Summary
Computational physical modeling is a key resource to complement theoretical and experimental methods in modern scientific research and engineering. While access to large amount of data has favored the use of Artificial Intelligence and Machine Learning (ML) techniques to enhance physical simulations, limitations of purely data-driven methods have emerged as concerns their generalization capability and their intelligibility. In particular, the latter feature promotes understanding, a fundamental driver for scientific and technical progress, and possibly allows to rigorously investigate the reliability of the models and the safety of the systems based on those models. To overcome these limitations, I propose a hybrid approach that originally combines ML methods and equation-based modeling to significantly improve generalization in small-data scenarios, while guaranteeing the intelligibility of the physical models through the use of symbolic representations. Core of the methodology are learning algorithms that reconstruct models with controllable complexity from data by consistently combining building blocks, which derive from a unifying mathematical framework for physical theories. The system will also incorporate novel human-inspired strategies for knowledge distillation, accumulation and reuse, which are missing in state-of-the-art physical model learning algorithms. To efficiently handle the computational cost associated with the proposed methods, I will implement them in a new software platform that seamlessly integrates automated model learning and high-performance simulation. Thanks to their general-purpose nature, the methods and algorithms developed in this project may be employed in all scientific disciplines and in engineering workflows. In particular, I plan to use them to advance biology and soft robotics by solving challenging modeling tasks.
Unfold all
/
Fold all
More information & hyperlinks
| Web resources: | https://cordis.europa.eu/project/id/101039481 |
| Start date: | 01-09-2022 |
| End date: | 31-08-2027 |
| Total budget - Public funding: | 1 315 000,00 Euro - 1 315 000,00 Euro |
Cordis data
Original description
Computational physical modeling is a key resource to complement theoretical and experimental methods in modern scientific research and engineering. While access to large amount of data has favored the use of Artificial Intelligence and Machine Learning (ML) techniques to enhance physical simulations, limitations of purely data-driven methods have emerged as concerns their generalization capability and their intelligibility. In particular, the latter feature promotes understanding, a fundamental driver for scientific and technical progress, and possibly allows to rigorously investigate the reliability of the models and the safety of the systems based on those models. To overcome these limitations, I propose a hybrid approach that originally combines ML methods and equation-based modeling to significantly improve generalization in small-data scenarios, while guaranteeing the intelligibility of the physical models through the use of symbolic representations. Core of the methodology are learning algorithms that reconstruct models with controllable complexity from data by consistently combining building blocks, which derive from a unifying mathematical framework for physical theories. The system will also incorporate novel human-inspired strategies for knowledge distillation, accumulation and reuse, which are missing in state-of-the-art physical model learning algorithms. To efficiently handle the computational cost associated with the proposed methods, I will implement them in a new software platform that seamlessly integrates automated model learning and high-performance simulation. Thanks to their general-purpose nature, the methods and algorithms developed in this project may be employed in all scientific disciplines and in engineering workflows. In particular, I plan to use them to advance biology and soft robotics by solving challenging modeling tasks.Status
SIGNEDCall topic
ERC-2021-STGUpdate Date
09-02-2023
Images
No images available.
Geographical location(s)
