Davide Staub

Graduate Teaching Assistant, Imperial College London: I help deliver the MSc Deep Learning course at Imperial College London. This involves running tutorials, answering students’ questions and supporting practical exercises on topics such as neural network fundamentals, optimisation and modern architectures.

Lecturer, ETH Zürich – Space Data: I teach the final third of the Space Data course in the Master in Space Systems programme at ETH Zürich. My block introduces convolutional neural networks and U-Net architectures for denoising images of the Moon’s permanently shadowed regions (PSRs). These techniques build on the HORUS framework developed by Ben Moseley and Valentin Bickel, and are key to reliable resource mapping and landing-site planning. Students learn the basics of deep learning (MLPs, CNNs, U-Nets, etc.) and then apply them to clean up PSR images using real training data.

Current research

My PhD project aims to develop a single, differentiable pipeline that converts all available JWST observations of a hot Jupiter—thermal emission spectra, phase curves, eclipses and transmission spectra—into a unified three-dimensional temperature field. The map uses a low-dimensional spherical-harmonics basis for horizontal structure and smooth vertical modes, and the inversion is regularised using physically motivated terms such as energy balance, global radiative closure and hydrostatic consistency. This framework treats the atmosphere as a shared state that must simultaneously explain observations across multiple viewing geometries.

Previous work

At the HomanLab in Zürich I explored how our brains follow the flow of stories. Working with a language model, we distilled two simple signals: a drift signal that captures the gradual build-up of meaning as a narrative unfolds, and a shift signal that spikes when the story moves to a new event or scene. When we compared these signals to high-resolution fMRI recordings from a volunteer listening to crime stories, we found that the burst-like shift signal lit up the brain’s speech and hearing centres, whereas the slow drift signal was strongest in the so-called default-mode network – regions like the angular gyrus, precuneus and posterior cingulate that support memory and imagination. This pattern suggests that auditory areas mark event boundaries while broader networks follow the slow evolution of context.

Below are brain maps showing where each signal explained neural responses. Bright colours indicate regions with stronger effects.

Master’s thesis: Physics-informed neural networks for seismology

During my master’s studies at ETH Zürich I investigated Physics-Informed Neural Networks (PINNs) as a way to solve the elastic wave equation, which describes how seismic waves propagate through the Earth. By embedding wave physics directly into the network architecture – for example using custom wavelet or plane-wave layers along with encoder and decoder components – I achieved solutions that were roughly twice as accurate as standard PINNs. I also conditioned the networks on the location of the seismic source, allowing them to infer the wavefield for countless source locations in a single forward pass and dramatically speeding up simulations compared with traditional finite-difference methods.

Astronomy & Geophysics . The report notes that incorporating wave physics improves accuracy and that once trained these networks are much faster than standard numerical solvers. You can access the full thesis online via its DOI: 10.3929/ethz-b-000668359.