physical modelling, Bayesian atmospheric retrieval, differentiable surrogates, and simulation-based inference plugins.
ESA Space Mission
ESA Ariel ML Working Group
ML Lead Coordinator · 2020 to present
Coordinating the Machine Learning Working Group at the ESA Ariel Space Mission for the 1,000-planet survey, developing pipelines for instrument detrending, noise mitigation, and retrieval.
-
Simulation-based inference & high-dimensional inverse problems. Building on the open source
TauREx framework that we co-develop, we design simulation-based inference (SBI), variational methods, and normalising flows alongside differentiable forward models for rapid, amortised Bayesian posterior estimation from noisy remote-sensing and spectroscopic instruments.
Simulation-based inference
Normalising flows
Variational inference
Differentiable physics
Remote sensing & retrieval
TauREx
-
Physics-informed neural surrogates & 3D simulation emulation. We design physics-informed neural surrogates that emulate complex radiative transfer, chemical networks, and multidimensional fluid and atmospheric dynamics at orders-of-magnitude higher speed than numerical solvers.
AI for Science
Neural surrogates
Physics-informed ML
3D fluid & radiative simulation
Chemical network emulation
-
Trustworthy scientific AI, benchmarking & uncertainty calibration. We investigate how deep learning models represent physical principles and fail under distribution shift, developing explainability tools, out-of-distribution diagnostics, and calibrated uncertainty quantification so ML inference remains transparent and reliable under real observational and sensor noise.
AI evaluation & safety
Calibrated uncertainty
Out-of-distribution diagnostics
Explainable AI
Stress-testing under noise
-
ExoHack. Organising multi-university hackathons training researchers and students in scientific machine learning, uncertainty quantification, and space mission data pipelines.
European university hackathons
Space mission AI training
