Machine Learning in Physical Sciences: Theory and Applications
May 26, 2025 – May 31, 2025
Registration deadline: April 26, 2025
Location: CECAM-FR-RA
Hosting node: CECAM-FR-RA
More information here: CECAM – Machine Learning in Physical Sciences: Theory and ApplicationsMachine Learning in Physical Sciences: Theory and Applications
Organisers
- Raffaela Cabriolu (Norwegian University of Science and Technology)
- Alex Hansen (Norwegian University of Science and Technology)
- Morten Hjorth-Jensen (University of Oslo)
- Hannes Jónsson (University of Iceland)
- Markus Miettinen (University of Bergen)
With this application, we propose a five-day school on machine learning (ML) techniques for physical sciences, focusing on modeling and direct applications in various fields. The aim of this school is to provide comprehensive ML training within physical science topics that extend beyond well-established applications—such as data processing and developing accurate potentials—and highlight emerging trends in ML applications that are significantly advancing several fields.
ML is increasingly being utilized in quantum chemistry, condensed matter physics, material science, nuclear physics, and other subfields of physics and chemistry. Our school will specifically focus on ML contributions to electronic structure calculations, complex porous media, biomolecular science, and method developments for atomistic modeling.
The rationale for integrating these four sessions into a single school lies in the fundamental principles of ML, which are broadly applicable across various domains of physical sciences. By emphasizing the commonalities of ML approaches and their versatile applications, we aim to equip researchers with the skills and knowledge to apply these techniques in innovative and non-traditional ways. This approach has the potential to significantly extend the boundaries of current research and open new avenues for scientific discovery.
By focusing on these diverse yet interconnected topics, we aim to showcase the robustness and adaptability of ML techniques. This interdisciplinary approach will not only foster cross-disciplinary collaboration but also enable participants to leverage ML techniques in novel contexts, driving innovation and expanding the horizons of physical sciences.
We are confident that this school has strong potential to attract a diverse audience of students, scientists, and researchers, both theorists and experimentalists. For new generations of scientists and researchers, the primary challenge is to understand and harness the full potential of ML techniques applied across various fields and domains that have not yet been covered in previous schools and workshops. We think that this school is needed to establish and maximize the potential of these novel trends for future developments and to keep students, researchers, and scientists up to date.
As experimental frontiers shift towards studying materials and phenomena at different scales, the predictive power of successful phenomenological approaches is increasingly challenged by the scarcity of nearby experimental data to constrain model parameters or assumptions.
In our school in particular, we will investigate the recent development of the ML techniques combined with Many-Body methods, which are increasingly playing a prominent role in enhancing the predictive power of “data-driven” approaches as experiments explore largely uncharted regions. For example, ML is increasingly being used in a broad spectrum of quantum mechanical many-body systems [1,2], with promising results compared with traditional implementations of the state-of-the-art methods like post-Hartree-Fock methods such as full configuration interaction theory and coupled cluster theory, many-body perturbation theory and Green’s function approaches or various Monte Carlo methods [3,4].
ML provides also significant contributions and insights into the field of complex porous media, focusing on the investigation of multiphase flow, transport, reaction, adsorption, and deformation in heterogeneous porous media and materials [5]. ML’s predictive modeling capabilities are enhancing characterization and analysis, optimizing design and processes, and exploring novel applications to better understand and manage fluid flow. Its ability to handle large datasets, uncover complex relationships, and adapt to new information makes ML a powerful tool for advancing the understanding and application of porous materials and media across various fields. These materials have applications in gas storage, catalysis, and separation processes. ML aids in various aspects of their modeling and characterization, including property prediction, structure-property relationships, materials discovery, and overall characterization.
ML plays also a crucial role in analyzing and utilizing vast, freely available simulation-generated datasets, for example in biomolecular modeling [6]. ML algorithms can identify patterns and correlations within these large datasets, leading to new insights into physical and chemical processes, conformational changes, and structural motifs. ML models applied to biomolecular molecular dynamics (MD) trajectories can predict protein folding, ligand binding, and other dynamic processes. They are also capable of identifying critical states and transitions within biomolecular systems. By training and testing ML models on these datasets, researchers can accelerate discoveries, optimize simulations, and gain deeper insights into complex molecular systems.
In classical atomistic simulations, ML introduces novel approaches and enhances traditional computational methods. For example, the creation of high-accuracy potential energy surfaces through new machine learning potentials, significantly accelerating predictions and calculations compared to traditional ab initio methods is now an established practice. ML techniques are also routinely employed to rapidly screen large libraries of materials, identifying candidates with optimal properties for specific applications such as batteries, catalysis, or photovoltaics. Furthermore, when combined with genetic algorithms, ML is used to predict the stability of crystal structures, fostering new discoveries in material science [7,8]. Additionally, ML is used in reaction pathway analysis and kinetics investigations. It facilitates the integration of classical atomistic simulations with quantum computing, thereby enhancing the capabilities and accuracy of these simulations.
The teaching team comprises scientists with expertise spanning many-body methods, quantum field theories, and classical and numerical simulations at the atomistic, meso, and macro scales. Our team members have extensive experience developing and teaching courses in these areas, alongside our ongoing research, which includes quantum engineering, algorithm development, and diverse applications. We believe this broad range of expertise is essential for providing students and researchers novel training in machine learning applications to various problems in the physical sciences, as well as insight into the limitations and potential of different algorithms and results.
We are confident that this program has strong potential to attract a wide audience of students, scientists, and researchers, both theorists and experimentalists.
References
[1] G. Carleo, I. Cirac, K. Cranmer, L. Daudet, M. Schuld, N. Tishby, L. Vogt-Maranto, L. Zdeborová, Rev. Mod. Phys., 91, 045002 (2019)
[2] K. Schütt, S. Chmiela, O. von Lilienfeld, A. Tkatchenko, K. Tsuda, K. Müller, Introduction, 2020
[3] G. Carleo, M. Troyer, Science, 355, 602-606 (2017)
[4] Y.L.A. Schmerwitz , L. Thirion et al., ArXiv:2406.08154v1 [physics.chem-ph]
[5] M. Sahimi, Phys. Rev. E, 109, 041001 (2024)
[6] A. Kiirikki, H. Antila, L. Bort, P. Buslaev, F. Favela-Rosales, T. Ferreira, P. Fuchs, R. Garcia-Fandino, I. Gushchin, B. Kav, N. Kučerka, P. Kula, M. Kurki, A. Kuzmin, A. Lalitha, F. Lolicato, J. Madsen, M. Miettinen, C. Mingham, L. Monticelli, R. Nencini, A. Nesterenko, T. Piggot, Á. Piñeiro, N. Reuter, S. Samantray, F. Suárez-Lestón, R. Talandashti, O. Ollila, Nat. Commun., 15, 1136 (2024)
[7] M. Ceriotti, The Journal of Chemical Physics, 150, (2019)