The rapid advancement of machine learning (ML) applications in chemistry and physics has been
driven by the increasing availability of comprehensive quantum-mechanical (QM) datasets.
Recently, we introduced high-fidelity property data for both small [Sci. Data 8, 43, (2021)] and large
[Sci. Data 11, 742, (2024)] drug-like molecules in equilibrium and non-equilibrium states. These
datasets have been instrumental in advancing QM-based ML interatomic potentials
[10.26434/chemrxiv-2024-bdfr0-v2, (2025)] and enhancing semi-empirical (SE) methods [J. Phys.
Chem. Lett., 11, 6835 (2020)], enabling accurate (bio)molecular simulations. In this seminar, I will
first discuss our recent efforts to improve the transferability and generalizability of the ML-
corrected density functional tight-binding method. This approach can now be used to efficiently
compute the energetic, structural, and vibrational properties of large drug-like molecules and
molecular dimers at a higher level of theory. Secondly, I will briefly discuss how the insights gained
from the demonstration of the “freedom of design” principle in the chemical compound space (CCS)
[Chem. Sci. 14, 10702 (2023)] are leveraged to design molecules with desired properties. To this end,
a variational autoencoder (VAE) approach is developed to parameterize the CCS using a finite set
of QM properties [Nat. Commun. 15, 6061 (2024)]. I will showcase the capabilities of this method by
conditionally generating de novo molecular structures, interpolating transition paths for chemical
reactions, and providing insightful insights into property–structure relationships
Leonardo Medrano Sandonas obtained his Bachelor’s and Master’s degrees in Physics from the
National University of San Marcos in Lima, Peru. In 2018, he earned his doctoral degree in
computational modeling of low-dimensional materials at the Dresden University of Technology as
a fellow of the International Max Planck Research School (IMPRS), under the supervision of Prof.
Gianaurelio Cuniberti. During his doctoral studies, he also received a fellowship from the
Deutscher Akademischer Austauschdienst (DAAD) within the Doctoral Programmes in Germany
2015/16. In 2019, he joined the group of Prof. Alexandre Tkatchenko at the University of
Luxembourg as a postdoctoral researcher. There, his work centered on integrating machine
learning with quantum and statistical mechanics to develop physics-inspired neural network
potentials for (bio)molecular simulations and generative AI frameworks for molecular design. In
addition to his theoretical contributions, he actively participates in multidisciplinary projects with
experimental and industrial partners to address current challenges in physics and chemistry (see
Google Scholar page). He also serves as a reviewer for high-impact scientific journals and has
organized several workshops and conferences in recent years. As of February 2024, he has rejoined
Prof. Cuniberti’s Chair as a research scientist.
The rapid advancement of machine learning (ML) applications in chemistry and physics has been
driven by the increasing availability of comprehensive quantum-mechanical (QM) datasets.
Recently, we introduced high-fidelity property data for both small [Sci. Data 8, 43, (2021)] and large
[Sci. Data 11, 742, (2024)] drug-like molecules in equilibrium and non-equilibrium states. These
datasets have been instrumental in advancing QM-based ML interatomic potentials
[10.26434/chemrxiv-2024-bdfr0-v2, (2025)] and enhancing semi-empirical (SE) methods [J. Phys.
Chem. Lett., 11, 6835 (2020)], enabling accurate (bio)molecular simulations. In this seminar, I will
first discuss our recent efforts to improve the transferability and generalizability of the ML-
corrected density functional tight-binding method. This approach can now be used to efficiently
compute the energetic, structural, and vibrational properties of large drug-like molecules and
molecular dimers at a higher level of theory. Secondly, I will briefly discuss how the insights gained
from the demonstration of the “freedom of design” principle in the chemical compound space (CCS)
[Chem. Sci. 14, 10702 (2023)] are leveraged to design molecules with desired properties. To this end,
a variational autoencoder (VAE) approach is developed to parameterize the CCS using a finite set
of QM properties [Nat. Commun. 15, 6061 (2024)]. I will showcase the capabilities of this method by
conditionally generating de novo molecular structures, interpolating transition paths for chemical
reactions, and providing insightful insights into property–structure relationships
Leonardo Medrano Sandonas obtained his Bachelor’s and Master’s degrees in Physics from the
National University of San Marcos in Lima, Peru. In 2018, he earned his doctoral degree in
computational modeling of low-dimensional materials at the Dresden University of Technology as
a fellow of the International Max Planck Research School (IMPRS), under the supervision of Prof.
Gianaurelio Cuniberti. During his doctoral studies, he also received a fellowship from the
Deutscher Akademischer Austauschdienst (DAAD) within the Doctoral Programmes in Germany
2015/16. In 2019, he joined the group of Prof. Alexandre Tkatchenko at the University of
Luxembourg as a postdoctoral researcher. There, his work centered on integrating machine
learning with quantum and statistical mechanics to develop physics-inspired neural network
potentials for (bio)molecular simulations and generative AI frameworks for molecular design. In
addition to his theoretical contributions, he actively participates in multidisciplinary projects with
experimental and industrial partners to address current challenges in physics and chemistry (see
Google Scholar page). He also serves as a reviewer for high-impact scientific journals and has
organized several workshops and conferences in recent years. As of February 2024, he has rejoined
Prof. Cuniberti’s Chair as a research scientist.
