Computational scientist: numerical methods and machine learning for problems governed by physics.
💡 I build efficient, mathematically grounded solvers, from the model to tested code that runs fast.
🎓 Marie Skłodowska-Curie Doctoral Fellow at the University of Pavia (IN-DEEP European Doctoral Network), working on deep learning for PDEs and inverse problems in computational oncology. MSc in Computational Science and Engineering at EPFL, BSc in Mathematics at the University of Pisa.
🏎️ Previously: reinforcement learning for live race strategy at the Alfa Romeo F1 Team.
🤝 Open to collaborations in research and industry: simulation, surrogate modelling, optimisation and machine learning for engineering and scientific problems.
Open-source optimisers for JAX, installable from PyPI, with documentation and examples. Both are drop-in Optimistix minimisers that work on any pytree of parameters.
Quasi-Newton optimisers for smooth, full-batch problems where Adam stalls and you need high accuracy: self-scaled Broyden (SSBroyden, SSBFGS), BFGS and L-BFGS with a strong-Wolfe line search.
pip install ssbroyden-optimistixsol = optx.minimise(loss, ssbroyden_optimistix.SSBroyden(rtol=1e-10, atol=1e-10), params)Natural gradient descent (Gauss–Newton) for least-squares training of neural networks, such as neural PDE solvers: matrix-free CG with randomized preconditioners (Nyström, randomly pivoted Cholesky), so NGD scales to larger networks. Automatic damping, no learning rate to tune.
pip install randngdmetric = randngd.metrics.gauss_newton([(residual, points)], params)
sol = optx.minimise(loss, randngd.ngd_solver(randngd.NystromNGD(metric)), params)| Repository | What it is | Paper |
|---|---|---|
| TT_for_GPE | Riemannian optimisation on low-rank tensor-train manifolds for Gross–Pitaevskii ground states | arXiv:2609.37103 |
| ldnets-tumor-inversion | Latent-dynamics surrogates for uncertainty-aware Bayesian inversion of tumor growth models | in preparation |
| riemannian-spdmatrixeq | Preconditioned low-rank Riemannian optimisation for SPD linear matrix equations | SISC 47(2), 2025 |
The libraries above also come with papers: ssbroyden-optimistix, arXiv:2603.10599; randngd, CMAME 464, 119456 (2027).
- Scientific machine learning: neural PDE solvers and optimisers for training them
- Inverse problems and UQ: Bayesian inversion and model calibration with neural surrogates
- Numerical linear algebra: low-rank methods, Riemannian optimisation, randomised preconditioning
- Numerical PDEs: finite elements and isogeometric analysis, discretisation and error analysis
- patrick-kidger/optimistix: fix of the sign of the Hestenes–Stiefel and Dai–Yuan β in nonlinear conjugate gradient (#241, under review)
- pnkraemer/matfree: randomly pivoted (block) partial Cholesky for low-rank approximation (#295, under review)
- pymanopt/pymanopt: constant line searcher for fixed step sizes, stopping on a NaN gradient norm, and logging of the initial and final iterates (#321, #322, #323, under review)
- JuliaNLSolvers/Optim.jl: fix to preconditioner updates in nonlinear conjugate gradient (merged)
- NicolasBoumal/manopt: new β-update rules for Riemannian conjugate gradient, after Sato (merged)
- elementifiniti2526: exercise material in Julia for the Finite Elements course at the University of Pavia (in Italian)
- Lecturer of Introduction to Julia Programming for Data Science and Scientific Computing, University of Pavia
Publications, talks and CV: ivanbioli.github.io
