The aim of this group is to explore techniques that combines machine learning models with numerical and physics-based simulations to tackle a wide range of problems in scientific disciplines. Some projects include embedding information from PDE systems inside ML models, modelization of dynamical systems, application to control systems, deployment on embedded electronics.
Quadrature Methods for Neural PDE
F. Calisto, S. Azeglio, L. Bottero and V. Pagliarino from MLJC (students at University of Turin) in 2021 collaborated with researchers from MIT Julia Lab and other important institutions for a paper describing and analyzing the performances of NeuralPDE library for PINNs training. (Page)
Main contacts: F. Calisto, S. Azeglio, L. Bottero, V. Pagliarino (francesco.calisto ‘at’ edu.unito.it)
Control Systems on real time electronics using SciML
Main contacts: V. Pagliarino, S. Azeglio, F. Calisto, V. Berta, L. Bottero (valerio.pagliarino ‘at’ edu.unito.it)
The aim of this project is to investigate some promising Scientific Machine Learning architectures that are suitable not only for physics-informed modelization of dynamical systems, but also allows the design of fast adaptive control systems based on such models. Our final goal is the deployment of this control system of cutting-edge electronics including FPGAs, NVIDIA Jetson and custom boards.
Main contacts: A. Semeraro, M. Falco, S. Azeglio (andrea.semeraro450 ‘at’ edu.unito.it)
We try to investigate problems related to ocean circulation by employing Physics Informed Neural Networks. Our aim is to devise techniques which can be helpful in tackling the problem of plastic debris in oceans. Currently, we are working on the Langmuir circulation problem with PINNs.
Main contacts: F. Calisto, L. Bottero, V. Pagliarino, S. Azeglio (francesco.calisto ‘at’ edu.unito.it)
The aim of this work is to evaluate the feasibility of re-implementing some key parts of the widely used Weather Research and Forecasting WRF-SFIRE simulator by replacing its core differential equations numerical solvers with state-of-the-art physics-informed machine learning techniques to solve ODEs and PDEs in order to increase its computational efficiency. The goal is to evaluate the feasibility of a real-time simulator for wildfire spread prediction based on PINNs. Our ML approach is based on Physics Informed Neural Networks implemented in the NeuralPDE.jl library, which turns an integration problem into a minimization one. Furthermore, it can successfully tackle ill-posed problems and it is efficient for high-dimensional PDEs.
High quality resources selected by MLJC on these topics:
- C. Rackauckas et al – Universal Differential Equations for Scientific Machine Learning – ArXiv – https://arxiv.org/abs/2001.04385
- G. E. Karniadakis et al – “Physics-informed machine learning” – Nature Review – Physics – view on www.nature.com