Offre en lien avec l’Action/le Réseau : DSChem/Doctorants
Laboratoire/Entreprise : Laboratoire d’Informatique et des Systèmes (LIS,
Durée : 6 mois
Contact : thierry.artieres@lis-lab.fr
Date limite de publication : 2024-02-16
Contexte :
This internship is part of a collaboration between the CT/ICR and QARMA teams at LIS that aims to promote artificial intelligence (AI) solutions in chemical research at Aix Marseille Université (AMU), a field that is still largely unexplored by the local chemical community. The simplified context of the study is as follows. The object of study for this project is the prediction of quantities of interest for a given molecule. The molecule corresponds to an assembly of atoms interacting via bonds, which possesses a certain energy E as a function of the geometry of the molecule (relative positions of the various atoms), and which is subjected to a force F. The aim is to predict not only the energy E but also the forces F as a function of G. Two features are important in devising a model for predicting the quantities E and F. Firstly, F is equal to the gradient of the energy E. Secondly, there are several energy surfaces of E as a function of G. These surfaces are continuous. During the evolution of a molecule, its geometry can evolve, and there can be a jump from one surface to another e,n a geometry G, all the more likely as the two curves are close for this geometry.
Sujet :
From a Machine Learning point of view, the problem may be viewed as a prediction task where one wants to predict an energy function from the geometric characteristics of a molecule, but also to predict the gradient of this energy finely, for which one also has supervision. Finally, it is a problem that can be modelled as multi-task learning, since it involves predicting several energy surfaces simultaneously.
The aim of the internship is, starting from recent approaches proposed in the field [Batalia et al., 2022, Batzner et al., 2022, Gilmer et al., 2017, Satoki et al., 2024, Thölke et al., 2022], to test and compare them experimentally and then to propose an innovative approach that will enable us to overcome their limitations. We will be able to draw on recent ideas such as multitask learning [Crawshaw, 2020] and gradient learning [Wu et al., 2010], approaches that have not received much attention in the literature.
On the one hand, multitask learning is a paradigm in which several tasks are learned simultaneously to improve the generalisation performance of a learning task using other related tasks. While the typical protocol is to train a model independently to predict energy, gradient it may be beneficial to exploit multitask extensions, which have not been employed to date in this type of domain.
On the other hand, gradient learning is a little-known but potentially valuable framework in which the objective is to learn the gradient of a classification or regression function, with or without supervision. In addition to conventional energy learning using gradient information, we will explore strategies based on explicit learning of the gradient function, starting with neural networks in a multi-output, multi-task framework and extending to other designs.
Références
[Batatia et al., 2022] Ilyes Batatia, Dávid Péter Kovács, Gregor N. C. Simm, Christoph Ortner, Gábor Csányi: MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields. NeurIPS 2022
[Batzner et al., 2022] Batzner, S., Musaelian, A., Sun, L. et al. E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nat Commun 13, 2453 (2022). https://doi.org/10.1038/s41467-022-29939-5
[Crawshaw, 2020] Crawshaw, M. arXiv 2020.Multi-Task Learning with Deep Neural Networks: A Survey, https://doi.org/10.48550/ARXIV.2009.09796
[Gilmer et al., 2017] Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, George E. Dahl: Neural Message Passing for Quantum Chemistry. ICML 2017: 1263-1272
[Satoki et al., 2024] Satoki Ishiai, Ikki Yasuda, Katsuhiro Endo, and Kenji Yasuoka, Graph-Neural-Network-Based Unsupervised Learning of the Temporal Similarity of Structural Features Observed in Molecular Dynamics Simulations, Journal of Chemical Theory and Computation 2024 20 (2), 819-831
[Thölke et al., 2022] Philipp Thölke, Gianni De Fabritiis, Equivariant Transformers for Neural Network based Molecular Potentials. ICLR 2022
[Wu et al., 2010] Qiang Wu, Justin Guinney, Mauro Maggioni, Sayan Mukherjee: Learning Gradients: Predictive Models that Infer Geometry and Statistical Dependence. J. Mach. Learn. Res. 11: 2175-2198 (2010)
Profil du candidat :
Computer science or data science master 2
Last year engineering school
Formation et compétences requises :
Programning : python and deep Learning platform (pytorch or tensorflow)
Machine Learning and deep learning basics
Adresse d’emploi :
Pole scientifique de Chateau Gombert
Marseille