Skip to main content
Go to the home page of the European Commission (opens in new window)
English en
CORDIS - EU research results
CORDIS

Phase Space Foundations for Scientific Machine Learning

Project description

Interpretable AI in phase space

Machine learning in science is rapidly shifting from interpretable physical models to highly parameterised neural networks, raising questions about what is learned from data versus underlying physics. The ERC-funded PhaseShift project addresses this gap by developing a theoretical foundation for scalable, interpretable scientific machine learning grounded in Hamiltonian phase spaces and microlocal analysis. It designs operators exploiting universal phase-space locality, improving sample efficiency and stability. It also introduces kinetic phase-space surrogate models that extend beyond traditional approaches based on partial differential equation. The project also studies scale interactions and inductive biases for learning physical systems and multivariate time series dynamics. Applications include cryo-electron tomography, seismic signal modelling and geophysical granular flow prediction, advancing imaging and Earth science.

Objective

Science and engineering have long relied on interpretable, parsimonious models of signals and systems. Today there is a shift to highly-parameterized neural networks trained on massive datasets. This often makes it hard to understand what is taken from physics and what from data, and what mechanisms determine when the methods succeed or fail. There is a need for solid foundations for scalable and interpretable scientific machine learning. A key idea is to work in phase spaces inspired by Hamilton’s and Boltzmann’s formalisms. Phase spaces lift problems into higher-dimensional representations but in return render descriptions of a broad range of phenomena simple and interpretable. They reveal the universal local structure which is key for efficient computation, learning, and interpretability. In PhaseShift, we will: - Develop theory and designs for operators that leverage the universal Hamiltonian structure underlying applications from tomography to fluid flows, as described by microlocal analysis. We will prove that phase-space locality significantly improves sample efficiency and Lipschitz stability over existing networks. - Introduce a new generation of interpretable surrogate models built on discrete kinetic phase spaces, going well beyond PDE-based physics to phenomena like granular flows where classic PDEs do not suffice. We will also adapt kinetic theory principles to model general multivariate time series - Study of the role of scale interactions in both data and physics, and the related inductive biases for learning physics. In parallel, we will apply the developed tools to outstanding challenges in imaging and Earth science: end-to-end learning for cryoelectron tomography to visualize cells at near-atomic resolution, kinetic modeling of multistation seismic signals, and learning the dynamics of geophysical granular flows.

Fields of science (EuroSciVoc)

Keywords

Programme(s)

Topic(s)

Funding Scheme

HORIZON-ERC - HORIZON ERC Grants

See all projects funded under this funding scheme

Call for proposal

(opens in new window) ERC-2025-COG

See all projects funded under this call

Host institution

UNIVERSITAT BASEL
Net EU contribution
€ 1 995 957,00
Address
PETERSPLATZ 1
4051 Basel
Switzerland
Region
Schweiz/Suisse/Svizzera Nordwestschweiz Basel-Stadt
Activity type
Higher or Secondary Education Establishments
Links
Total cost
€ 1 995 957,00

Beneficiaries (1)