Vortragende Person/Vortragende Personen:
Janosh Riebesell

Graph neural network interatomic potentials have emerged as powerful tools for accelerating materials simulation and property prediction.

The latest generation of models approach about ab initio accuracy while maintaining linear scaling of compute cost with system size, promising high quality molecular dynamics at unprecedented time and length scales.

In this talk, he will discuss the development of CHGNet and MACE-MP, two models released in 2023, as well as the Matbench Discovery leaderboard which quantifies the utility of ML in guiding prospective materials discovery.

He will highlight some of the 35 use cases across various chemistry domains we subjected MACE-MP to (the current open SOTA on Matbench Discovery), pointing out what worked and what did not.

The focus will be on an issue that all current models appear to suffer from which we refer to as potential energy surface (PES) over-softening. Recent results indicate PES softening is due to inadequate training data.

He will conclude with highlighting efforts to generate new and better datasets specifically designed as universal ML potential training sets and what users can do to extract maximum performance from existing models on their tasks until models trained on these new datasets become available.

Further information and registration