Hartmut Maennel

Hartmut Maennel

Auto-generated profile, based on publicly available information

Hartmut Maennel

Google DeepMind, Zurich, Switzerland

Research Focus & Metrics

Academic Career Length

7Years (2020-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
5 April 2024 in Science Advances
Verified
Journal Article
Open access
OUOliver T. Unke
Oliver T. Unke
MSMartin Stöhr
Martin Stöhr
SGStefan Ganscha
Stefan Ganscha
TUThomas Unterthiner
Thomas Unterthiner
HMHartmut Maennel
Hartmut Maennel
KMKlaus-Robert Müller
Klaus-Robert Müller
The GEMS method enables molecular dynamics simulations of large heterogeneous systems at ab initio quality.

8 March 2025 in Scientific Data
Verified
Journal Article
Open access
PDF file
SGStefan Ganscha
Stefan Ganscha
OUOliver T. Unke
Oliver T. Unke
DADaniel Ahlin
Daniel Ahlin
HMHartmut Maennel
Hartmut Maennel
SKSergii Kashubin
Sergii Kashubin
KMKlaus-Robert Müller
Klaus-Robert Müller
Machine learning (ML) methods enable prediction of the properties of chemical structures without computationally expensive ab initio calculations. The quality of such predictions depends on the reference data that was used to train the model. In this work, we introduce the QCML dataset: A comprehensive dataset for training ML models for quantum chemistry. The QCML dataset systematically covers chemical space with small molecules consisting of up to 8 heavy atoms and includes elements from a large fraction of the periodic table, as well as different electronic states. Starting from chemical graphs, conformer search and normal mode sampling are used to generate both equilibrium and off-equilibrium 3D structures, for which various properties are calculated with semi-empirical methods (14.7 billion entries) and density functional theory (33.5 million entries). The covered properties include energies, forces, multipole moments, and other quantities, e.g., Kohn-Sham matrices. We provide a first demonstration of the utility of our dataset by training ML-based force fields on the data and applying them to run molecular dynamics simulations.

13 December 2024 in The Journal of Physical Chemistry Letters
Verified
Journal Article
Open access
HMHartmut Maennel
Hartmut Maennel
OUOliver T. Unke
Oliver T. Unke
KMKlaus-Robert Müller
Klaus-Robert Müller
When physical properties of molecules are being modeled with machine learning, it is desirable to incorporate SO(3)-covariance. While such models based on low body order features are not complete, we formulate and prove general completeness properties for higher order methods and show that 6k - 5 of these features are enough for up to k atoms. We also find that the Clebsch-Gordan operations commonly used in these methods can be replaced by matrix multiplications without sacrificing completeness, lowering the scaling from O(l6) to O(l3) in the degree of the features. We apply this to quantum chemistry, but the proposed methods are generally applicable for problems involving three-dimensional point configurations.

11 November 2020 in arXiv
Verified
Preprint
Open access
PDF file
HMHartmut Maennel
Hartmut Maennel
IAIbrahim Alabdulmohsin
Ibrahim Alabdulmohsin
ITIlya Tolstikhin
Ilya Tolstikhin
RBRobert J. N. Baldock
Robert J. N. Baldock
OBOlivier Bousquet
Olivier Bousquet
DKDaniel Keysers
Daniel Keysers
We study deep neural networks (DNNs) trained on natural image data with entirely random labels. Despite its popularity in the literature, where it is often used to study memorization, generalization, and other phenomena, little is known about what DNNs learn in this setting. In this paper, we show analytically for convolutional and fully connected networks that an alignment between the principal components of network parameters and data takes place when training with random labels. We study this alignment effect by investigating neural networks pre-trained on randomly labelled image data and subsequently fine-tuned on disjoint datasets with random or real labels. We show how this alignment produces a positive transfer: networks pre-trained with random labels train faster downstream compared to training from scratch even after accounting for simple effects, such as weight scaling. We analyze how competing effects, such as specialization at later layers, may hide the positive transfer. These effects are studied in several network architectures, including VGG16 and ResNet18, on CIFAR10 and ImageNet.

4 September 2024 in arXiv
Verified
Preprint
Open access
PDF file
HMHartmut Maennel
Hartmut Maennel
OUOliver T. Unke
Oliver T. Unke
KMKlaus-Robert Müller
Klaus-Robert Müller
When modeling physical properties of molecules with machine learning, it is desirable to incorporate $SO(3)$-covariance. While such models based on low body order features are not complete, we formulate and prove general completeness properties for higher order methods, and show that $6k-5$ of these features are enough for up to $k$ atoms. We also find that the Clebsch--Gordan operations commonly used in these methods can be replaced by matrix multiplications without sacrificing completeness, lowering the scaling from $O(l^6)$ to $O(l^3)$ in the degree of the features. We apply this to quantum chemistry, but the proposed methods are generally applicable for problems involving 3D point configurations.