[ENCCS Webinar] Machine Learning Interatomic Potentials on HPC: From Fine-tuning to Distillation

Thursday 21 Jan 2027, 14:00 → 15:30 Europe/Stockholm

Online

A joint webinar by ENCCS and the Sweden AI Factory.

About this webinar

Pretrained machine-learning interatomic potentials (MLIPs) are a good starting point, but research often needs more: higher accuracy for a specific system, or much faster models for long and large molecular dynamics (MD) simulations. In this second part of the oh-my-mlip series, we will show two workflows that address these needs on HPC.

The first is fine-tuning: adapting a pretrained model to your own DFT data, using each framework’s own trainer behind a common interface. The second is distillation: training a small, fast student model (NN-MTP, which runs in LAMMPS on CPU nodes) from a large foundation model. On-the-fly active learning adds new teacher calculations wherever the student breaks down, until the student runs stable MD. The result is a tailored MLIP built for your own system: small, fast and accurate where it matters for your simulations.

We will discuss how to split these workflows between GPU and CPU resources, and close with ongoing work on distillation for reactive systems. As in Part 1, hands-on materials will be provided after the webinar. Attending Part 1 is helpful but not required.

Part 1, "Machine Learning Interatomic Potentials on HPC: From Installation to Benchmarking", is on 1 December 2026: see Part 1.

Who is the webinar for

This webinar is for:

• Researchers who already use pretrained MLIPs and want to adapt them to their own systems
• Computational chemists and materials scientists who need long or large MD simulations at a lower cost than a foundation model
• HPC users who want to plan GPU and CPU resources for MLIP training and production MD
• HPC support staff and research software engineers who support MLIP workflows

Basic familiarity with Linux, Python and MD simulations is helpful.

Key takeaways

After this webinar, participants will know:

• When fine-tuning a pretrained MLIP is worth it, and how to prepare training and validation data from their own DFT calculations
• How to fine-tune different MLIP frameworks with the same workflow, using each framework’s own trainer
• How distillation turns a large foundation model (teacher) into a small, fast and accurate MLIP tailored to their own system, running in LAMMPS on CPU nodes
• How on-the-fly active learning checks the student’s stability and accuracy against the teacher
• How to divide a distillation workflow between GPU (teacher calculations and training) and CPU (production MD) resources
• What the trade-offs are between accuracy, cost and speed when choosing among pretrained, fine-tuned and distilled models
• Where this is heading: ongoing work on distillation for reactive systems

Speaker and moderators

Jinuk Moon (Seoul National University)

Jinuk Moon is a PhD student in the Department of Materials Science and Engineering at Seoul National University, working in Prof. Jeong Woo Han’s group. His research combines density functional theory and machine-learning interatomic potentials (MLIPs) for computational catalysis. He develops CatBench, a benchmark framework for evaluating MLIPs on heterogeneous catalysis, and oh-my-mlip, an open-source hub for installing, running, fine-tuning and distilling MLIPs with LLM agents. More about his work on his website.

Moderators: Yonglei Wang and Wei Li (ENCCS)

More events & contact

Check out more upcoming events from ENCCS and our European network at https://enccs.se/events/, as well as available ENCCS lesson materials, suitable also for self-learning.

For questions regarding this webinar or general questions about ENCCS training events, please contact training@enccs.se.