Cosmological analysis of the DESI DR1 Lyman-α 1D power spectrum
Hydrodynamical-simulation emulation for precision cosmological inference.
Cross-cutting methods
Scientific AI
I build fast, physically informed models that make expensive simulations usable for inference—and apply modern machine learning where it can preserve uncertainty, improve scale, or reveal structure that conventional methods cannot.
My approach treats machine learning as part of the scientific model, not as a black-box prediction layer. Physical parameterizations, simulation validation, uncertainty quantification, and out-of-distribution tests are central to deciding whether a model is useful for cosmology.
These methods span intergalactic-medium emulation, differentiable galaxy and halo histories, spectral modelling, photometric-redshift inference, and robust source classification. The complete publication record is grouped by methodological strand below.
01
Simulation emulators & cosmological inference
Learning costly hydrodynamical predictions while retaining explicit cosmological and astrophysical parameters.
Lyman-α forest holography: 3D predictions from 1D measurements
ForestFlow translation of DESI P1D constraints into nonlinear 3D predictions.
ForestFlow: predicting Lyman-α clustering from linear to nonlinear scales
Conditional normalizing flows trained on paired hydrodynamical simulations.
A neural-network emulator for the Lyman-α 1D flux power spectrum
Sub-percent predictions across cosmology and intergalactic-medium physics.
02
Differentiable & surrogate physical models
Compact, interpretable descriptions of halo growth, star formation, spectra, and galaxy colours.
Diffstar: a fully parametric model for galaxy assembly history
Differentiable star-formation histories tied to halo growth and quenching.
DSPS: Differentiable stellar population synthesis
JAX-native spectra and photometry with gradients through physical parameters.
A differentiable model of dark-matter halo assembly
Population-level halo growth with assembly-bias correlations in JAX.
Surrogate modelling the Baryonic Universe II
Forward modelling the colours of individual galaxies and populations.
Surrogate modelling the Baryonic Universe I
Using Galaxpy and dimensional reduction to isolate the colour of star formation.
03
Machine learning for survey data
Probabilistic inference and domain adaptation for photometric surveys.
J-PAS: semi-supervised sim-to-obs transfer
Domain adaptation for robust star–galaxy–quasar classification.
SYTH-Z: machine-learning synthetic spectra for probabilistic redshifts
Mixture-density networks trained on physically motivated synthetic data.
Scalable statistical inference of photometric redshift via data subsampling
Graph-guided subsampling and Gaussian-process ensembles with predictive uncertainty.
04
Perspectives on Scientific AI
Reviews and roadmaps connecting machine-learning methodology to trustworthy computational cosmology.
Machine Learning Techniques for Astrophysics and Cosmology: Lyman-α forest
A review spanning emulation, simulation-based inference, and field-level reconstruction.
Scientific AI Approaches in Computational Cosmology
A roadmap for emulation, uncertainty quantification, physical symmetries, and validation.