The 4th SMLQC seminar will be given by Pascal Friederich on March 16, 2023 (15:00 Paris | 22:00 Beijing | 10:00 New York).

Title

Adversarial Sampling and Extrapolation Trends in Neural Network Potentials

Abstract

In the last few years, several neural network interatomic potentials (NNIPs) have been proposed to improve model accuracy in a variety of benchmarks. Despite these advances, NNIPs still struggle to generalize outside of its training domain. In this talk, I will describe how neural network generalization can be improved with adversarial sampling and analyzed using loss landscapes. First, I will show how NNIP generalization can be improved by combining uncertainty quantification, active learning, and adversarial attacks. This differentiable sampling strategy leads to an increase in model robustness in production simulations despite using fewer data points than traditional active learning strategies. Furthermore, I will show how extrapolation trends can be derived from loss landscapes of NNIPs, with sharper loss landscapes relating to lower robust generalization. On the other hand, some training routines are shown to improve the loss landscape and generalization ability of the models. This work provides deep learning-based justifications for NNIP extrapolation and can inform the development of next-generation NNIPs. Prepared by LLNL under Contract DE-AC52-07NA27344.

Introduction to the speaker

Daniel Schwalbe-Koda is a Lawrence Fellow at the Lawrence Livermore National Laboratory (California, US), where he leads a project for accelerating materials discovery using high-performance computing and machine learning. He obtained a PhD in Materials Science and Engineering from MIT in 2022.

Seminars on Machine Learning in Quantum Chemistry and Quantum Computing for Quantum Chemistry (SMLQC) are flexible online lectures and tutorials for highlighting recent developments and providing hands-on experience on the title topics as well as networking opportunities. SMLQC sessions are organized to bridge Symposia on Machine Learning and Quantum Chemistry, the first one held in 2021 in Xiamen, China (online) and the second one to be held in 2023 in Uppsala, Sweden (hybrid). The real-time interaction is enabled via a dedicated Slack workspace which already has many researchers active in the title fields.

Format

Format is flexible with roughly biweekly sessions online usually divided into two sections:

– Scientific section

focused on recent developments which can be, e.g., highlighting recently published papers.

Talks are limited to 20 min so that people can attend them during their “coffee break”. The question part will be up to 5 min if scientific section is followed by a tutorial session where the speaker can elaborate more on the questions. Otherwise, if no tutorial session is given, then question part will be open until the speaker or participants give up.

– Tutorial/hands-on section

Dedicated to providing instructions on program use and practical understanding of the underlying ML theory. The materials used in the talks (e.g., Jupyter notebooks, Python scripts, etc.) could be made available for the participants via GitHub, Google Colab, etc. or the SMLQC website under the allowance of the speaker. No strict time limitations but we recommend up to 1 hour.

We particularly encourage and give platform to young researchers (e.g., postdocs, PhDs, etc.) to give seminars.

All our sessions will also have break out rooms in Zoom where participants can have further exchange or fix some specific problems during hands-on session.

If speakers agree, the seminars will be recorded and uploaded on SMLQC YouTube channel.

Speakers

Speakers are invited by the organizing committee and can be selected according to the suggestions of SMLQC participants, e.g., via voting in our Slack workspace. Self-nominations are also encouraged!

Organizers

Current organizers in alphabetic order:

• Pavlo O. Dral
• Alexei A. Kananenka
• Zhenggang Lan
• Max Pinheiro Jr
• Arif Ullah
• Julia Westermayr