Louis De Neve
  • Home
  • Travel
  • Academic
  • Rowing

    • Academic

    I am a PhD researcher at the intersection of machine learning and climate science. My work focuses on developing interpretable, data-efficient models for weather prediction and climate analysis, with a secondary interest in causal inference applied to geophysical systems.

    Research Projects

    Machine Learning for Subseasonal-to-Seasonal Forecasting

    Active

    Developing deep learning architectures that can skillfully predict climate variability on the 2–6 week timescale, where useful signal still exists in the ocean and sea-ice state.

    Deep Learning Climate Forecasting Transformers

    Mechanistic Interpretability Applied to Climate Models

    Active

    Borrowing sparse autoencoder techniques from LLM interpretability to understand what representations physics-informed neural networks learn from atmospheric data.

    Interpretability Sparse Autoencoders Climate Modelling

    Causal Discovery in Climate Time Series

    Completed

    Applying constraint-based and score-based causal discovery to multivariate climate datasets to identify robust causal relationships between climate indices.

    Causal Inference Time Series Climate Variability

    Publications

    2024

    Journal

    Towards Interpretable Climate Models: Sparse Autoencoders for Disentangling Atmospheric Circulation Patterns

    De Neve, L., Hartmann, A., Schulz, M., & Chen, J.

    Journal of Climate, 2024

    DOI: 10.1175/JCLI-D-24-0001

    Abstract

    We present a framework for applying sparse autoencoders to reanalysis data with the goal of discovering disentangled representations of large-scale atmospheric dynamics. The learned features recover known teleconnection patterns and reveal previously undercharacterised modes of variability in the mid-latitude circulation.

    Preprint

    Uncertainty Quantification in Neural Network Weather Forecasts via Conformal Prediction

    De Neve, L., Patel, R., & Hartmann, A.

    arXiv preprint, 2024

    arXiv: 2401.12345

    Abstract

    Reliable uncertainty estimates are critical for operational weather forecasting. We adapt split conformal prediction to produce distribution-free, marginally valid prediction intervals for neural network weather forecasts, demonstrating coverage guarantees on held-out test sets spanning multiple seasons.

    2023

    Conference

    Scaling Laws for Data-Driven Weather Prediction at Kilometre Resolution

    De Neve, L. & Schulz, M.

    NeurIPS Workshop on Tackling Climate Change with Machine Learning, 2023

    Abstract

    We investigate how prediction skill scales with model capacity and training data volume for deep learning-based numerical weather prediction at convection-permitting resolutions. Our results suggest that current architectures are significantly data-limited rather than parameter-limited.

    Journal

    Extreme Precipitation Attribution Using Causal Inference and High-Resolution Climate Ensembles

    Chen, J., De Neve, L., & Müller, K.

    Geophysical Research Letters, 2023

    DOI: 10.1029/2023GL105678

    Abstract

    We combine structural causal models with large climate ensembles to attribute changes in the frequency and intensity of extreme precipitation events to anthropogenic forcing. The method provides sharper attribution statements than regression-based approaches and scales to global analysis.

    Contact

    For research enquiries, collaboration proposals, or general contact, please email lfcd2@cantab.ac.uk.

    © Louis De Neve
    • Travel
    • Academic
    • Rowing
    • Contact