Machine Learning and AI for the Sciences: toward understanding

Klaus-Robert Müller

Slide at 41:38

Machine Learning for chemical compound space (IPAM)
Laura Maria Sangalli
{Z1,R1}
Ansatz:
instead of
A({Z1,R1})
HY = EY
[Rupp, Tkatchenko, 2012, 2015, Montavon Chmiela et al
Müller & V Lilienfeld 2012, et al Hansen 2013, vLilienfeld et al 2013, et
2015, al 2020, Snyder Keith et al et al 2021, Unke et al forthcoming] 2021, 2024,
2023, Frank et al 2024

References

1.
M. Rupp et al. (2012) Fast and Accurate Modeling of Molecular Atomization Energies with Machine Learning. Physical Review Letters
2.
K. Hansen et al. (2013) Assessment and Validation of Machine Learning Methods for Predicting Molecular Atomization Energies. Journal of Chemical Theory and Computation
3.
J. C. Snyder et al. (2012) Finding Density Functionals with Machine Learning. Physical Review Letters
4.
J. C. Snyder et al. (2015) Nonlinear gradient denoising: Finding accurate extrema from inaccurate functional derivatives. International Journal of Quantum Chemistry
5.
G. Montavon et al. (2013) Machine learning of molecular electronic properties in chemical compound space. New Journal of Physics
6.
O. A. v. Lilienfeld et al. (2020) Exploring chemical compound space with quantum-based machine learning. Nature Reviews Chemistry
7.
K. Hansen et al. (2015) Machine Learning Predictions of Molecular Properties: Accurate Many-Body Potentials and Nonlocality in Chemical Space. The Journal of Physical Chemistry Letters
8.
J. A. Keith et al. (2021) Combining Machine Learning and Computational Chemistry for Predictive Insights Into Chemical Systems. Chemical Reviews
9.
O. T. Unke et al. (2021) Machine Learning Force Fields. Chemical Reviews
10.
S. Chmiela et al. (2023) Accurate global machine learning force fields for molecules with hundreds of atoms. Science Advances
11.
O. T. Unke et al. (2024) Biomolecular dynamics with machine-learned quantum-mechanical force fields trained on diverse chemical fragments. Science Advances
12.
J. T. Frank et al. (2024) A Euclidean transformer for fast and stable machine learned force fields. Nature Communications

Share slide

Summary (AI generated)

In 2011, during another sabbatical, I was invited to IPAM, the Institute for Pure and Applied Mathematics at UCLA. IPAM is known for running workshops and long-term programs, typically lasting three months. During these programs, around 15 faculty members bring along postdocs or students, creating a group of 40-50 individuals who work together in the same building. This environment fosters collaboration and productivity, as everyone is on sabbatical and ready to try new ideas.

During my time at IPAM, I learned about a program focused on understanding the chemical compound space, specifically the Schrödinger equation. My background is in theoretical physics and computer science, so I was intrigued by the challenge of solving this complex equation. Traditionally, methods like density functional theory have been used to approximate solutions to the Schrödinger equation, earning Nobel prizes for their contributions to chemistry.

However, I proposed a different approach. Instead of simplifying and approximating the equation, I suggested treating it as a stochastic problem from the beginning. By using data generated from density functional theory, we could train machine learning models to predict the outcomes of the Schrödinger equation. While this idea was unconventional, I believed it could revolutionize quantum chemistry and improve our understanding of materials, drugs, and chemical interactions.

Quantum chemistry research requires significant computational power, with small molecules taking up to four hours to analyze and materials requiring a month's worth of computing time. This expensive data is crucial for advancing our knowledge of quantum properties and their applications. My goal was to find a way to integrate this data into machine learning models, paving the way for new discoveries in quantum chemistry.