Knowledge-Informed AI group IDSIA USI-SUPSI

Research lines

Physics-informed machine learning

Models that learn from data while respecting the differential equations that describe physical, chemical and physiological processes.

What we do

We embed mechanistic knowledge, expressed as ordinary or partial differential equations, into machine learning models. The resulting models are consistent with known mechanisms, can estimate unobserved states and parameters when these are identifiable, and require fewer observations than purely data-driven models.

Fast surrogates of complex mechanistic models make patient-specific or process-specific simulation possible in real time. The same methods are used for glucose and insulin dynamics in healthcare and for the prediction of the properties of industrial materials.

Applications. Glucose and insulin dynamics in prediabetes and diabetes, disease progression models, characterisation of industrial materials.

Methods

  • Physics-informed neural networks
  • Hybrid mechanistic and data-driven models
  • Inverse problems and parameter estimation
  • Surrogate models and digital twins
  • Meta-learning across experiments and individuals
  • Spatial regression with differential regularisation

Projects

All projects
    • Physics-informed ML
    • Industry and engineering
    • Ongoing

    AI for materials characterisation

    Physics-informed machine learning models for the prediction of the properties of industrial materials, in a research mandate with an international manufacturing company.

    2025–2027 · Industrial research mandate

    • Physics-informed ML
    • Causal AI and graphical models
    • Healthcare

    PRAESIIDIUM

    Patient digital twins that combine differential equation models of glucose metabolism with clinical and wearable data, to predict individual risk and suggest personalised interventions.

    2023–2025 · Horizon Europe consortium

Publications

All publications

Software

All software

  • Physics-informed neural networks for hidden insulin dynamics

    Physics-informed neural networks that estimate insulin dynamics, which are not measured directly, from glucose data.

    GitLabPaper

  • Reduced model of type 2 diabetes progression

    Simulation code for a two-scale model of the long-term effects of physical activity on type 2 diabetes and for its reduced (homogenised) version, which allows fast long-term predictions.

    GitLabPaper

  • Surrogate model of prediabetes progression

    Deep learning surrogate of a mechanistic model of prediabetes progression, for fast patient-specific simulations. Data and trained model are not included.

    GitLabPaper

  • Hierarchical Bayesian inference for ODE systems

    Hierarchical Bayesian estimation of the parameters of ordinary differential equations from multiple sparse data sets, which share information across individuals or experiments.

    GitLabPreprint

  • Spatial regression with PDE penalisation

    Code for spatial regression with partial differential equation penalisation, used to estimate blood flow velocity fields.

    Code

  • Counterfactual inference with differential equation models

    Counterfactual inference for ordinary differential equation models, applied to a model of glucose regulation to assess personalised physical activity plans for the prevention of type 2 diabetes.

    GitLabPaper

Other research lines

  • Causal AI and probabilistic graphical models

    Causal models that estimate the effect of interventions and answer what-if questions, built on probabilistic graphical models that combine data with expert knowledge.

  • Federated learning

    Methods and software to train models on data that remain at the institutions that collected them, such as hospitals and research centres.